CN115272250A - Method, apparatus, computer equipment and storage medium for determining lesion location - Google Patents

Abstract

The application relates to a method, a device, a computer device and a storage medium for determining a lesion position. The method comprises the following steps: inputting the magnetic resonance image to a detection model; the detection model comprises a backbone network, a feature fusion network and a decoupling head based on an anchor-free detection framework; extracting features of the magnetic resonance image through each feature extraction module in the backbone network to obtain feature maps with different sizes; respectively carrying out semantic enhancement processing and position enhancement processing on the feature maps with the corresponding sizes through the feature fusion network to obtain enhanced feature maps with all sizes; classifying and regressing the enhanced feature maps of all sizes based on the decoupling heads to obtain processing results corresponding to the enhanced feature maps; and performing decoding processing on each processing result, and determining the position of the focus in the magnetic resonance image based on the decoding result. The method can accurately determine the position of the focus.

Description

Translated fromChinese

确定病灶位置方法、装置、计算机设备和存储介质Method, device, computer equipment and storage medium for determining lesion location

技术领域technical field

本申请涉及图像处理领域，特别是涉及一种确定病灶位置方法、装置、计算机设备和存储介质。The present application relates to the field of image processing, in particular to a method, device, computer equipment and storage medium for determining the location of a lesion.

背景技术Background technique

随着现代医学的蓬勃发展，磁共振成像(Magnetic Resonance Imaging，MRI)作为一种非侵入性的医学图像分析技术被广泛应用。这种医学成像模式对于包括中风、癌症、手术计划、急性损伤等在内的广泛的临床诊断任务至关重要，依据磁共振成像生成的磁共振图像可用于确定病灶位置。With the vigorous development of modern medicine, Magnetic Resonance Imaging (MRI) is widely used as a non-invasive medical image analysis technique. This medical imaging modality is crucial for a wide range of clinical diagnostic tasks including stroke, cancer, surgical planning, acute injury, etc. Magnetic resonance images generated from magnetic resonance imaging can be used to determine the location of lesions.

传统确定病灶位置的方式，通常是医学专业人员根据积累的经验对磁共振图像进行判断来确定，在临床环境下，放射科医生的诊断准确率只能达到64-70％，随着工作量的增加，医生自身的状态可能会受到很大影响，再加上磁共振图像中，不同病灶之间存在大小不一、灰度不均匀、灰度接近等原因，有些病灶在诊断过程中容易被忽略或误判，不可避免的存在确定病灶位置不精确的问题。The traditional way to determine the location of the lesion is usually determined by medical professionals based on the accumulated experience of magnetic resonance images. In the clinical environment, the diagnostic accuracy of radiologists can only reach 64-70%. In addition, in the magnetic resonance image, there are reasons such as different sizes, uneven gray levels, and close gray levels between different lesions, and some lesions are easy to be ignored in the diagnosis process. Or misjudgment, there is inevitably the problem of inaccurate determination of the lesion location.

发明内容Contents of the invention

基于此，有必要针对上述技术问题，提供一种能够提升病灶位置精确度的确定病灶位置方法、装置、计算机设备和计算机可读存储介质。Based on this, it is necessary to address the above technical problems and provide a method, device, computer equipment and computer-readable storage medium for determining the location of a lesion that can improve the accuracy of the location of a lesion.

第一方面，本申请提供了一种确定病灶位置方法。所述方法包括：In a first aspect, the present application provides a method for determining the location of a lesion. The methods include:

将磁共振图像输入至检测模型；所述检测模型包括主干网络、特征融合网络和基于无锚点检测框架的解耦头；The magnetic resonance image is input to the detection model; the detection model includes a backbone network, a feature fusion network and a decoupling head based on an anchor-free detection framework;

通过所述主干网络中的各特征提取模块对磁共振图像进行特征提取，得到不同尺寸的特征图；Perform feature extraction on the magnetic resonance image through each feature extraction module in the backbone network to obtain feature maps of different sizes;

通过所述特征融合网络分别对相应尺寸的特征图进行语义增强处理和位置增强处理，得到各尺寸的增强特征图；Perform semantic enhancement processing and position enhancement processing on feature maps of corresponding sizes through the feature fusion network to obtain enhanced feature maps of each size;

基于所述解耦头对各尺寸的所述增强特征图做分类和回归处理，得到各所述增强特征图对应的处理结果；Perform classification and regression processing on the enhanced feature maps of each size based on the decoupling head to obtain processing results corresponding to each of the enhanced feature maps;

对各所述处理结果进行解码处理，并基于解码的结果确定所述磁共振图像中的病灶位置。Perform decoding processing on each of the processing results, and determine the location of the lesion in the magnetic resonance image based on the decoding result.

在其中一个实施例中，所述主干网络中的特征提取模块包括第一特征提取模块、第二特征提取模块、第三特征提取模块和第四特征提取模块；In one of the embodiments, the feature extraction module in the backbone network includes a first feature extraction module, a second feature extraction module, a third feature extraction module and a fourth feature extraction module;

所述通过所述主干网络中的各特征提取模块对磁共振图像进行特征提取，得到不同尺寸的特征图包括：The feature extraction of the magnetic resonance image by each feature extraction module in the backbone network to obtain feature maps of different sizes includes:

通过所述第一特征提取模块，从所述磁共振图像中提取第一尺寸的特征图；Extracting a feature map of a first size from the magnetic resonance image through the first feature extraction module;

在基于所述第一特征提取模块对所述第一尺寸的特征图进行图像融合后，通过所述第二特征提取模块，从融合后的特征图中提取第二尺寸的特征图；After image fusion is performed on the feature map of the first size based on the first feature extraction module, a feature map of a second size is extracted from the fused feature map by the second feature extraction module;

在基于所述第二特征提取模块对所述第二尺寸的特征图进行图像融合后，通过所述第三特征提取模块，从融合后的特征图中提取第三尺寸的特征图；After image fusion is performed on the feature map of the second size based on the second feature extraction module, a feature map of a third size is extracted from the fused feature map by the third feature extraction module;

在基于所述第三特征提取模块对所述第三尺寸的特征图进行图像融合后，通过所述第四特征提取模块，从融合后的特征图中提取第四尺寸的特征图。After image fusion is performed on the feature map of the third size based on the third feature extraction module, a feature map of a fourth size is extracted from the fused feature map by the fourth feature extraction module.

在其中一个实施例中，其特征在于，所述通过所述第一特征提取模块，从所述磁共振图像中提取第一尺寸的特征图之前，所述方法还包括：In one of the embodiments, it is characterized in that, before extracting the feature map of the first size from the magnetic resonance image through the first feature extraction module, the method further includes:

通过所述主干网络对磁共振图像进行卷积嵌入处理，得到嵌入特征图；Performing convolution embedding processing on the magnetic resonance image through the backbone network to obtain an embedding feature map;

所述通过所述第一特征提取模块，从所述磁共振图像中提取第一尺寸的特征图包括：The extracting a feature map of a first size from the magnetic resonance image through the first feature extraction module includes:

通过所述第一特征提取模块中的动态位置嵌入层对所述嵌入特征图进行动态位置嵌入，得到第一动态特征图；performing dynamic position embedding on the embedded feature map through the dynamic position embedding layer in the first feature extraction module to obtain a first dynamic feature map;

通过所述第一特征提取模块中的多头相关性聚合层对所述第一动态特征图进行局部聚合处理，得到第一聚合特征图；performing local aggregation processing on the first dynamic feature map through the multi-head correlation aggregation layer in the first feature extraction module to obtain a first aggregated feature map;

将所述第一聚合特征图输入所述第一特征提取模块中的前馈网络层进行图像块表征，得到第一尺寸的特征图。Inputting the first aggregated feature map into the feed-forward network layer in the first feature extraction module to perform image block representation to obtain a feature map of the first size.

在其中一个实施例中，所述通过所述第一特征提取模块，从所述磁共振图像中提取第一尺寸的特征图之后，所述方法还包括：In one of the embodiments, after extracting the feature map of the first size from the magnetic resonance image through the first feature extraction module, the method further includes:

对所述第一尺寸的特征图进行降维，得到第一降维特征图；performing dimensionality reduction on the feature map of the first size to obtain a first dimensionality reduction feature map;

对所述第一降维特征图进行归一化，得到第一归一化特征图；normalizing the first dimensionality reduction feature map to obtain a first normalized feature map;

所述通过所述第二特征提取模块，从融合后的特征图中提取第二尺寸的特征图包括：Said extracting the feature map of the second size from the fused feature map by said second feature extraction module includes:

通过所述第二特征提取模块，从所述第一归一化特征图中提取第二尺寸的特征图。A feature map of a second size is extracted from the first normalized feature map by the second feature extraction module.

在其中一个实施例中，所述特征融合网络包括特征金字塔网络和路径聚合网络，所述通过所述特征融合网络分别对相应尺寸的特征图进行语义增强处理和位置增强处理，得到各尺寸的增强特征图包括：In one of the embodiments, the feature fusion network includes a feature pyramid network and a path aggregation network, and the feature fusion network performs semantic enhancement processing and position enhancement processing on feature maps of corresponding sizes through the feature fusion network to obtain enhancements of each size Feature maps include:

通过所述特征金字塔网络分别对相应尺寸的特征图进行上采样处理，得到各尺寸的中间增强图；respectively performing upsampling processing on feature maps of corresponding sizes through the feature pyramid network to obtain intermediate enhanced maps of each size;

通过所述路径聚合网络分别对相应尺寸的所述中间增强图进行下采样处理，得到各尺寸的增强特征图。The intermediate enhancement maps of corresponding sizes are respectively down-sampled through the path aggregation network to obtain enhanced feature maps of each size.

在其中一个实施例中，所述各尺寸的中间增强图包括第一金字塔特征、第二金字塔特征、第三金字塔特征和第四金字塔特征；In one of the embodiments, the intermediate enhancement map of each size includes a first pyramid feature, a second pyramid feature, a third pyramid feature and a fourth pyramid feature;

所述通过所述特征金字塔网络分别对相应尺寸的特征图进行上采样处理，得到各尺寸的中间增强图包括：The feature map of the corresponding size is respectively upsampled through the feature pyramid network, and the intermediate enhanced map of each size is obtained including:

通过所述特征金字塔网络的第一金字塔层，对所述第四尺寸的特征图进行上采样处理，得到第一金字塔特征；Through the first pyramid layer of the feature pyramid network, the feature map of the fourth size is upsampled to obtain the first pyramid feature;

通过所述特征金字塔网络的第二金字塔层，对所述第三尺寸的特征图和所述第一金字塔特征之间的融合特征进行上采样处理，得到第二金字塔特征；Through the second pyramid layer of the feature pyramid network, the fusion feature between the feature map of the third size and the first pyramid feature is upsampled to obtain the second pyramid feature;

通过所述特征金字塔网络的第三金字塔层，对所述第二尺寸的特征图和所述第二金字塔特征之间的融合特征进行上采样处理，得到第三金字塔特征；Through the third pyramid layer of the feature pyramid network, the fusion feature between the feature map of the second size and the second pyramid feature is upsampled to obtain the third pyramid feature;

通过所述特征金字塔网络的第四金字塔层，对所述第一尺寸的特征图和所述第三金字塔特征之间的融合特征进行上采样处理，得到第四金字塔特征。Through the fourth pyramid layer of the feature pyramid network, the fusion feature between the feature map of the first size and the feature of the third pyramid is upsampled to obtain the fourth pyramid feature.

在其中一个实施例中，所述各尺寸的增强特征图包括第一路径聚合特征、第二路径聚合特征、第三路径聚合特征和第四路径聚合特征；所述通过所述路径聚合网络分别对相应尺寸的中间增强图进行下采样处理，得到各尺寸的增强特征图包括：In one of the embodiments, the enhanced feature map of each size includes a first path aggregation feature, a second path aggregation feature, a third path aggregation feature and a fourth path aggregation feature; The intermediate enhancement map of the corresponding size is down-sampled, and the enhanced feature map of each size includes:

通过所述路径聚合网络的第一路径聚合层，对所述第四金字塔特征进行下采样处理，得到所述第一路径聚合特征；Through the first path aggregation layer of the path aggregation network, the fourth pyramid feature is down-sampled to obtain the first path aggregation feature;

通过所述路径聚合网络的第二路径聚合层，对所述第三金字塔特征和所述第一路径聚合特征之间的融合特征进行下采样处理，得到所述第二路径聚合特征；Through the second path aggregation layer of the path aggregation network, the fusion feature between the third pyramid feature and the first path aggregation feature is down-sampled to obtain the second path aggregation feature;

通过所述路径聚合网络的第三路径聚合层，对所述第二金字塔特征和所述第二路径聚合特征之间的融合特征进行下采样处理，得到所述第三路径聚合特征；Through the third path aggregation layer of the path aggregation network, the fusion feature between the second pyramid feature and the second path aggregation feature is down-sampled to obtain the third path aggregation feature;

通过所述路径聚合网络的第四路径聚合层，对所述第一金字塔特征和所述第三路径聚合特征之间的融合特征进行下采样处理，得到所述第四路径聚合特征。Through the fourth path aggregation layer of the path aggregation network, the fusion feature between the first pyramid feature and the third path aggregation feature is down-sampled to obtain the fourth path aggregation feature.

第二方面，本申请还提供了一种确定病灶位置装置。所述装置包括：In a second aspect, the present application also provides a device for determining the location of a lesion. The devices include:

输入模块，用于将磁共振图像输入至检测模型；所述检测模型包括主干网络、特征融合网络和基于无锚点检测框架的解耦头；The input module is used to input the magnetic resonance image into the detection model; the detection model includes a backbone network, a feature fusion network and a decoupling head based on an anchor-free detection framework;

特征提取模块，用于通过所述主干网络中的各特征提取模块对磁共振图像进行特征提取，得到不同尺寸的特征图；The feature extraction module is used to extract the features of the magnetic resonance image through each feature extraction module in the backbone network to obtain feature maps of different sizes;

增强模块，用于通过所述特征融合网络分别对相应尺寸的特征图进行语义增强处理和位置增强处理，得到各尺寸的增强特征图；An enhancement module, configured to perform semantic enhancement processing and position enhancement processing on feature maps of corresponding sizes through the feature fusion network to obtain enhanced feature maps of each size;

分类与回归模块，用于基于所述解耦头对各尺寸的所述增强特征图做分类和回归处理，得到各所述增强特征图对应的处理结果；A classification and regression module, configured to perform classification and regression processing on the enhanced feature maps of each size based on the decoupling head, to obtain processing results corresponding to each of the enhanced feature maps;

解码模块，用于对各所述处理结果进行解码处理，并基于解码的结果确定所述磁共振图像中的病灶位置。The decoding module is configured to perform decoding processing on each of the processing results, and determine the location of the lesion in the magnetic resonance image based on the decoding results.

在其中一个实施例中，所述主干网络中的特征提取模块包括第一特征提取模块、第二特征提取模块、第三特征提取模块和第四特征提取模块；In one of the embodiments, the feature extraction module in the backbone network includes a first feature extraction module, a second feature extraction module, a third feature extraction module and a fourth feature extraction module;

所述特征提取模块还用于通过所述第一特征提取模块，从所述磁共振图像中提取第一尺寸的特征图；在基于所述第一特征提取模块对所述第一尺寸的特征图进行图像融合后，通过所述第二特征提取模块，从融合后的特征图中提取第二尺寸的特征图；在基于所述第二特征提取模块对所述第二尺寸的特征图进行图像融合后，通过所述第三特征提取模块，从融合后的特征图中提取第三尺寸的特征图；在基于所述第三特征提取模块对所述第三尺寸的特征图进行图像融合后，通过所述第四特征提取模块，从融合后的特征图中提取第四尺寸的特征图。The feature extraction module is also used to extract a feature map of a first size from the magnetic resonance image through the first feature extraction module; After performing image fusion, extract a feature map of a second size from the fused feature map by the second feature extraction module; perform image fusion on the feature map of the second size based on the second feature extraction module Finally, extract a feature map of a third size from the fused feature map through the third feature extraction module; after performing image fusion on the feature map of the third size based on the third feature extraction module, by The fourth feature extraction module extracts a feature map of a fourth size from the fused feature map.

在其中一个实施例中，其特征在于，所述特征提取模块还用于通过所述主干网络对磁共振图像进行卷积嵌入处理，得到嵌入特征图；通过所述第一特征提取模块中的动态位置嵌入层对所述嵌入特征图进行动态位置嵌入，得到第一动态特征图；通过所述第一特征提取模块中的多头相关性聚合层对所述第一动态特征图进行局部聚合处理，得到第一聚合特征图；将所述第一聚合特征图输入所述第一特征提取模块中的前馈网络层进行图像块表征，得到第一尺寸的特征图。In one of the embodiments, it is characterized in that the feature extraction module is also used to perform convolution embedding processing on the magnetic resonance image through the backbone network to obtain an embedded feature map; through the dynamic The position embedding layer performs dynamic position embedding on the embedded feature map to obtain a first dynamic feature map; through the multi-head correlation aggregation layer in the first feature extraction module, the first dynamic feature map is partially aggregated to obtain A first aggregated feature map: inputting the first aggregated feature map into a feed-forward network layer in the first feature extraction module to perform image block characterization to obtain a feature map of a first size.

在其中一个实施例中，所述特征提取模块还用于对所述第一尺寸的特征图进行降维，得到第一降维特征图；对所述第一降维特征图进行归一化，得到第一归一化特征图；通过所述第二特征提取模块，从所述第一归一化特征图中提取第二尺寸的特征图。In one of the embodiments, the feature extraction module is further configured to perform dimensionality reduction on the feature map of the first size to obtain a first dimensionality reduction feature map; normalize the first dimensionality reduction feature map, A first normalized feature map is obtained; a feature map of a second size is extracted from the first normalized feature map by the second feature extraction module.

在其中一个实施例中，所述增强模块还用于通过所述特征金字塔网络分别对相应尺寸的特征图进行上采样处理，得到各尺寸的中间增强图；通过所述路径聚合网络分别对相应尺寸的所述中间增强图进行下采样处理，得到各尺寸的增强特征图。In one of the embodiments, the enhancement module is also used to perform upsampling processing on feature maps of corresponding sizes through the feature pyramid network to obtain intermediate enhancement maps of each size; The intermediate enhancement map of the above is subjected to down-sampling processing to obtain enhanced feature maps of each size.

在其中一个实施例中，所述各尺寸的中间增强图包括第一金字塔特征、第二金字塔特征、第三金字塔特征和第四金字塔特征；In one of the embodiments, the intermediate enhancement map of each size includes a first pyramid feature, a second pyramid feature, a third pyramid feature and a fourth pyramid feature;

所述增强模块还用于通过所述特征金字塔网络的第一金字塔层，对所述第四尺寸的特征图进行上采样处理，得到第一金字塔特征；通过所述特征金字塔网络的第二金字塔层，对所述第三尺寸的特征图和所述第一金字塔特征之间的融合特征进行上采样处理，得到第二金字塔特征；通过所述特征金字塔网络的第三金字塔层，对所述第二尺寸的特征图和所述第二金字塔特征之间的融合特征进行上采样处理，得到第三金字塔特征；通过所述特征金字塔网络的第四金字塔层，对所述第一尺寸的特征图和所述第三金字塔特征之间的融合特征进行上采样处理，得到第四金字塔特征。The enhancement module is also used to perform upsampling processing on the feature map of the fourth size through the first pyramid layer of the feature pyramid network to obtain the first pyramid feature; through the second pyramid layer of the feature pyramid network , performing upsampling processing on the fusion feature between the feature map of the third size and the first pyramid feature to obtain the second pyramid feature; through the third pyramid layer of the feature pyramid network, the second The fusion feature between the feature map of the size and the second pyramid feature is subjected to upsampling processing to obtain the third pyramid feature; through the fourth pyramid layer of the feature pyramid network, the feature map of the first size and the feature map of the first size are obtained. The fusion features between the above-mentioned third pyramid features are subjected to up-sampling processing to obtain the fourth pyramid features.

在其中一个实施例中，所述各尺寸的增强特征图包括第一路径聚合特征、第二路径聚合特征、第三路径聚合特征和第四路径聚合特征；所述增强模块还用于通过所述路径聚合网络的第一路径聚合层，对所述第四金字塔特征进行下采样处理，得到所述第一路径聚合特征；通过所述路径聚合网络的第二路径聚合层，对所述第三金字塔特征和所述第一路径聚合特征之间的融合特征进行下采样处理，得到所述第二路径聚合特征；通过所述路径聚合网络的第三路径聚合层，对所述第二金字塔特征和所述第二路径聚合特征之间的融合特征进行下采样处理，得到所述第三路径聚合特征；通过所述路径聚合网络的第四路径聚合层，对所述第一金字塔特征和所述第三路径聚合特征之间的融合特征进行下采样处理，得到所述第四路径聚合特征。In one of the embodiments, the enhanced feature map of each size includes a first path aggregation feature, a second path aggregation feature, a third path aggregation feature and a fourth path aggregation feature; the enhancement module is also used to pass the The first path aggregation layer of the path aggregation network performs down-sampling processing on the fourth pyramid feature to obtain the first path aggregation feature; through the second path aggregation layer of the path aggregation network, the third pyramid The fusion feature between the feature and the first path aggregation feature is down-sampled to obtain the second path aggregation feature; through the third path aggregation layer of the path aggregation network, the second pyramid feature and the The fusion feature between the second path aggregation features is down-sampled to obtain the third path aggregation feature; through the fourth path aggregation layer of the path aggregation network, the first pyramid feature and the third The fusion features between the path aggregation features are down-sampled to obtain the fourth path aggregation feature.

第三方面，本申请还提供了一种计算机设备。所述计算机设备包括存储器和处理器，所述存储器存储有计算机程序，所述处理器执行所述计算机程序时实现上述方法的步骤。In a third aspect, the present application also provides a computer device. The computer device includes a memory and a processor, the memory stores a computer program, and the processor implements the steps of the above method when executing the computer program.

第四方面，本申请还提供了一种计算机可读存储介质。所述计算机可读存储介质，其上存储有计算机程序，所述计算机程序被处理器执行时实现上述方法的步骤。In a fourth aspect, the present application also provides a computer-readable storage medium. The computer-readable storage medium stores a computer program thereon, and when the computer program is executed by a processor, the steps of the above method are realized.

第五方面，本申请还提供了一种计算机程序产品。所述计算机程序产品，包括计算机程序，该计算机程序被处理器执行时实现上述方法的步骤。In a fifth aspect, the present application also provides a computer program product. The computer program product includes a computer program, and when the computer program is executed by a processor, the steps of the above method are realized.

上述确定病灶位置方法、装置、计算机设备和存储介质，通过将磁共振图像输入至检测模型；通过主干网络中的各特征提取模块对磁共振图像进行特征提取，得到不同尺寸的特征图；通过特征融合网络分别对相应尺寸的特征图进行语义增强处理和位置增强处理，得到各尺寸的增强特征图；基于解耦头对各尺寸的增强特征图做分类和回归处理，得到各增强特征图对应的处理结果；对各处理结果进行解码处理，并基于解码的结果确定磁共振图像中的病灶位置。灵活的无锚点的检测框架，可以很好适应大小不一的病灶区域且处理高效，不仅如此，通过检测模型中的主干网络的特征提取、以及特征融合网络的语义增强与位置增强、以及基于无锚点检测框架的解耦头的分类和回归处理，能够精确确定病灶位置。The above method, device, computer equipment and storage medium for determining the location of a lesion, by inputting the magnetic resonance image into the detection model; performing feature extraction on the magnetic resonance image through each feature extraction module in the backbone network, to obtain feature maps of different sizes; The fusion network performs semantic enhancement processing and position enhancement processing on the feature maps of corresponding sizes to obtain the enhanced feature maps of each size; based on the decoupling head, the enhanced feature maps of each size are classified and regressed to obtain the corresponding enhanced feature maps. Processing results: performing decoding processing on each processing result, and determining the location of the lesion in the magnetic resonance image based on the decoding result. The flexible anchor-free detection framework can well adapt to lesion areas of different sizes and is efficient in processing. Not only that, through the feature extraction of the backbone network in the detection model, semantic enhancement and position enhancement of the feature fusion network, and based on Classification and regression processing of a decoupled head without an anchor detection framework enables precise determination of lesion location.

附图说明Description of drawings

图1为一个实施例中确定病灶位置方法的应用环境图；Fig. 1 is an application environment diagram of the method for determining the location of a lesion in one embodiment;

图2为一个实施例中确定病灶位置方法的流程示意图；Figure 2 is a schematic flow chart of a method for determining the location of a lesion in an embodiment;

图3为一个实施例中检测模型的结构示意图；Fig. 3 is a schematic structural diagram of a detection model in an embodiment;

图4为一个实施例中特征提取步骤的流程示意图；Fig. 4 is a schematic flow chart of the feature extraction step in an embodiment;

图5为一个实施例中特征金字塔网络处理步骤的流程示意图；Fig. 5 is a schematic flow chart of a feature pyramid network processing step in one embodiment;

图6为一个实施例中路径增强网络处理步骤的流程示意图；FIG. 6 is a schematic flow diagram of the processing steps of the path enhancement network in one embodiment;

图7为一个实施例中确定病灶位置装置的结构框图；Fig. 7 is a structural block diagram of a device for determining a lesion position in an embodiment;

图8为一个实施例中计算机设备的内部结构图。Figure 8 is a diagram of the internal structure of a computer device in one embodiment.

具体实施方式Detailed ways

为了使本申请的目的、技术方案及优点更加清楚明白，以下结合附图及实施例，对本申请进行进一步详细说明。应当理解，此处描述的具体实施例仅仅用以解释本申请，并不用于限定本申请。In order to make the purpose, technical solution and advantages of the present application clearer, the present application will be further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described here are only used to explain the present application, and are not intended to limit the present application.

本申请实施例提供的确定病灶位置方法，可以应用于如图1所示的应用环境中。其中，终端102通过网络与服务器104进行通信。数据存储系统可以存储服务器104需要处理的数据。数据存储系统可以集成在服务器104上，也可以放在云上或其他网络服务器上。本申请可由终端102或服务器104执行，本实施例以终端102执行为例进行说明。The method for determining the location of a lesion provided in the embodiment of the present application can be applied to the application environment shown in FIG. 1 . Wherein, the terminal 102 communicates with theserver 104 through the network. The data storage system can store data that needs to be processed by theserver 104 . The data storage system can be integrated on theserver 104, or placed on the cloud or other network servers. This application can be executed by the terminal 102 or theserver 104, and this embodiment is described by taking the execution by the terminal 102 as an example.

终端102将磁共振图像输入至检测模型；检测模型包括主干网络、特征融合网络和基于无锚点检测框架的解耦头；终端102通过主干网络中的各特征提取模块对磁共振图像进行特征提取，得到不同尺寸的特征图；终端102通过特征融合网络分别对相应尺寸的特征图进行语义增强处理和位置增强处理，得到各尺寸的增强特征图；终端102基于解耦头对各尺寸的增强特征图做分类和回归处理，得到各增强特征图对应的处理结果；终端102对各处理结果进行解码处理，并基于解码的结果确定磁共振图像中的病灶位置。The terminal 102 inputs the magnetic resonance image into the detection model; the detection model includes a backbone network, a feature fusion network, and a decoupling head based on an anchor-free detection framework; the terminal 102 performs feature extraction on the magnetic resonance image through each feature extraction module in the backbone network , to obtain feature maps of different sizes; terminal 102 respectively performs semantic enhancement processing and position enhancement processing on feature maps of corresponding sizes through the feature fusion network to obtain enhanced feature maps of each size;terminal 102 enhances features of each size based on the decoupling head Classification and regression processing are performed on the maps to obtain processing results corresponding to each enhanced feature map; the terminal 102 performs decoding processing on each processing result, and determines the location of the lesion in the magnetic resonance image based on the decoding result.

其中，终端102可以但不限于是各种个人计算机、笔记本电脑、智能手机、平板电脑、物联网设备和便携式可穿戴设备，物联网设备可为智能音箱、智能电视、智能空调、智能车载设备等。便携式可穿戴设备可为智能手表、智能手环、头戴设备等。服务器104可以用独立的服务器或者是多个服务器组成的服务器集群来实现。Among them, the terminal 102 can be, but not limited to, various personal computers, notebook computers, smart phones, tablet computers, Internet of Things devices and portable wearable devices, and the Internet of Things devices can be smart speakers, smart TVs, smart air conditioners, smart vehicle-mounted devices, etc. . Portable wearable devices can be smart watches, smart bracelets, head-mounted devices, and the like. Theserver 104 can be implemented by an independent server or a server cluster composed of multiple servers.

在一个实施例中，如图2所示，提供了一种确定病灶位置方法，以该方法应用于图1中的终端102为例进行说明，包括以下步骤：In one embodiment, as shown in FIG. 2 , a method for determining the location of a lesion is provided. The method is applied to the terminal 102 in FIG. 1 as an example for illustration, including the following steps:

S202，将磁共振图像输入至检测模型；检测模型包括主干网络、特征融合网络和基于无锚点检测框架的解耦头。S202. Input the magnetic resonance image into the detection model; the detection model includes a backbone network, a feature fusion network, and a decoupling head based on an anchor-free detection framework.

其中，磁共振图像可以指通过磁共振成像技术生成的图像。检测模型可以指病灶检测模型，检测模型可用于检测磁共振图像中的病灶位置。图3为一个实施例中检测模型的结构示意图；如图所示，主干网络可以包括卷积嵌入层和特征提取模块，其中，特征提取模块包括第一特征提取模块、第一特征提取模块、第二特征提取模块、第三特征提取模块和第四特征提取模块，第一特征提取模块包括局部相关性聚合模块和图像融合模块，第二特征提取模块包括局部相关性聚合模块和图像融合模块，第三特征提取模块包括混合相关性聚合模块和图像融合模块，第四特征提取模块包括全局相关性聚合模块。特征融合网络包括特征金字塔网络(Feature Pyramid Networks，FPN)和路径聚合网络(Path AggregationNetwork，PAN)。Wherein, the magnetic resonance image may refer to an image generated by magnetic resonance imaging technology. The detection model may refer to a lesion detection model, and the detection model may be used to detect a lesion position in a magnetic resonance image. Fig. 3 is a schematic structural diagram of a detection model in an embodiment; as shown in the figure, the backbone network may include a convolution embedding layer and a feature extraction module, wherein the feature extraction module includes a first feature extraction module, a first feature extraction module, a second feature extraction module Two feature extraction modules, the third feature extraction module and the fourth feature extraction module, the first feature extraction module includes a local correlation aggregation module and an image fusion module, the second feature extraction module includes a local correlation aggregation module and an image fusion module, and the second feature extraction module includes a local correlation aggregation module and an image fusion module. The three feature extraction modules include a hybrid correlation aggregation module and an image fusion module, and the fourth feature extraction module includes a global correlation aggregation module. Feature fusion networks include Feature Pyramid Networks (FPN) and Path Aggregation Networks (Path AggregationNetwork, PAN).

无锚点检测框架可以指YOLOX检测框架，YOLOX是指YOLO(you only look once，只需一眼)目标检测算法的改进算法，YOLOX构建了一种无锚点(anchor-free)的端到端目标检测框架，并且达到了一流的检测水平。基于无锚点检测框架的解耦头可以指基于YOLOX的解耦头。The anchor-free detection framework can refer to the YOLOX detection framework. YOLOX refers to the improved algorithm of the YOLO (you only look once) target detection algorithm. YOLOX builds an anchor-free end-to-end target The detection framework has reached a first-class detection level. The decoupled head based on the anchor-free detection framework can refer to the decoupled head based on YOLOX.

具体地，终端将磁共振图像输入至检测模型Specifically, the terminal inputs the magnetic resonance image to the detection model

S204，通过主干网络中的各特征提取模块对磁共振图像进行特征提取，得到不同尺寸的特征图。S204, performing feature extraction on the magnetic resonance image through each feature extraction module in the backbone network to obtain feature maps of different sizes.

其中，特征图可以指由主干网络输入至特征金字塔网络的图，如图3所示，第一特征提取模块中经由三个局部相关性聚合模块输出的特征图可为第一尺寸的特征图，第二特征提取模块中经由四个局部相关性聚合模块输出的特征图可为第二尺寸的特征图；第三特征提取模块中经由八个混合相关性聚合模块输出的特征图可为第三尺寸的特征图，第四特征提取模块中经由三个全局相关性聚合模块输出的特征图可为第四尺寸的特征图。Wherein, the feature map can refer to the map input from the backbone network to the feature pyramid network, as shown in Figure 3, the feature map output by the three local correlation aggregation modules in the first feature extraction module can be a feature map of the first size, The feature map output via the four local correlation aggregation modules in the second feature extraction module can be a feature map of the second size; the feature map output via the eight hybrid correlation aggregation modules in the third feature extraction module can be a third size The feature map of the fourth feature extraction module output via the three global correlation aggregation modules may be a feature map of the fourth size.

S206，通过特征融合网络分别对相应尺寸的特征图进行语义增强处理和位置增强处理，得到各尺寸的增强特征图。S206. Perform semantic enhancement processing and position enhancement processing on the feature maps of corresponding sizes through the feature fusion network to obtain enhanced feature maps of each size.

其中，特征融合网络包括特征金字塔网络和路径聚合网络；增强特征图可以指经由特征融合网络中的特征金字塔网络和路径聚合网络处理后输出的特征图。各尺寸的增强特征图包括第一路径聚合特征、第二路径聚合特征、第三路径聚合特征和第四路径聚合特征。Wherein, the feature fusion network includes a feature pyramid network and a path aggregation network; the enhanced feature map may refer to a feature map output after being processed by the feature pyramid network and the path aggregation network in the feature fusion network. The enhanced feature map for each size includes a first path aggregation feature, a second path aggregation feature, a third path aggregation feature and a fourth path aggregation feature.

具体地，终端可以通过特征金字塔网络分别对相应尺寸的特征图进行上采样处理，得到各尺寸的中间增强图；通过路径聚合网络分别对相应尺寸的中间增强图进行下采样处理，得到各尺寸的增强特征图。Specifically, the terminal can respectively upsample the feature maps of corresponding sizes through the feature pyramid network to obtain intermediate enhanced maps of each size; through the path aggregation network, respectively down-sample the intermediate enhanced maps of corresponding sizes to obtain the Enhanced feature maps.

其中，中间增强图可以指不同尺寸的特征图经由特征金字塔网络处理后输出的特征图。各尺寸的中间增强图包括第一金字塔特征、第二金字塔特征、第三金字塔特征和第四金字塔特征。Wherein, the intermediate enhanced map may refer to a feature map output after feature maps of different sizes are processed by a feature pyramid network. The intermediate enhancement map for each size includes a first pyramid feature, a second pyramid feature, a third pyramid feature, and a fourth pyramid feature.

S208，基于解耦头对各尺寸的增强特征图做分类和回归处理，得到各增强特征图对应的处理结果。S208, perform classification and regression processing on the enhanced feature maps of each size based on the decoupling head, and obtain processing results corresponding to each enhanced feature map.

其中，处理结果可以指增强特征图分别进行分类处理和回归处理后的结果。Wherein, the processing result may refer to the result after classification processing and regression processing are respectively performed on the enhanced feature map.

具体地，如图3所示，终端可以基于解耦头对第一路径聚合特征、第二路径聚合特征、第三路径聚合特征和第四路径聚合特征做分类和回归处理，得到第一路径聚合特征对应的第一处理结果，第二路径聚合特征对应的第二处理结果，第三路径聚合特征对应的第三处理结果，第四路径聚合特征对应的第四处理结果。Specifically, as shown in Figure 3, the terminal can perform classification and regression processing on the first path aggregation feature, the second path aggregation feature, the third path aggregation feature, and the fourth path aggregation feature based on the decoupling head to obtain the first path aggregation feature The first processing result corresponding to the feature, the second processing result corresponding to the second path aggregation feature, the third processing result corresponding to the third path aggregation feature, and the fourth processing result corresponding to the fourth path aggregation feature.

其中，第一处理结果，第二处理结果，第三处理结果和第四处理结果是指不同的处理结果。Wherein, the first processing result, the second processing result, the third processing result and the fourth processing result refer to different processing results.

在一个实施例中，在S208之后，终端可以对处理结果进行合并，得到合并结果，再对合并结果进行数组展平操作、拼接操作和数组转置操作，再对数组转置后的结果进行解码处理。In one embodiment, after S208, the terminal can combine the processing results to obtain the combined result, and then perform array flattening, splicing and array transposition operations on the combined results, and then decode the result after the array transposition deal with.

S210，对各处理结果进行解码处理，并基于解码的结果确定磁共振图像中的病灶位置。S210, perform decoding processing on each processing result, and determine a lesion position in the magnetic resonance image based on the decoding result.

其中，病灶位置可以指存在病灶的位置。Wherein, the location of a lesion may refer to a location where a lesion exists.

具体地，终端对各处理结果进行解码处理，得到解码的结果，再对解码的结果进行非极大抑制处理，并基于非极大抑制处理后的结果确定磁共振图形中的病灶位置。Specifically, the terminal performs decoding processing on each processing result to obtain the decoding result, then performs non-maximum suppression processing on the decoded result, and determines the location of the lesion in the magnetic resonance image based on the non-maximum suppression processing result.

在一个实施例中，解码处理可以是转换成对应的预测框格式，也可以指处理结果的在磁共振图像中的坐标还原过程。In an embodiment, the decoding process may be converted into a corresponding prediction frame format, and may also refer to a process of restoring the coordinates of the processing result in the magnetic resonance image.

在一个实施例中，检测网络训练时可以基于目标损失、IoU(交并比)损失以及分类损失，目标损失与分类损失的损失函数均为交叉熵损失，最后使用自适应梯度法优化模型参数直到收敛。In one embodiment, the detection network training can be based on target loss, IoU (intersection over union) loss and classification loss, the loss functions of target loss and classification loss are both cross entropy loss, and finally use the adaptive gradient method to optimize the model parameters until convergence.

上述确定病灶位置方法中，通过将磁共振图像输入至检测模型；通过主干网络中的各特征提取模块对磁共振图像进行特征提取，得到不同尺寸的特征图；通过特征融合网络分别对相应尺寸的特征图进行语义增强处理和位置增强处理，得到各尺寸的增强特征图；基于解耦头对各尺寸的增强特征图做分类和回归处理，得到各增强特征图对应的处理结果；对各处理结果进行解码处理，并基于解码的结果确定磁共振图像中的病灶位置。灵活的无锚点的检测框架，可以很好适应大小不一的病灶区域且处理高效，不仅如此，通过检测模型中的主干网络的特征提取、以及特征融合网络的语义增强与位置增强、以及基于无锚点检测框架的解耦头的分类和回归处理，能够精确确定病灶位置。In the above method for determining the location of the lesion, the magnetic resonance image is input into the detection model; the feature extraction modules of the backbone network are used to extract the features of the magnetic resonance image to obtain feature maps of different sizes; The feature map is subjected to semantic enhancement processing and position enhancement processing to obtain the enhanced feature map of each size; based on the decoupling head, the enhanced feature map of each size is classified and regressed to obtain the processing results corresponding to each enhanced feature map; the processing results of each Decoding processing is performed, and the location of the lesion in the magnetic resonance image is determined based on the decoding result. The flexible anchor-free detection framework can well adapt to lesion areas of different sizes and is efficient in processing. Not only that, through the feature extraction of the backbone network in the detection model, semantic enhancement and position enhancement of the feature fusion network, and based on Classification and regression processing of a decoupled head without an anchor detection framework enables precise determination of lesion locations.

在一个实施例中，如图4所示，特征提取步骤包括：In one embodiment, as shown in Figure 4, the feature extraction step includes:

S402，通过第一特征提取模块，从磁共振图像中提取第一尺寸的特征图。S402. Using the first feature extraction module, extract a feature map of a first size from the magnetic resonance image.

在一个实施例中，在S402之前，终端通过主干网络的卷积嵌入层对磁共振图像进行卷积嵌入处理，得到嵌入特征图。In one embodiment, before S402, the terminal performs convolutional embedding processing on the magnetic resonance image through the convolutional embedding layer of the backbone network to obtain an embedding feature map.

其中，嵌入特征图可以指进行卷积嵌入处理后的特征图。例如，如图3所示，终端可将磁共振图像输入至主干网络的卷积嵌入层进行卷积嵌入处理，得到嵌入特征图。Wherein, the embedded feature map may refer to a feature map after convolutional embedding processing. For example, as shown in FIG. 3 , the terminal may input the magnetic resonance image to the convolutional embedding layer of the backbone network for convolutional embedding processing to obtain an embedded feature map.

具体地，终端可以通过第一特征提取模块中的动态位置嵌入层对嵌入特征图进行动态位置嵌入，得到第一动态特征图；通过第一特征提取模块中的多头相关性聚合层对第一动态特征图进行局部聚合处理，得到第一聚合特征图；将第一聚合特征图输入第一特征提取模块中的前馈网络层进行图像块表征，得到第一尺寸的特征图。Specifically, the terminal can perform dynamic position embedding on the embedded feature map through the dynamic position embedding layer in the first feature extraction module to obtain the first dynamic feature map; through the multi-head correlation aggregation layer in the first feature extraction module, the first dynamic The feature map is locally aggregated to obtain the first aggregated feature map; the first aggregated feature map is input to the feedforward network layer in the first feature extraction module for image block representation to obtain the feature map of the first size.

其中，动态位置嵌入层可用于对图像进行动态位置嵌入，其中动态位置嵌入可通过零填充的深度卷积位置编码来实现，且适用于任意分辨率。第一动态特征图可以指进行动态位置嵌入后输出的特征图，第一动态特征图、第二动态特征图、第三动态特征图和第四动态特征图是不同的动态特征图。多头相关性聚合层可实现图像的上下文编码与相关性学习，多头指的是将图像块分成多组，每个头都分别处理一组通道的信息。第一聚合特征图可以指第一动态特征图进行局部聚合处理后输出的特征图。前馈网络层可用于进一步表征图像。Among them, the dynamic position embedding layer can be used to perform dynamic position embedding on the image, wherein the dynamic position embedding can be realized by zero-filled deep convolutional position coding, and is applicable to any resolution. The first dynamic feature map may refer to a feature map output after dynamic position embedding, and the first dynamic feature map, the second dynamic feature map, the third dynamic feature map, and the fourth dynamic feature map are different dynamic feature maps. The multi-head correlation aggregation layer can realize the context coding and correlation learning of the image. The multi-head refers to dividing the image block into multiple groups, and each head processes the information of a group of channels separately. The first aggregated feature map may refer to a feature map output after local aggregation processing is performed on the first dynamic feature map. Feedforward network layers can be used to further characterize images.

S404，在基于第一特征提取模块对第一尺寸的特征图进行图像融合后，通过第二特征提取模块，从融合后的特征图中提取第二尺寸的特征图。S404. After image fusion is performed on the feature map of the first size based on the first feature extraction module, a feature map of the second size is extracted from the fused feature map by the second feature extraction module.

在一个实施例中，基于第一特征提取模块对第一尺寸的特征图进行图像融合包括：终端对第一尺寸的特征图进行降维，得到第一降维特征图；对第一降维特征图进行归一化，得到第一归一化特征图。In one embodiment, the image fusion of the feature map of the first size based on the first feature extraction module includes: the terminal performs dimensionality reduction on the feature map of the first size to obtain the first dimensionality reduction feature map; the first dimensionality reduction feature The graph is normalized to obtain the first normalized feature map.

其中，第一降维特征图可以指第一尺寸的特征图进行降维后的特征图。第一归一化特征图可以指第一降维特征图进行归一化后的特征图。Wherein, the first dimensionality reduction feature map may refer to a feature map after dimensionality reduction is performed on the feature map of the first size. The first normalized feature map may refer to a feature map after normalization of the first dimensionality reduction feature map.

在一个实施例中，通过第二特征提取模块，从融合后的特征图中提取第二尺寸的特征图包括通过第二特征提取模块，从第一归一化特征图中提取第二尺寸的特征图。In one embodiment, extracting the feature map of the second size from the fused feature map by the second feature extraction module includes extracting the feature of the second size from the first normalized feature map by the second feature extraction module picture.

在一个实施例中，通过第二特征提取模块，从第一归一化特征图中提取第二尺寸的特征图包括终端可以通过第二特征提取模块中的动态位置嵌入层对第一归一化特征图进行动态位置嵌入，得到第二动态特征图；通过第二特征提取模块中的多头相关性聚合层对第二动态特征图进行局部聚合处理，得到第二聚合特征图；将第二聚合特征图输入第二特征提取模块中的前馈网络层进行图像块表征，得到第二尺寸的特征图。In one embodiment, extracting the feature map of the second size from the first normalized feature map through the second feature extraction module includes that the terminal can use the dynamic position embedding layer in the second feature extraction module to perform the first normalization The feature map is embedded in the dynamic position to obtain the second dynamic feature map; the second dynamic feature map is locally aggregated through the multi-head correlation aggregation layer in the second feature extraction module to obtain the second aggregated feature map; the second aggregated feature The image is input to the feed-forward network layer in the second feature extraction module for image block representation to obtain a feature map of the second size.

其中，第二聚合特征图可以指第二动态特征图进行局部聚合处理后输出的特征图。Wherein, the second aggregated feature map may refer to a feature map output after local aggregation processing is performed on the second dynamic feature map.

S406，在基于第二特征提取模块对第二尺寸的特征图进行图像融合后，通过第三特征提取模块，从融合后的特征图中提取第三尺寸的特征图。S406. After image fusion is performed on the feature map of the second size based on the second feature extraction module, a feature map of a third size is extracted from the fused feature map by a third feature extraction module.

具体地，基于第二特征提取模块对第二尺寸的特征图进行图像融合，得到第二归一化特征图；通过第三特征提取模块，依据第一预设注意力机制对第二归一化特征图进行混合聚合处理，得到第三尺寸的特征图。Specifically, image fusion is performed on the feature map of the second size based on the second feature extraction module to obtain the second normalized feature map; the second normalized feature map is obtained according to the first preset attention mechanism through the third feature extraction module The feature map is mixed and aggregated to obtain a feature map of the third dimension.

其中，第一预设注意力机制可以指用于混合聚合处理的注意力机制，第一预设注意力机制可以是窗口自注意力机制(Window Self-Attention)或全局自注意力机制(Self-Attention)，第一预设注意力机制是与第二预设注意力机制不同的注意力机制。Wherein, the first preset attention mechanism can refer to the attention mechanism used for hybrid aggregation processing, and the first preset attention mechanism can be a window self-attention mechanism (Window Self-Attention) or a global self-attention mechanism (Self-Attention) Attention), the first default attention mechanism is an attention mechanism different from the second default attention mechanism.

在一个实施例中，基于第二特征提取模块对第二尺寸的特征图进行图像融合，得到第二归一化特征图包括终端对第二尺寸的特征图进行降维，得到第二降维特征图；对第二降维特征图进行归一化，得到第二归一化特征图。In one embodiment, performing image fusion on the feature map of the second size based on the second feature extraction module to obtain the second normalized feature map includes performing dimensionality reduction on the feature map of the second size by the terminal to obtain the second dimensionality reduction feature Figure; normalize the second dimensionality reduction feature map to obtain the second normalized feature map.

其中，第二归一化特征图可以指对第二尺寸的特征图进行降维后的特征图。Wherein, the second normalized feature map may refer to a feature map obtained by performing dimensionality reduction on the feature map of the second size.

在一个实施例中，通过第三特征提取模块，依据第一预设注意力机制对第二归一化特征图进行混合聚合处理，得到第三尺寸的特征图包括终端可以通过第三特征提取模块中的动态位置嵌入层对第二归一化特征图进行动态位置嵌入，得到第三动态特征图；通过第三特征提取模块中的多头相关性聚合层依据第一预设注意力机制对第二动态特征图进行混合聚合处理，得到第三聚合特征图；将第三聚合特征图输入第三特征提取模块中的前馈网络层进行图像块表征，得到第三尺寸的特征图。In one embodiment, through the third feature extraction module, the second normalized feature map is mixed and aggregated according to the first preset attention mechanism, and the feature map of the third size is obtained. The terminal can pass the third feature extraction module. The dynamic position embedding layer in the second normalized feature map performs dynamic position embedding to obtain the third dynamic feature map; through the multi-head correlation aggregation layer in the third feature extraction module, the second The dynamic feature map is mixed and aggregated to obtain a third aggregated feature map; the third aggregated feature map is input to the feedforward network layer in the third feature extraction module for image block representation to obtain a feature map of a third size.

S408，在基于第三特征提取模块对第三尺寸的特征图进行图像融合后，通过第四特征提取模块，从融合后的特征图中提取第四尺寸的特征图。S408. After image fusion is performed on the feature map of the third size based on the third feature extraction module, a feature map of the fourth size is extracted from the fused feature map by the fourth feature extraction module.

具体地，终端在基于第三特征提取模块对第三尺寸的特征图进行图像融合后，得到第三归一化特征图；通过第四特征提取模块，依据第二预设注意力机制对第三归一化特征图进行特征提取，得到第四尺寸的特征图。Specifically, the terminal obtains the third normalized feature map after performing image fusion on the feature map of the third size based on the third feature extraction module; through the fourth feature extraction module, according to the second preset attention mechanism, the third The normalized feature map is used for feature extraction to obtain a feature map of the fourth dimension.

第二预设注意力机制可以指用于全局聚合处理的注意力机制。第二预设注意力机制可以是窗口自注意力机制或全局自注意力机制。The second preset attention mechanism may refer to an attention mechanism for global aggregation processing. The second preset attention mechanism may be a window self-attention mechanism or a global self-attention mechanism.

在一个实施例中，基于第三特征提取模块对第三尺寸的特征图进行图像融合后，得到第三归一化特征图包括终端对第三尺寸的特征图进行降维，得到第三降维特征图；对第三降维特征图进行归一化，得到第三归一化特征图。In one embodiment, after performing image fusion on the feature map of the third size based on the third feature extraction module, obtaining the third normalized feature map includes performing dimension reduction on the feature map of the third size by the terminal to obtain the third dimension reduction A feature map; normalizing the third dimensionality reduction feature map to obtain a third normalized feature map.

其中，第三归一化特征图可以指对第三尺寸的特征图进行降维后的特征图。Wherein, the third normalized feature map may refer to a feature map obtained by performing dimensionality reduction on the feature map of the third size.

在一个实施例中，通过第四特征提取模块，依据第二预设注意力机制对第三归一化特征图进行特征提取，得到第四尺寸的特征图包括终端可以通过第四特征提取模块中的动态位置嵌入层对第三归一化特征图进行动态位置嵌入，得到第四动态特征图；通过第四特征提取模块中的多头相关性聚合层依据第二预设注意力机制对第四动态特征图进行混合聚合处理，得到第四聚合特征图；将第四聚合特征图输入第四特征提取模块中的前馈网络层进行图像块表征，得到第四尺寸的特征图。In one embodiment, through the fourth feature extraction module, feature extraction is performed on the third normalized feature map according to the second preset attention mechanism, and the feature map of the fourth size is obtained, including that the terminal can pass through the fourth feature extraction module. The dynamic position embedding layer of the third normalized feature map performs dynamic position embedding to obtain the fourth dynamic feature map; through the multi-head correlation aggregation layer in the fourth feature extraction module, the fourth dynamic The feature map is mixed and aggregated to obtain a fourth aggregated feature map; the fourth aggregated feature map is input to the feedforward network layer in the fourth feature extraction module for image block representation, and a feature map of a fourth size is obtained.

本实施例中，通过第一特征提取模块，从磁共振图像中提取第一尺寸的特征图，在基于第一特征提取模块对第一尺寸的特征图进行图像融合后，通过第二特征提取模块，从融合后的特征图中提取第二尺寸的特征图，在基于第二特征提取模块对第二尺寸的特征图进行图像融合后，通过第三特征提取模块，从融合后的特征图中提取第三尺寸的特征图，在基于第三特征提取模块对第三尺寸的特征图进行图像融合后，通过第四特征提取模块，从融合后的特征图中提取第四尺寸的特征图。能够实现对特征进行精确提取。In this embodiment, the feature map of the first size is extracted from the magnetic resonance image through the first feature extraction module, and after image fusion is performed on the feature map of the first size based on the first feature extraction module, the second feature extraction module , extract the feature map of the second size from the fused feature map, after image fusion of the feature map of the second size based on the second feature extraction module, extract from the fused feature map through the third feature extraction module For the feature map of the third size, after image fusion is performed on the feature map of the third size based on the third feature extraction module, the feature map of the fourth size is extracted from the fused feature map through the fourth feature extraction module. Accurate extraction of features can be achieved.

在一个实施例中，如图5所示，特征金字塔网络处理步骤包括：In one embodiment, as shown in Figure 5, the feature pyramid network processing steps include:

S502，通过特征金字塔网络的第一金字塔层，对第四尺寸的特征图进行上采样处理，得到第一金字塔特征。S502. Perform upsampling processing on the feature map of the fourth dimension through the first pyramid layer of the feature pyramid network to obtain the first pyramid feature.

其中，第一金字塔层可以指特征金字塔网络中用于处理第四尺寸的特征图的金字塔层。Wherein, the first pyramid level may refer to the pyramid level used to process the feature map of the fourth dimension in the feature pyramid network.

S504，通过特征金字塔网络的第二金字塔层，对第三尺寸的特征图和第一金字塔特征之间的融合特征进行上采样处理，得到第二金字塔特征。S504. Through the second pyramid layer of the feature pyramid network, perform upsampling processing on the fusion feature between the feature map of the third size and the first pyramid feature to obtain the second pyramid feature.

其中，第二金字塔层可以指特征金字塔网络中用于处理第三尺寸的特征图和第一金字塔特征之间的融合特征的金字塔层。Wherein, the second pyramid level may refer to the pyramid level in the feature pyramid network for processing fusion features between the feature map of the third dimension and the features of the first pyramid.

S506，通过特征金字塔网络的第三金字塔层，对第二尺寸的特征图和第二金字塔特征之间的融合特征进行上采样处理，得到第三金字塔特征。S506. Through the third pyramid layer of the feature pyramid network, perform upsampling processing on the fusion feature between the feature map of the second size and the second pyramid feature to obtain the third pyramid feature.

其中，第三金字塔层可以指特征金字塔网络中用于处理第二尺寸的特征图和第二金字塔特征之间的融合特征的金字塔层。Wherein, the third pyramid level may refer to the pyramid level in the feature pyramid network for processing fusion features between the feature map of the second size and the features of the second pyramid.

S508，通过特征金字塔网络的第四金字塔层，对第一尺寸的特征图和第三金字塔特征之间的融合特征进行上采样处理，得到第四金字塔特征。S508. Through the fourth pyramid layer of the feature pyramid network, perform upsampling processing on the fusion feature between the feature map of the first size and the third pyramid feature to obtain the fourth pyramid feature.

其中，第四金字塔层可以指特征金字塔网络中用于处理第二尺寸的特征图和第二金字塔特征之间的融合特征的金字塔层。Wherein, the fourth pyramid level may refer to the pyramid level in the feature pyramid network for processing fusion features between the feature map of the second size and the features of the second pyramid.

本实施例中，通过特征金字塔网络的第一金字塔层，第二金字塔层，第三金字塔层和第四金字塔层对相应尺寸的特征图，以及相应尺寸的特征图和相应金字塔特征之间的融合特征进行上采样处理，得到相应的金字塔特征，能够增强不同尺寸的特征图中的语义信息与位置信息。In this embodiment, through the first pyramid layer, the second pyramid layer, the third pyramid layer and the fourth pyramid layer of the feature pyramid network, the feature map of the corresponding size, and the fusion between the feature map of the corresponding size and the corresponding pyramid feature The features are up-sampled to obtain the corresponding pyramid features, which can enhance the semantic information and position information in feature maps of different sizes.

在一个实施例中，如图6所示，路径增强网络处理步骤包括：In one embodiment, as shown in Figure 6, the path enhancement network processing steps include:

S602，通过路径聚合网络的第一路径聚合层，对第四金字塔特征进行下采样处理，得到第一路径聚合特征。S602. Through the first path aggregation layer of the path aggregation network, perform down-sampling processing on the fourth pyramid feature to obtain the first path aggregation feature.

其中，第一路径聚合层可以指路径聚合网络中用于处理第四金字塔特征的路径聚合层。Wherein, the first path aggregation layer may refer to the path aggregation layer used to process the fourth pyramid feature in the path aggregation network.

S604，通过路径聚合网络的第二路径聚合层，对第三金字塔特征和第一路径聚合特征之间的融合特征进行下采样处理，得到第二路径聚合特征。S604. Through the second path aggregation layer of the path aggregation network, perform down-sampling processing on the fusion feature between the third pyramid feature and the first path aggregation feature to obtain the second path aggregation feature.

其中，第二路径聚合层可以指路径聚合网络中用于处理第三金字塔特征和第一路径聚合特征之间的融合特征的路径聚合层。Wherein, the second path aggregation layer may refer to the path aggregation layer in the path aggregation network for processing the fusion feature between the third pyramid feature and the first path aggregation feature.

S606，通过路径聚合网络的第三路径聚合层，对第二金字塔特征和第二路径聚合特征之间的融合特征进行下采样处理，得到第三路径聚合特征。S606. Through the third path aggregation layer of the path aggregation network, perform down-sampling processing on the fusion feature between the second pyramid feature and the second path aggregation feature to obtain a third path aggregation feature.

其中，第三路径聚合层可以指路径聚合网络中用于处理第二金字塔特征和第二路径聚合特征之间的融合特征的路径聚合层。Wherein, the third path aggregation layer may refer to the path aggregation layer in the path aggregation network for processing fusion features between the second pyramid features and the second path aggregation features.

S608，通过路径聚合网络的第四路径聚合层，对第一金字塔特征和第三路径聚合特征之间的融合特征进行下采样处理，得到第四路径聚合特征。S608. Through the fourth path aggregation layer of the path aggregation network, perform down-sampling processing on the fusion feature between the first pyramid feature and the third path aggregation feature to obtain a fourth path aggregation feature.

其中，第四路径聚合层可以指路径聚合网络中用于处理第一金字塔特征和第三路径聚合特征之间的融合特征的路径聚合层。Wherein, the fourth path aggregation layer may refer to the path aggregation layer in the path aggregation network for processing fusion features between the first pyramid feature and the third path aggregation feature.

本实施例中，通过路径聚合网络的第一路径聚合层、第二路径聚合层、第三路径聚合层和第四路径聚合层对相应金字塔特征，以及相应金字塔特征和相应路径聚合特征之间的融合特征进行下采样处理，得到相应路径聚合特征，能够增强不同尺寸的特征图中的语义信息与位置信息。In this embodiment, through the first path aggregation layer, the second path aggregation layer, the third path aggregation layer and the fourth path aggregation layer of the path aggregation network, the corresponding pyramid features, and the corresponding pyramid features and the corresponding path aggregation features The fusion features are down-sampled to obtain the corresponding path aggregation features, which can enhance the semantic information and position information in feature maps of different sizes.

作为一个示例，本实施例如下：As an example, this embodiment is as follows:

主干网络的卷积嵌入层实现方法是：以二维图像(磁共振图像)

作为输入，为模型设计一个可学习的图像块嵌入函数f(·)，从而将x作为输入，得到图像特征

f(·)是卷积核大小为k×k，步长s，padding为p的二维卷积运算。图像特征f(x)的通道数为嵌入维度Dim，高和宽分别为：

接着f(x)被展平为hw×Dim的图像块序列，并通过层归一化方法(LayerNorm)进行归一化，以输入到后续模块中。The implementation method of the convolutional embedding layer of the backbone network is: a two-dimensional image (magnetic resonance image)

As input, design a learnable image patch embedding function f( ) for the model, so that x is used as input to obtain image features

f( ) is a two-dimensional convolution operation with a convolution kernel size of k×k, step size s, and padding of p. The number of channels of the image feature f(x) is the embedding dimension Dim, and the height and width are:

Then f(x) is flattened into a hw×Dim sequence of image patches and normalized by a layer normalization method (LayerNorm) for input into subsequent modules.

主干网络的相关性聚合模块由三个关键组件组成：动态位置嵌入(DynamicPosition Embedding，DPE)层、多头相关性聚合(Multi-Head Relation Aggregator，MHRA)层、前馈网络(Feed-Forward Network，FFN)层。对于输入X_in∈R^C×H×W，首先引入动态位置嵌入将位置信息动态集成进上文得到的图像块序列中，动态位置嵌入通常使用的是零填充的深度卷积位置编码，它适用于任意输入分辨率，可充分利用特征顺序进行更好视觉识别，式子表达为：X＝DPE(X_in)+X_in；接着使用多头相关性聚合来实现图像块的上下文编码与相关性学习，多头指的是将图像块分成多组，每个头都分别处理一组通道的信息，局部相关性聚合通常使用大卷积核实现，全局相关性聚合通常使用自注意力机制实现，式子表达为：Y＝MHRA(Norm(X))+X；最后，类似于Transformer(转换)结构，我们添加前馈网络进一步表征图像块，式子表达为：Z＝FFN(Norm(Y))+Y。The correlation aggregation module of the backbone network consists of three key components: Dynamic Position Embedding (DPE) layer, Multi-Head Relation Aggregator (MHRA) layer, Feed-Forward Network (FFN) )Floor. For the input X_in ∈ R^C×H×W , the dynamic position embedding is firstly introduced to dynamically integrate the position information into the sequence of image blocks obtained above. The dynamic position embedding usually uses a zero-filled deep convolutional position code, which is suitable for For any input resolution, the feature order can be fully utilized for better visual recognition, the formula is expressed as: X=DPE(X_in )+X_in ; then multi-head correlation aggregation is used to realize context coding and correlation learning of image blocks , multi-head refers to dividing the image block into multiple groups, and each head processes the information of a group of channels separately. Local correlation aggregation is usually implemented using a large convolution kernel, and global correlation aggregation is usually implemented using a self-attention mechanism. The expression It is: Y=MHRA(Norm(X))+X; finally, similar to the Transformer (transformation) structure, we add a feed-forward network to further characterize the image block, and the expression is: Z=FFN(Norm(Y))+Y .

图像融合模块可以使用卷积层或者全连接层实现，其功能为图像特征的降维。The image fusion module can be implemented using a convolutional layer or a fully connected layer, and its function is to reduce the dimensionality of image features.

图3为一个实施例中检测模型的结构示意图；如图3所示，输入一张大小为H×W的MRI图，首先会经过卷积嵌入层，在其内部使用内核大小为4×4，步长为4×4，输出通道数为64的卷积来对图像做动态位置嵌入，并且在卷积的后面紧接一个正则化LN层对数据进行归一化。卷积嵌入层输出得到H/4×W/4×64的特征图，随后依次输入到三个局部相关性聚合模块中，对特征进行局部相关性聚合，输出的特征图记为F1，这里的核心在于使用内核大小为5×5，padding为2，groups为64的卷积来做局部相关性聚合。图像融合模块都采用内核大小和步长都为2×2×2的卷积。F1先输入到图像融合模块中做特征下采样得到H/8×W/8×128的特征图，下采样卷积后接正则化LN层。接着步骤与前面类似，将下采样得到的特征图依次输入到四个局部相关性聚合模块中得到特征图记为F2，这一层局部相关性聚合模块使用内核大小为5×5，padding为2，groups为128的卷积来做局部相关性聚合。然后F2在图像块融合模块中做特征下采样得到H/16×W/16×320的特征图，将其依次输入到8个混合相关性聚合模块中，进行全局特征建模，输出的特征图记为F3。这8个混合模块分为2组，每组中包含4个模块，前3个为窗口自注意力机制(Window Self-Attention)模块，最后1个为全局自注意力机制(Global Self-Attention)模块。最后，F3经过特征下采样得到4×4×4×512的特征图，依次输入到三个全局相关性聚合模块中，输出结果记为F4。接着，通过主干网络得到了4个不同尺度的特征图F1，F2，F3，F4，输入到YOLOX检测框架中的特征融合网络中，先是在特征金字塔网络中自顶向下将高层的特征信息，利用上采样的方式进行传递到低层次特征中并与之融合，随后通过路径聚合网络自底向上的路径增强特征，使用低层中准确的定位信息来增强整个特征层次结构，从而缩短了低层与高层特征之间的信息路径。Fig. 3 is a schematic diagram of the structure of the detection model in an embodiment; as shown in Fig. 3, inputting an MRI image with a size of H×W will first pass through the convolution embedding layer, and use a kernel size of 4×4 inside it, The step size is 4×4, and the convolution with the output channel number of 64 is used to embedding the dynamic position of the image, and a regularized LN layer is followed by the convolution to normalize the data. The output of the convolutional embedding layer is H/4×W/4×64 feature map, which is then input into three local correlation aggregation modules in turn to perform local correlation aggregation on the features, and the output feature map is marked as F1, where The core is to use a convolution with a kernel size of 5×5, padding of 2, and groups of 64 for local correlation aggregation. The image fusion modules all use convolutions with a kernel size and a stride of 2×2×2. F1 is first input to the image fusion module for feature downsampling to obtain the feature map of H/8×W/8×128, and the downsampling convolution is followed by a regularized LN layer. The next step is similar to the previous one. The feature map obtained by downsampling is input into four local correlation aggregation modules in turn to obtain a feature map marked as F2. The local correlation aggregation module of this layer uses a kernel size of 5×5 and a padding of 2. , groups are 128 convolutions for local correlation aggregation. Then F2 performs feature down-sampling in the image block fusion module to obtain the feature map of H/16×W/16×320, which is input into 8 hybrid correlation aggregation modules in turn for global feature modeling, and the output feature map Denote it as F3. These 8 hybrid modules are divided into 2 groups, each group contains 4 modules, the first 3 are window self-attention mechanism (Window Self-Attention) modules, and the last one is global self-attention mechanism (Global Self-Attention) module. Finally, F3 obtains a 4×4×4×512 feature map through feature downsampling, which is input into three global correlation aggregation modules in turn, and the output result is recorded as F4. Then, four feature maps F1, F2, F3, and F4 of different scales are obtained through the backbone network, which are input into the feature fusion network in the YOLOX detection framework. First, the high-level feature information is top-down in the feature pyramid network. The upsampling method is used to transfer to the low-level features and fuse with them, and then through the bottom-up path enhancement feature of the path aggregation network, the accurate positioning information in the low-level layer is used to enhance the entire feature hierarchy, thereby shortening the low-level and high-level features. Information paths between features.

对于每一层的路径聚合层输出的特征都会输入到一个解耦头中做分类和回归，具体来说，每个解耦头采用1×1的卷积层，将这些输入特征的通道减少到256个，然后增加两个平行分支，每个分支有两个3×3卷积层，分别用于分类和回归任务，最后在回归分支上增加了IoU(交并比)分支来辅助训练。每一层特征经过合并，把分类和回归结果按通道维度进行拼接。For each layer, the features output by the path aggregation layer will be input into a decoupling head for classification and regression. Specifically, each decoupling head uses a 1×1 convolutional layer to reduce the channels of these input features to 256, and then add two parallel branches, each branch has two 3×3 convolutional layers, which are used for classification and regression tasks respectively, and finally add the IoU (intersection-over-union ratio) branch to the regression branch to assist training. The features of each layer are merged, and the classification and regression results are spliced according to the channel dimension.

然后用数组展平操作将特征图展平成向量，从(batch size,channel,height,width)变成(batch size,channel,height×width)，而height×width为这一尺度特征预测的Anchor个数。随后按第三维度做拼接操作合并不同尺度下的结果，让(batch size,channel,height×width)变成(batch size,channel,anchors)。接着是数组转置操作，(batch size,channel,anchors)转置成(batch size,anchors,channel)，其中每一行是一张图像预测的Anchor信息。最后将这个数组进行解码，也就是转换成对应的预测框格式，简单来说解码过程就是预测的结果相对于网格左上角偏移的坐标加上网格的坐标，再乘以下采样倍数，从而还原到原图坐标。Then use the array flattening operation to flatten the feature map into a vector, from (batch size, channel, height, width) to (batch size, channel, height×width), and height×width is the Anchor predicted by this scale feature number. Then perform a stitching operation according to the third dimension to merge the results at different scales, so that (batch size, channel, height×width) becomes (batch size, channel, anchors). Then there is the array transposition operation, (batch size, channel, anchors) is transposed into (batch size, anchors, channel), where each row is the Anchor information of an image prediction. Finally, this array is decoded, that is, converted into the corresponding prediction box format. Simply put, the decoding process is the coordinates of the predicted result offset from the upper left corner of the grid plus the coordinates of the grid, and then multiplied by the downsampling multiple to restore to the original image coordinates.

应该理解的是，虽然如上的各实施例所涉及的流程图中的各个步骤按照箭头的指示依次显示，但是这些步骤并不是必然按照箭头指示的顺序依次执行。除非本文中有明确的说明，这些步骤的执行并没有严格的顺序限制，这些步骤可以以其它的顺序执行。而且，如上的各实施例所涉及的流程图中的至少一部分步骤可以包括多个步骤或者多个阶段，这些步骤或者阶段并不必然是在同一时刻执行完成，而是可以在不同的时刻执行，这些步骤或者阶段的执行顺序也不必然是依次进行，而是可以与其它步骤或者其它步骤中的步骤或者阶段的至少一部分轮流或者交替地执行。It should be understood that although the steps in the flow charts involved in the above embodiments are shown sequentially as indicated by the arrows, these steps are not necessarily executed sequentially in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order restriction on the execution of these steps, and these steps can be executed in other orders. Moreover, at least some of the steps in the flow charts involved in the above embodiments may include multiple steps or stages, and these steps or stages are not necessarily executed at the same time, but may be executed at different times, The execution order of these steps or stages is not necessarily performed sequentially, but may be performed in turn or alternately with other steps or at least a part of steps or stages in other steps.

基于同样的发明构思，本申请实施例还提供了一种用于实现上述所涉及的确定病灶位置方法的确定病灶位置装置。该装置所提供的解决问题的实现方案与上述方法中所记载的实现方案相似，故下面所提供的一个或多个确定病灶位置装置实施例中的具体限定可以参见上文中对于确定病灶位置方法的限定，在此不再赘述。Based on the same inventive concept, an embodiment of the present application further provides a device for determining the position of a lesion for implementing the above-mentioned method for determining a position of a lesion. The solution to the problem provided by the device is similar to the implementation described in the above-mentioned method, so for the specific limitations in one or more embodiments of the device for determining the position of the lesion provided below, please refer to the above-mentioned method for determining the position of the lesion limited and will not be repeated here.

在一个实施例中，如图7所示，提供了一种确定病灶位置装置，包括：输入模块702、特征提取模块704、增强模块706、分类与回归模块708和解码模块710，其中：In one embodiment, as shown in FIG. 7 , a device for determining the location of a lesion is provided, including: aninput module 702, afeature extraction module 704, anenhancement module 706, a classification andregression module 708, and adecoding module 710, wherein:

输入模块702，用于将磁共振图像输入至检测模型；检测模型包括主干网络、特征融合网络和基于无锚点检测框架的解耦头；Theinput module 702 is used to input the magnetic resonance image into the detection model; the detection model includes a backbone network, a feature fusion network and a decoupling head based on an anchor-free detection framework;

特征提取模块704，用于通过主干网络中的各特征提取模块对磁共振图像进行特征提取，得到不同尺寸的特征图；Thefeature extraction module 704 is used to perform feature extraction on the magnetic resonance image through each feature extraction module in the backbone network to obtain feature maps of different sizes;

增强模块706，用于通过特征融合网络分别对相应尺寸的特征图进行语义增强处理和位置增强处理，得到各尺寸的增强特征图；Theenhancement module 706 is used to perform semantic enhancement processing and position enhancement processing on the feature maps of corresponding sizes through the feature fusion network to obtain enhanced feature maps of each size;

分类与回归模块708，用于基于解耦头对各尺寸的增强特征图做分类和回归处理，得到各增强特征图对应的处理结果；The classification andregression module 708 is used to perform classification and regression processing on the enhanced feature maps of each size based on the decoupling head, and obtain the corresponding processing results of each enhanced feature map;

解码模块710，用于对各处理结果进行解码处理，并基于解码的结果确定磁共振图像中的病灶位置。Thedecoding module 710 is configured to perform decoding processing on each processing result, and determine the location of the lesion in the magnetic resonance image based on the decoding result.

在一个实施例中，主干网络中的特征提取模块包括第一特征提取模块、第二特征提取模块、第三特征提取模块和第四特征提取模块；特征提取模块704还用于通过第一特征提取模块，从磁共振图像中提取第一尺寸的特征图；在基于第一特征提取模块对第一尺寸的特征图进行图像融合后，通过第二特征提取模块，从融合后的特征图中提取第二尺寸的特征图；在基于第二特征提取模块对第二尺寸的特征图进行图像融合后，通过第三特征提取模块，从融合后的特征图中提取第三尺寸的特征图；在基于第三特征提取模块对第三尺寸的特征图进行图像融合后，通过第四特征提取模块，从融合后的特征图中提取第四尺寸的特征图。In one embodiment, the feature extraction modules in the backbone network include a first feature extraction module, a second feature extraction module, a third feature extraction module and a fourth feature extraction module; thefeature extraction module 704 is also used to extract module, extracting a feature map of the first size from the magnetic resonance image; after image fusion is performed on the feature map of the first size based on the first feature extraction module, the second feature extraction module is used to extract the first feature map from the fused feature map A feature map of two dimensions; after image fusion is performed on the feature map of the second size based on the second feature extraction module, a feature map of the third size is extracted from the fused feature map through the third feature extraction module; based on the second feature map After the three-feature extraction module performs image fusion on the feature map of the third size, the feature map of the fourth size is extracted from the fused feature map through the fourth feature extraction module.

在其中一个实施例中，其特征在于，特征提取模块704还用于通过主干网络对磁共振图像进行卷积嵌入处理，得到嵌入特征图；通过第一特征提取模块中的动态位置嵌入层对嵌入特征图进行动态位置嵌入，得到第一动态特征图；通过第一特征提取模块中的多头相关性聚合层对第一动态特征图进行局部聚合处理，得到第一聚合特征图；将第一聚合特征图输入第一特征提取模块中的前馈网络层进行图像块表征，得到第一尺寸的特征图。In one of the embodiments, it is characterized in that thefeature extraction module 704 is also used to perform convolution and embedding processing on the magnetic resonance image through the backbone network to obtain an embedded feature map; The feature map is embedded in the dynamic position to obtain the first dynamic feature map; the first dynamic feature map is locally aggregated through the multi-head correlation aggregation layer in the first feature extraction module to obtain the first aggregated feature map; the first aggregated feature The image is input to the feed-forward network layer in the first feature extraction module to perform image block representation to obtain a feature map of the first size.

在其中一个实施例中，特征提取模块704还用于对第一尺寸的特征图进行降维，得到第一降维特征图；对第一降维特征图进行归一化，得到第一归一化特征图；通过第二特征提取模块，从第一归一化特征图中提取第二尺寸的特征图。In one of the embodiments, thefeature extraction module 704 is also used to perform dimensionality reduction on the feature map of the first size to obtain the first dimensionality reduction feature map; to normalize the first dimensionality reduction feature map to obtain the first normalization The feature map is normalized; through the second feature extraction module, the feature map of the second size is extracted from the first normalized feature map.

在其中一个实施例中，特征提取模块704还用于基于第二特征提取模块对第二尺寸的特征图进行图像融合，得到第二归一化特征图；通过第三特征提取模块，依据第一预设注意力机制对第二归一化特征图进行混合聚合处理，得到第三尺寸的特征图；在基于第三特征提取模块对第三尺寸的特征图进行图像融合后，得到第三归一化特征图；通过第四特征提取模块，依据第二预设注意力机制对第三归一化特征图进行特征提取，得到第四尺寸的特征图。In one of the embodiments, thefeature extraction module 704 is also used to perform image fusion on the feature map of the second size based on the second feature extraction module to obtain the second normalized feature map; through the third feature extraction module, according to the first The preset attention mechanism performs mixed aggregation processing on the second normalized feature map to obtain the feature map of the third size; after image fusion is performed on the feature map of the third size based on the third feature extraction module, the third normalized feature map is obtained. feature map; through the fourth feature extraction module, feature extraction is performed on the third normalized feature map according to the second preset attention mechanism to obtain a feature map of a fourth size.

在其中一个实施例中，增强模块706还用于通过特征金字塔网络分别对相应尺寸的特征图进行上采样处理，得到各尺寸的中间增强图；通过路径聚合网络分别对相应尺寸的中间增强图进行下采样处理，得到各尺寸的增强特征图。In one of the embodiments, theenhancement module 706 is also used to perform upsampling processing on the feature maps of corresponding sizes through the feature pyramid network to obtain intermediate enhanced maps of each size; respectively perform upsampling on the intermediate enhanced maps of corresponding sizes through the path aggregation network Downsampling processing to obtain enhanced feature maps of each size.

在其中一个实施例中，各尺寸的中间增强图包括第一金字塔特征、第二金字塔特征、第三金字塔特征和第四金字塔特征；增强模块706还用于通过特征金字塔网络的第一金字塔层，对第四尺寸的特征图进行上采样处理，得到第一金字塔特征；通过特征金字塔网络的第二金字塔层，对第三尺寸的特征图和第一金字塔特征之间的融合特征进行上采样处理，得到第二金字塔特征；通过特征金字塔网络的第三金字塔层，对第二尺寸的特征图和第二金字塔特征之间的融合特征进行上采样处理，得到第三金字塔特征；通过特征金字塔网络的第四金字塔层，对第一尺寸的特征图和第三金字塔特征之间的融合特征进行上采样处理，得到第四金字塔特征。In one of the embodiments, the intermediate enhancement map of each size includes the first pyramid feature, the second pyramid feature, the third pyramid feature and the fourth pyramid feature; theenhancement module 706 is also used to pass through the first pyramid layer of the feature pyramid network, Perform upsampling processing on the feature map of the fourth dimension to obtain the first pyramid feature; through the second pyramid layer of the feature pyramid network, perform upsampling processing on the fusion feature between the feature map of the third dimension and the first pyramid feature, Obtain the second pyramid feature; through the third pyramid layer of the feature pyramid network, perform upsampling processing on the fusion feature between the feature map of the second size and the second pyramid feature, and obtain the third pyramid feature; through the feature pyramid network The first The fourth pyramid layer performs upsampling processing on the fusion feature between the feature map of the first size and the third pyramid feature to obtain the fourth pyramid feature.

在其中一个实施例中，各尺寸的增强特征图包括第一路径聚合特征、第二路径聚合特征、第三路径聚合特征和第四路径聚合特征；增强模块706还用于通过路径聚合网络的第一路径聚合层，对第四金字塔特征进行下采样处理，得到第一路径聚合特征；通过路径聚合网络的第二路径聚合层，对第三金字塔特征和第一路径聚合特征之间的融合特征进行下采样处理，得到第二路径聚合特征；通过路径聚合网络的第三路径聚合层，对第二金字塔特征和第二路径聚合特征之间的融合特征进行下采样处理，得到第三路径聚合特征；通过路径聚合网络的第四路径聚合层，对第一金字塔特征和第三路径聚合特征之间的融合特征进行下采样处理，得到第四路径聚合特征。In one of the embodiments, the enhanced feature map of each size includes the first path aggregation feature, the second path aggregation feature, the third path aggregation feature and the fourth path aggregation feature; theenhancement module 706 is also used to pass the path aggregation network's A path aggregation layer performs down-sampling processing on the fourth pyramid feature to obtain the first path aggregation feature; through the second path aggregation layer of the path aggregation network, the fusion feature between the third pyramid feature and the first path aggregation feature is performed Downsampling processing to obtain the second path aggregation feature; through the third path aggregation layer of the path aggregation network, downsampling the fusion feature between the second pyramid feature and the second path aggregation feature to obtain the third path aggregation feature; Through the fourth path aggregation layer of the path aggregation network, the fusion feature between the first pyramid feature and the third path aggregation feature is down-sampled to obtain the fourth path aggregation feature.

上述实施例，通过将磁共振图像输入至检测模型；通过主干网络中的各特征提取模块对磁共振图像进行特征提取，得到不同尺寸的特征图；通过特征融合网络分别对相应尺寸的特征图进行语义增强处理和位置增强处理，得到各尺寸的增强特征图；基于解耦头对各尺寸的增强特征图做分类和回归处理，得到各增强特征图对应的处理结果；对各处理结果进行解码处理，并基于解码的结果确定磁共振图像中的病灶位置。灵活的无锚点的检测框架，可以很好适应大小不一的病灶区域且处理高效，不仅如此，通过检测模型中的主干网络的特征提取、以及特征融合网络的语义增强与位置增强、以及基于无锚点检测框架的解耦头的分类和回归处理，能够精确确定病灶位置。In the above embodiment, by inputting the magnetic resonance image into the detection model; performing feature extraction on the magnetic resonance image through each feature extraction module in the backbone network, to obtain feature maps of different sizes; Semantic enhancement processing and position enhancement processing to obtain enhanced feature maps of each size; based on the decoupling head, perform classification and regression processing on the enhanced feature maps of each size to obtain the processing results corresponding to each enhanced feature map; decode each processing result , and determine the location of the lesion in the magnetic resonance image based on the decoding result. The flexible anchor-free detection framework can well adapt to lesion areas of different sizes and is efficient in processing. Not only that, through the feature extraction of the backbone network in the detection model, semantic enhancement and position enhancement of the feature fusion network, and based on Classification and regression processing of a decoupled head without an anchor detection framework enables precise determination of lesion location.

上述确定病灶位置装置中的各个模块可全部或部分通过软件、硬件及其组合来实现。上述各模块可以硬件形式内嵌于或独立于计算机设备中的处理器中，也可以以软件形式存储于计算机设备中的存储器中，以便于处理器调用执行以上各个模块对应的操作。Each module in the above-mentioned device for determining the position of a lesion may be fully or partially realized by software, hardware or a combination thereof. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, and can also be stored in the memory of the computer device in the form of software, so that the processor can invoke and execute the corresponding operations of the above-mentioned modules.

在一个实施例中，提供了一种计算机设备，该计算机设备可以是终端或服务器，本实施例以计算机设备为终端为例进行说明，其内部结构图可以如图8所示。该计算机设备包括处理器、存储器、输入/输出接口、通信接口、显示单元和输入装置。其中，处理器、存储器和输入/输出接口通过系统总线连接，通信接口、显示单元和输入装置通过输入/输出接口连接到系统总线。其中，该计算机设备的处理器用于提供计算和控制能力。该计算机设备的存储器包括非易失性存储介质、内存储器。该非易失性存储介质存储有操作系统和计算机程序。该内存储器为非易失性存储介质中的操作系统和计算机程序的运行提供环境。该计算机设备的输入/输出接口用于处理器与外部设备之间交换信息。该计算机设备的通信接口用于与外部的终端进行有线或无线方式的通信，无线方式可通过WIFI、移动蜂窝网络、NFC(近场通信)或其他技术实现。该计算机程序被处理器执行时以实现一种确定病灶位置方法。该计算机设备的显示单元用于形成视觉可见的画面，可以是显示屏、投影装置或虚拟现实成像装置，显示屏可以是液晶显示屏或电子墨水显示屏，该计算机设备的输入装置可以是显示屏上覆盖的触摸层，也可以是计算机设备外壳上设置的按键、轨迹球或触控板，还可以是外接的键盘、触控板或鼠标等。In one embodiment, a computer device is provided, and the computer device may be a terminal or a server. In this embodiment, the computer device is used as a terminal as an example for illustration, and its internal structure diagram may be shown in FIG. 8 . The computer device includes a processor, a memory, an input/output interface, a communication interface, a display unit and an input device. Wherein, the processor, the memory and the input/output interface are connected through the system bus, and the communication interface, the display unit and the input device are connected to the system bus through the input/output interface. Wherein, the processor of the computer device is used to provide calculation and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The input/output interface of the computer device is used for exchanging information between the processor and external devices. The communication interface of the computer device is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner can be realized through WIFI, mobile cellular network, NFC (Near Field Communication) or other technologies. When the computer program is executed by the processor, a method for determining the location of the lesion is realized. The display unit of the computer equipment is used to form a visually visible picture, and may be a display screen, a projection device or a virtual reality imaging device, the display screen may be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer equipment may be a display screen The touch layer covered on the upper surface may also be a button, a trackball or a touch pad arranged on the casing of the computer device, or an external keyboard, touch pad or mouse.

本领域技术人员可以理解，图8中示出的结构，仅仅是与本申请方案相关的部分结构的框图，并不构成对本申请方案所应用于其上的计算机设备的限定，具体的计算机设备可以包括比图中所示更多或更少的部件，或者组合某些部件，或者具有不同的部件布置。Those skilled in the art can understand that the structure shown in FIG. 8 is only a block diagram of a partial structure related to the solution of this application, and does not constitute a limitation on the computer equipment to which the solution of this application is applied. The specific computer equipment can be More or fewer components than shown in the figures may be included, or some components may be combined, or have a different arrangement of components.

在一个实施例中，提供了一种计算机设备，包括存储器和处理器，存储器中存储有计算机程序，该处理器执行计算机程序时实现上述各实施例。In one embodiment, a computer device is provided, including a memory and a processor, where a computer program is stored in the memory, and the above embodiments are realized when the processor executes the computer program.

在一个实施例中，提供了一种计算机可读存储介质，其上存储有计算机程序，计算机程序被处理器执行时实现上述各实施例。In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored, and when the computer program is executed by a processor, the above-mentioned embodiments are realized.

在一个实施例中，提供了一种计算机程序产品，包括计算机程序，该计算机程序被处理器执行时实现上述各实施例。In one embodiment, a computer program product is provided, including a computer program, which implements the above-mentioned embodiments when executed by a processor.

需要说明的是，本申请所涉及的用户信息(包括但不限于用户设备信息、用户个人信息等)和数据(包括但不限于用于分析的数据、存储的数据、展示的数据等)，均为经用户授权或者经过各方充分授权的信息和数据，且相关数据的收集、使用和处理需要遵守相关国家和地区的相关法律法规和标准。It should be noted that the user information (including but not limited to user equipment information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all It is information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data need to comply with relevant laws, regulations and standards of relevant countries and regions.

本领域普通技术人员可以理解实现上述实施例方法中的全部或部分流程，是可以通过计算机程序来指令相关的硬件来完成，的计算机程序可存储于一非易失性计算机可读取存储介质中，该计算机程序在执行时，可包括如上述各方法的实施例的流程。其中，本申请所提供的各实施例中所使用的对存储器、数据库或其它介质的任何引用，均可包括非易失性和易失性存储器中的至少一种。非易失性存储器可包括只读存储器(Read-OnlyMemory，ROM)、磁带、软盘、闪存、光存储器、高密度嵌入式非易失性存储器、阻变存储器(ReRAM)、磁变存储器(Magnetoresistive Random Access Memory，MRAM)、铁电存储器(Ferroelectric Random Access Memory，FRAM)、相变存储器(Phase Change Memory，PCM)、石墨烯存储器等。易失性存储器可包括随机存取存储器(Random Access Memory，RAM)或外部高速缓冲存储器等。作为说明而非局限，RAM可以是多种形式，比如静态随机存取存储器(Static Random Access Memory，SRAM)或动态随机存取存储器(Dynamic RandomAccess Memory，DRAM)等。本申请所提供的各实施例中所涉及的数据库可包括关系型数据库和非关系型数据库中至少一种。非关系型数据库可包括基于区块链的分布式数据库等，不限于此。本申请所提供的各实施例中所涉及的处理器可为通用处理器、中央处理器、图形处理器、数字信号处理器、可编程逻辑器、基于量子计算的数据处理逻辑器等，不限于此。Those of ordinary skill in the art can understand that realizing all or part of the processes in the methods of the above embodiments can be completed by instructing related hardware through computer programs, and the computer programs can be stored in a non-volatile computer-readable storage medium , when the computer program is executed, it may include the procedures of the embodiments of the above-mentioned methods. Wherein, any reference to storage, database or other media used in the various embodiments provided in the present application may include at least one of non-volatile and volatile storage. Non-volatile memory can include read-only memory (Read-Only Memory, ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive variable memory (ReRAM), magnetic variable memory (Magnetoresistive Random Access Memory, MRAM), Ferroelectric Random Access Memory (FRAM), Phase Change Memory (Phase Change Memory, PCM), graphene memory, etc. The volatile memory may include random access memory (Random Access Memory, RAM) or external cache memory. As an illustration and not a limitation, the RAM can be in various forms, such as static random access memory (Static Random Access Memory, SRAM) or dynamic random access memory (Dynamic Random Access Memory, DRAM). The databases involved in the various embodiments provided in this application may include at least one of a relational database and a non-relational database. The non-relational database may include a blockchain-based distributed database, etc., but is not limited thereto. The processors involved in the various embodiments provided by this application can be general-purpose processors, central processing units, graphics processors, digital signal processors, programmable logic devices, data processing logic devices based on quantum computing, etc., and are not limited to this.

以上实施例的各技术特征可以进行任意的组合，为使描述简洁，未对上述实施例中的各个技术特征所有可能的组合都进行描述，然而，只要这些技术特征的组合不存在矛盾，都应当认为是本说明书记载的范围。The technical features of the above embodiments can be combined arbitrarily. To make the description concise, all possible combinations of the technical features in the above embodiments are not described. However, as long as there is no contradiction in the combination of these technical features, they should be It is considered to be within the range described in this specification.

以上实施例仅表达了本申请的几种实施方式，其描述较为具体和详细，但并不能因此而理解为对本申请专利范围的限制。应当指出的是，对于本领域的普通技术人员来说，在不脱离本申请构思的前提下，还可以做出若干变形和改进，这些都属于本申请的保护范围。因此，本申请的保护范围应以所附权利要求为准。The above examples only express several implementation modes of the present application, and the description thereof is relatively specific and detailed, but should not be construed as limiting the patent scope of the present application. It should be noted that those skilled in the art can make several modifications and improvements without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application should be determined by the appended claims.

Claims (10)

Translated fromChinese

1.一种确定病灶位置方法，其特征在于，所述方法包括：1. A method for determining the location of a lesion, characterized in that the method comprises:将磁共振图像输入至检测模型；所述检测模型包括主干网络、特征融合网络和基于无锚点检测框架的解耦头；The magnetic resonance image is input to the detection model; the detection model includes a backbone network, a feature fusion network and a decoupling head based on an anchor-free detection framework;通过所述主干网络中的各特征提取模块对磁共振图像进行特征提取，得到不同尺寸的特征图；Perform feature extraction on the magnetic resonance image through each feature extraction module in the backbone network to obtain feature maps of different sizes;通过所述特征融合网络分别对相应尺寸的特征图进行语义增强处理和位置增强处理，得到各尺寸的增强特征图；Perform semantic enhancement processing and position enhancement processing on feature maps of corresponding sizes through the feature fusion network to obtain enhanced feature maps of each size;基于所述解耦头对各尺寸的所述增强特征图做分类和回归处理，得到各所述增强特征图对应的处理结果；Perform classification and regression processing on the enhanced feature maps of each size based on the decoupling head to obtain processing results corresponding to each of the enhanced feature maps;对各所述处理结果进行解码处理，并基于解码的结果确定所述磁共振图像中的病灶位置。Perform decoding processing on each of the processing results, and determine the location of the lesion in the magnetic resonance image based on the decoding result.2.根据权利要求1所述的方法，其特征在于，所述主干网络中的特征提取模块包括第一特征提取模块、第二特征提取模块、第三特征提取模块和第四特征提取模块；2. The method according to claim 1, wherein the feature extraction module in the backbone network comprises a first feature extraction module, a second feature extraction module, a third feature extraction module and a fourth feature extraction module;所述通过所述主干网络中的各特征提取模块对磁共振图像进行特征提取，得到不同尺寸的特征图包括：The feature extraction of the magnetic resonance image by each feature extraction module in the backbone network to obtain feature maps of different sizes includes:通过所述第一特征提取模块，从所述磁共振图像中提取第一尺寸的特征图；Extracting a feature map of a first size from the magnetic resonance image through the first feature extraction module;在基于所述第一特征提取模块对所述第一尺寸的特征图进行图像融合后，通过所述第二特征提取模块，从融合后的特征图中提取第二尺寸的特征图；After image fusion is performed on the feature map of the first size based on the first feature extraction module, a feature map of a second size is extracted from the fused feature map by the second feature extraction module;在基于所述第二特征提取模块对所述第二尺寸的特征图进行图像融合后，通过所述第三特征提取模块，从融合后的特征图中提取第三尺寸的特征图；After image fusion is performed on the feature map of the second size based on the second feature extraction module, a feature map of a third size is extracted from the fused feature map by the third feature extraction module;在基于所述第三特征提取模块对所述第三尺寸的特征图进行图像融合后，通过所述第四特征提取模块，从融合后的特征图中提取第四尺寸的特征图。After image fusion is performed on the feature map of the third size based on the third feature extraction module, a feature map of a fourth size is extracted from the fused feature map by the fourth feature extraction module.3.根据权利要求2所述的方法，其特征在于，所述通过所述第一特征提取模块，从所述磁共振图像中提取第一尺寸的特征图之前，所述方法还包括：3. The method according to claim 2, characterized in that, before the feature map of the first size is extracted from the magnetic resonance image by the first feature extraction module, the method further comprises:通过所述主干网络对磁共振图像进行卷积嵌入处理，得到嵌入特征图；Performing convolution embedding processing on the magnetic resonance image through the backbone network to obtain an embedding feature map;所述通过所述第一特征提取模块，从所述磁共振图像中提取第一尺寸的特征图包括：The extracting a feature map of a first size from the magnetic resonance image through the first feature extraction module includes:通过所述第一特征提取模块中的动态位置嵌入层对所述嵌入特征图进行动态位置嵌入，得到第一动态特征图；performing dynamic position embedding on the embedded feature map through the dynamic position embedding layer in the first feature extraction module to obtain a first dynamic feature map;通过所述第一特征提取模块中的多头相关性聚合层对所述第一动态特征图进行局部聚合处理，得到第一聚合特征图；performing local aggregation processing on the first dynamic feature map through the multi-head correlation aggregation layer in the first feature extraction module to obtain a first aggregated feature map;将所述第一聚合特征图输入所述第一特征提取模块中的前馈网络层进行图像块表征，得到第一尺寸的特征图。Inputting the first aggregated feature map into the feed-forward network layer in the first feature extraction module to perform image block representation to obtain a feature map of the first size.4.根据权利要求2所述的方法，其特征在于，所述通过所述第一特征提取模块，从所述磁共振图像中提取第一尺寸的特征图之后，所述方法还包括：4. method according to claim 2, is characterized in that, described by described first feature extraction module, after extracting the feature map of first size from described magnetic resonance image, described method also comprises:对所述第一尺寸的特征图进行降维，得到第一降维特征图；performing dimensionality reduction on the feature map of the first size to obtain a first dimensionality reduction feature map;对所述第一降维特征图进行归一化，得到第一归一化特征图；normalizing the first dimensionality reduction feature map to obtain a first normalized feature map;所述通过所述第二特征提取模块，从融合后的特征图中提取第二尺寸的特征图包括：Said extracting the feature map of the second size from the fused feature map by said second feature extraction module includes:通过所述第二特征提取模块，从所述第一归一化特征图中提取第二尺寸的特征图。A feature map of a second size is extracted from the first normalized feature map by the second feature extraction module.5.根据权利要求2所述的方法，其特征在于，所述特征融合网络包括特征金字塔网络和路径聚合网络，所述通过所述特征融合网络分别对相应尺寸的特征图进行语义增强处理和位置增强处理，得到各尺寸的增强特征图包括：5. The method according to claim 2, wherein the feature fusion network includes a feature pyramid network and a path aggregation network, and the feature map of the corresponding size is respectively subjected to semantic enhancement processing and location by the feature fusion network. Enhanced processing, the enhanced feature maps of each size are obtained including:通过所述特征金字塔网络分别对相应尺寸的特征图进行上采样处理，得到各尺寸的中间增强图；respectively performing upsampling processing on feature maps of corresponding sizes through the feature pyramid network to obtain intermediate enhanced maps of each size;通过所述路径聚合网络分别对相应尺寸的所述中间增强图进行下采样处理，得到各尺寸的增强特征图。The intermediate enhancement maps of corresponding sizes are respectively down-sampled through the path aggregation network to obtain enhanced feature maps of each size.6.根据权利要求5所述的方法，其特征在于，所述各尺寸的中间增强图包括第一金字塔特征、第二金字塔特征、第三金字塔特征和第四金字塔特征；6. method according to claim 5, it is characterized in that, the intermediate enhanced figure of described each size comprises the first pyramid feature, the second pyramid feature, the 3rd pyramid feature and the 4th pyramid feature;所述通过所述特征金字塔网络分别对相应尺寸的特征图进行上采样处理，得到各尺寸的中间增强图包括：The feature map of the corresponding size is respectively upsampled through the feature pyramid network, and the intermediate enhanced map of each size is obtained including:通过所述特征金字塔网络的第一金字塔层，对所述第四尺寸的特征图进行上采样处理，得到第一金字塔特征；Through the first pyramid layer of the feature pyramid network, the feature map of the fourth size is upsampled to obtain the first pyramid feature;通过所述特征金字塔网络的第二金字塔层，对所述第三尺寸的特征图和所述第一金字塔特征之间的融合特征进行上采样处理，得到第二金字塔特征；Through the second pyramid layer of the feature pyramid network, the fusion feature between the feature map of the third size and the first pyramid feature is upsampled to obtain the second pyramid feature;通过所述特征金字塔网络的第三金字塔层，对所述第二尺寸的特征图和所述第二金字塔特征之间的融合特征进行上采样处理，得到第三金字塔特征；Through the third pyramid layer of the feature pyramid network, the fusion feature between the feature map of the second size and the second pyramid feature is upsampled to obtain the third pyramid feature;通过所述特征金字塔网络的第四金字塔层，对所述第一尺寸的特征图和所述第三金字塔特征之间的融合特征进行上采样处理，得到第四金字塔特征。Through the fourth pyramid layer of the feature pyramid network, the fusion feature between the feature map of the first size and the feature of the third pyramid is upsampled to obtain the fourth pyramid feature.7.根据权利要求5所述的方法，其特征在于，所述各尺寸的增强特征图包括第一路径聚合特征、第二路径聚合特征、第三路径聚合特征和第四路径聚合特征；所述通过所述路径聚合网络分别对相应尺寸的中间增强图进行下采样处理，得到各尺寸的增强特征图包括：7. The method according to claim 5, wherein the enhanced feature map of each size comprises a first path aggregation feature, a second path aggregation feature, a third path aggregation feature and a fourth path aggregation feature; The intermediate enhanced maps of corresponding sizes are respectively down-sampled through the path aggregation network, and the enhanced feature maps of each size obtained include:通过所述路径聚合网络的第一路径聚合层，对所述第四金字塔特征进行下采样处理，得到所述第一路径聚合特征；Through the first path aggregation layer of the path aggregation network, the fourth pyramid feature is down-sampled to obtain the first path aggregation feature;通过所述路径聚合网络的第二路径聚合层，对所述第三金字塔特征和所述第一路径聚合特征之间的融合特征进行下采样处理，得到所述第二路径聚合特征；Through the second path aggregation layer of the path aggregation network, the fusion feature between the third pyramid feature and the first path aggregation feature is down-sampled to obtain the second path aggregation feature;通过所述路径聚合网络的第三路径聚合层，对所述第二金字塔特征和所述第二路径聚合特征之间的融合特征进行下采样处理，得到所述第三路径聚合特征；Through the third path aggregation layer of the path aggregation network, the fusion feature between the second pyramid feature and the second path aggregation feature is down-sampled to obtain the third path aggregation feature;通过所述路径聚合网络的第四路径聚合层，对所述第一金字塔特征和所述第三路径聚合特征之间的融合特征进行下采样处理，得到所述第四路径聚合特征。Through the fourth path aggregation layer of the path aggregation network, the fusion feature between the first pyramid feature and the third path aggregation feature is down-sampled to obtain the fourth path aggregation feature.8.一种确定病灶位置装置，其特征在于，所述装置包括：8. A device for determining the location of a lesion, characterized in that the device comprises:输入模块，用于将磁共振图像输入至检测模型；所述检测模型包括主干网络、特征融合网络和基于无锚点检测框架的解耦头；The input module is used to input the magnetic resonance image into the detection model; the detection model includes a backbone network, a feature fusion network and a decoupling head based on an anchor-free detection framework;特征提取模块，用于通过所述主干网络中的各特征提取模块对磁共振图像进行特征提取，得到不同尺寸的特征图；The feature extraction module is used to extract the features of the magnetic resonance image through each feature extraction module in the backbone network to obtain feature maps of different sizes;增强模块，用于通过所述特征融合网络分别对相应尺寸的特征图进行语义增强处理和位置增强处理，得到各尺寸的增强特征图；An enhancement module, configured to perform semantic enhancement processing and position enhancement processing on feature maps of corresponding sizes through the feature fusion network to obtain enhanced feature maps of each size;分类与回归模块，用于基于所述解耦头对各尺寸的所述增强特征图做分类和回归处理，得到各所述增强特征图对应的处理结果；A classification and regression module, configured to perform classification and regression processing on the enhanced feature maps of each size based on the decoupling head, to obtain processing results corresponding to each of the enhanced feature maps;解码模块，用于对各所述处理结果进行解码处理，并基于解码的结果确定所述磁共振图像中的病灶位置。The decoding module is configured to perform decoding processing on each of the processing results, and determine the location of the lesion in the magnetic resonance image based on the decoding results.9.一种计算机设备，包括存储器和处理器，所述存储器存储有计算机程序，其特征在于，所述处理器执行所述计算机程序时实现权利要求1至7中任一项所述的方法的步骤。9. A computer device comprising a memory and a processor, the memory stores a computer program, wherein the processor implements the method according to any one of claims 1 to 7 when executing the computer program step.10.一种计算机可读存储介质，其上存储有计算机程序，其特征在于，所述计算机程序被处理器执行时实现权利要求1至7中任一项所述的方法的步骤。10. A computer-readable storage medium, on which a computer program is stored, characterized in that, when the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images