技术领域technical field
本发明涉及医学图像分析领域及机器学习领域，特别涉及一种超声影像慢性肾脏病辅助筛查方法，属于基于深度学习的医学影像分析领域。The invention relates to the field of medical image analysis and the field of machine learning, in particular to an ultrasonic imaging auxiliary screening method for chronic kidney disease, and belongs to the field of deep learning-based medical image analysis.
背景技术Background technique
慢性肾脏病(Chronic Kidney Disease,CKD)是一种较为常见的泌尿系统疾病。由于该病因隐匿，无明显早期症状，造成低知晓率、早期诊断率低和早期干预率低。一旦发病，通常难以痊愈，如果得不到有效治疗最终易进入尿毒症期。目前CKD尚无强有效的治疗手段，早发现、早治疗、早干预，对延缓病情进程很重要。超声作为一种无创性的检查方法能够显示肾脏的形态、结构，是现阶段肾脏疾病的首选影像检查方法。医生通过动态地观察肾脏的大小、形态、边缘等来对病情做出判断。Chronic Kidney Disease (CKD) is a relatively common urinary system disease. Due to the hidden etiology and no obvious early symptoms, the low awareness rate, early diagnosis rate and early intervention rate are low. Once the disease occurs, it is usually difficult to recover, and if there is no effective treatment, it is easy to enter the uremia stage. At present, there is no strong and effective treatment for CKD. Early detection, early treatment, and early intervention are very important to delay the progression of the disease. Ultrasound, as a non-invasive examination method, can display the shape and structure of the kidney, and is the preferred imaging method for renal disease at this stage. Doctors make judgments about the condition by dynamically observing the size, shape, edge, etc. of the kidneys.
在基于超声的慢性肾病诊断过程中，临床医生的诊断率和自身的工作经验关系重大。在部分地区，由于条件限制，从事慢性肾病筛查的医生有限。近年来，计算机视觉领域的深度学习技术日渐成熟，在医学影像的自动分析及辅助智能诊断方面展现出巨大的应用潜力。其中卷积神经网络(CNN)在自然图像分类、目标检测等研究领域获得了广泛应用。而目前并没有成熟的利用超声影像对慢性肾脏病进行判断并分期的算法模型，利用深度学习技术对基于超声影像的慢性肾脏病的诊断研究较少。此外，肾脏超声影像相比自然图像，色彩单一、边缘模糊，与其他人体组织纹理相似度高，普遍存在伪影噪声。这些不利因素使得观测难度高，难以使用统计方法获得有效特征，加上个体差异性大，使得传统方法的筛查诊断难度大。In the diagnosis of chronic kidney disease based on ultrasound, the diagnosis rate of clinicians is closely related to their own work experience. In some areas, doctors who are engaged in chronic kidney disease screening are limited due to limitations. In recent years, the deep learning technology in the field of computer vision has become more and more mature, and it has shown great application potential in the automatic analysis of medical images and auxiliary intelligent diagnosis. Among them, convolutional neural network (CNN) has been widely used in natural image classification, object detection and other research fields. At present, there is no mature algorithm model for judging and staging chronic kidney disease using ultrasound images, and there are few researches on the diagnosis of chronic kidney disease based on ultrasound images using deep learning technology. In addition, compared with natural images, kidney ultrasound images have single color, blurred edges, and high texture similarity with other human tissues, and artifact noise is common. These unfavorable factors make observation difficult, and it is difficult to use statistical methods to obtain effective characteristics. In addition, the large individual differences make the screening and diagnosis of traditional methods difficult.
发明内容SUMMARY OF THE INVENTION
为了克服现有超声影像慢性肾脏病筛查方式难度大、效率低下、精度较低的不足，本发明提出了一种速度快、效率高、精度较高的基于深度学习的超声影像慢性肾脏病辅助筛查方法，实现了肾脏超声影像的自动分析，可有效对超声影像慢性肾脏病患病情况进行辅助筛查判断。In order to overcome the shortcomings of the existing ultrasonic imaging chronic kidney disease screening methods, such as difficulty, low efficiency and low precision, the present invention proposes a deep learning-based ultrasonic imaging chronic kidney disease assistant with high speed, high efficiency and high precision. The screening method realizes the automatic analysis of renal ultrasound images, and can effectively assist in the screening and judgment of the prevalence of chronic kidney disease in ultrasound images.
为了解决其技术问题本发明所采用的技术方案是：In order to solve its technical problem, the technical scheme adopted by the present invention is:
一种基于深度学习的超声影像慢性肾脏病辅助筛查方法，包括以下步骤：A deep learning-based ultrasound imaging-assisted screening method for chronic kidney disease, comprising the following steps:
步骤一，对原始超声影像进行中心裁剪，保留超声影像区的主体部分，去除图像周围包含无关文字信息的区域，并进行RGB图像灰度转换得到单通道灰度影像；Step 1, performing center cropping on the original ultrasound image, retaining the main part of the ultrasound image area, removing the area containing irrelevant text information around the image, and performing RGB image grayscale conversion to obtain a single-channel grayscale image;
步骤二，对步骤一中得到的影像设置合适阈值进行二值化处理得到标签图像，根据标签图像对肾脏部位进行粗定位，获得粗定位肾脏影像；Step 2, set an appropriate threshold for the image obtained in step 1 and perform binarization processing to obtain a label image, and perform coarse positioning of the kidney part according to the label image to obtain a coarsely positioned kidney image;
步骤三，去除影像中的人工标记，并对其进行修复；Step 3, remove the artificial mark in the image and repair it;
步骤四，通过直方图均衡化方法增强影像对比度，缩放影像将尺寸统一为224*224，并将数据归一化为[0,1]；Step 4: Enhance the image contrast through the histogram equalization method, scale the image to unify the size to 224*224, and normalize the data to [0,1];
步骤五，构建卷积神经网络，并利用影像训练集进行训练。得到能够筛查是否患有慢性肾脏病的二分类神经网络；Step 5: Build a convolutional neural network and use the image training set for training. Obtain a binary neural network capable of screening for chronic kidney disease;
步骤六，利用训练得到的卷积神经网络对超声影像慢性肾脏病进行辅助筛查。Step 6, using the trained convolutional neural network to perform auxiliary screening for chronic kidney disease in ultrasound images.
进一步，所述步骤二中，获取肾脏部位粗定位的过程为：对高于阈值的像素点标记为255，其余标记为0，得到二值化标签图像。对其进行形态学闭运算操作，减少独立噪声，填充连通区域的孔洞。之后提取图像中各连通域的外轮廓，以面积最大的区域计算正外接矩形，根据该矩形的坐标在步骤一中得到的影像中截取得到粗定位的肾脏部位超声影像。Further, in the second step, the process of obtaining the coarse location of the kidney part is as follows: the pixels higher than the threshold are marked as 255, and the rest are marked as 0 to obtain a binarized label image. Perform morphological closure operations on it to reduce independent noise and fill holes in connected regions. After that, the outer contour of each connected domain in the image is extracted, the area with the largest area is used to calculate the right circumscribed rectangle, and the roughly positioned ultrasound image of the kidney is obtained by intercepting the image obtained in step 1 according to the coordinates of the rectangle.
再进一步，步骤二中，对于人工标记了肾脏区域的影像，可根据标记亮度较高的特点设置阈值，约为最大灰度值的85％，根据标记来获得更精确的肾脏定位影像；对于未标记肾脏区域的影像，阈值设置约为最大灰度值的40％。Still further, in step 2, for the images with the kidney area manually marked, a threshold value can be set according to the feature of high brightness of the marking, which is about 85% of the maximum gray value, and a more accurate kidney positioning image can be obtained according to the marking; An image of the kidney region was marked with a threshold set at approximately 40% of the maximum grayscale value.
更进一步，所述步骤三中，影像中的人工标记会遮挡部分纹理，影响待分析影像的完整性，去除影像中人工标记并进行修复的过程为：Further, in the third step, the artificial mark in the image will block part of the texture and affect the integrity of the image to be analyzed. The process of removing the artificial mark in the image and repairing is as follows:
步骤3.1根据超声影像人工标记偏黄色的特点，即R,G通道数值远高于B通道数置，在原始影像中筛选疑似人工标记的像素区域，进行标签化得到疑似标记区域；Step 3.1 According to the characteristic that the manual marking of ultrasound images is yellowish, that is, the value of R and G channels is much higher than the number of B channels, screen the pixel area suspected of manual marking in the original image, and carry out labeling to obtain the suspected marked area;
步骤3.2在粗定位影像中将疑似标记区域的像素点数值置为0，去除人工标记；Step 3.2 Set the pixel value of the suspected marked area to 0 in the coarse positioning image, and remove the artificial mark;
步骤3.3对去除标记后的区域进行修复，假设待修复区域的边缘某点为p，其近邻区域内某已知像素点为q，q为p点提供的近似值，计算公式为：Step 3.3 Repair the area after removing the mark. Suppose a point on the edge of the area to be repaired is p, and a known pixel point in the adjacent area is q, and q is the approximate value provided by point p. The calculation formula is:
所述步骤五中，构建卷积神经网络的过程为：In the step 5, the process of constructing the convolutional neural network is:
步骤5.1输入一组大小为224*224*1的肾脏超声影像；Step 5.1 Input a set of kidney ultrasound images with a size of 224*224*1;
步骤5.2先经过3*3大小的卷积操作提取特征，再经过2*2大小的最大池化，重复两次；Step 5.2 First extract features through a 3*3 convolution operation, and then go through a 2*2 maximum pooling, repeat twice;
步骤5.3经过多尺度残差模块Block模块提取特征，该模块具有三个分支，分别为1*1，3*3，5*5大小的卷积核，分别提取后进行特征拼接，再经过3*3大小的卷积操作再次提取特征。Step 5.3 Extract features through the multi-scale residual module Block module. This module has three branches, which are convolution kernels of 1*1, 3*3, and 5*5 sizes. After extraction, feature splicing is performed, and then 3* A convolution operation of size 3 extracts features again.
步骤5.4重复三次步骤5.3，在经过7*7大小的全局平均池化；Step 5.4 Repeat step 5.3 three times, after the global average pooling of size 7*7;
步骤5.5经过全连接层fc1，dropout层，再经过全连接层fc2，输出二分类筛查结果。Step 5.5 Pass through the fully connected layer fc1, the dropout layer, and then through the fully connected layer fc2, and output the two-category screening results.
本发明基于超声影像对慢性肾脏病进行早期智能辅助筛查，利用卷积神经网络提取影像特征，来实现对慢性肾脏病患病情况的判断。与现有方法相比，其技术优势在于：The invention performs early intelligent auxiliary screening for chronic kidney disease based on ultrasonic images, and uses convolutional neural network to extract image features to realize the judgment of chronic kidney disease. Compared with existing methods, its technical advantages are:
1.通过卷积神经网络对超声影像进行分析，自动辅助筛查判断慢性肾脏病情况，相比传统方法筛查效率高，速度快。1. Convolutional neural network is used to analyze ultrasound images and automatically assist screening to judge chronic kidney disease. Compared with traditional methods, screening is more efficient and faster.
2.在超声影像的预处理上，方法引入了人工标记去除及修复方法，避免了人工标记遮挡图像中的纹理部分，提高待分析区域图像的完整性。2. In the preprocessing of ultrasound images, the method introduces manual mark removal and repair methods, which avoids manual marks blocking the texture part of the image and improves the integrity of the image in the area to be analyzed.
3.通过引入多尺度残差卷积模块结构，分别使用不同尺寸的卷积核来提取不同粒度的卷积特征图，获取更为丰富的肾脏影像特征。同时残差快捷支路的存在降低了网络的训练难度。3. By introducing a multi-scale residual convolution module structure, convolution kernels of different sizes are used to extract convolution feature maps of different granularities, and more abundant kidney image features are obtained. At the same time, the existence of the residual shortcut branch reduces the training difficulty of the network.
附图说明Description of drawings
图1基于深度学习的超声影像慢性肾脏病辅助筛查方法的流程图。Figure 1. Flow chart of the deep learning-based ultrasound imaging-assisted screening method for chronic kidney disease.
图2用于超声影像慢性肾脏病筛查的神经网络结构示意图。Figure 2 is a schematic diagram of the neural network structure used for ultrasound imaging for chronic kidney disease screening.
图3神经网络中Block模块结构示意图。Figure 3 is a schematic diagram of the block module structure in the neural network.
具体实施方式Detailed ways
下面结合附图对本发明作进一步描述。The present invention will be further described below in conjunction with the accompanying drawings.
如图1所示，本发明为一种基于深度学习的超声影像慢性肾脏病辅助筛查方法，包含以下步骤：As shown in FIG. 1 , the present invention is a deep learning-based ultrasound imaging assisted screening method for chronic kidney disease, comprising the following steps:
步骤一，对原始超声影像进行中心裁剪，保留超声影像区的主体部分，去除图像周围包含无关文字信息的区域，并进行RGB图像灰度转换得到单通道灰度影像；Step 1, performing center cropping on the original ultrasound image, retaining the main part of the ultrasound image area, removing the area containing irrelevant text information around the image, and performing RGB image grayscale conversion to obtain a single-channel grayscale image;
步骤二，对步骤一中得到的影像设置合适阈值进行二值化处理得到标签图像，根据标签图像对肾脏部位进行粗定位，获得粗定位肾脏影像；Step 2, set an appropriate threshold for the image obtained in step 1 and perform binarization processing to obtain a label image, and perform coarse positioning of the kidney part according to the label image to obtain a coarsely positioned kidney image;
步骤三，去除影像师在影像中所做的人工标记，并进行修复；Step 3: Remove the artificial marks made by the videographer in the image and repair it;
步骤四，通过直方图均衡化方法增强影像对比度，缩放影像将尺寸统一为224*224，并将数据归一化为[0,1]；Step 4: Enhance the image contrast through the histogram equalization method, scale the image to unify the size to 224*224, and normalize the data to [0,1];
步骤五，构建卷积神经网络，并利用影像训练集进行训练。得到能够筛查是否患有慢性肾脏病的二分类神经网络；Step 5: Build a convolutional neural network and use the image training set for training. Obtain a binary neural network capable of screening for chronic kidney disease;
步骤六，利用训练得到的卷积神经网络对超声影像慢性肾脏病进行筛查。Step 6, using the trained convolutional neural network to screen chronic kidney disease in ultrasound images.
进一步，所述步骤二中，获取肾脏部位粗定位的过程为：对高于阈值的像素点标记为255，其余标记为0，得到二值化标签图像。对其进行形态学闭运算操作，减少独立噪声，填充连通区域的孔洞。之后提取图像中各连通域的外轮廓，以面积最大的区域计算正外接矩形，根据该矩形的坐标在步骤一中得到的影像中截取得到粗定位的肾脏部位超声影像。Further, in the second step, the process of obtaining the coarse location of the kidney part is as follows: the pixels higher than the threshold are marked as 255, and the rest are marked as 0 to obtain a binarized label image. Perform morphological closure operations on it to reduce independent noise and fill holes in connected regions. After that, the outer contour of each connected domain in the image is extracted, the area with the largest area is used to calculate the right circumscribed rectangle, and the roughly positioned ultrasound image of the kidney is obtained by intercepting the image obtained in step 1 according to the coordinates of the rectangle.
再进一步，步骤二中，对于人工标记了肾脏区域的影像，可根据标记亮度较高的特点设置阈值，约为最大灰度值的85％，根据标记来获得更精确的肾脏定位影像；对于未标记肾脏区域的影像，阈值设置约为最大灰度值的40％。Still further, in step 2, for the images with the kidney area manually marked, a threshold value can be set according to the feature of high brightness of the marking, which is about 85% of the maximum gray value, and a more accurate kidney positioning image can be obtained according to the marking; An image of the kidney region was marked with a threshold set at approximately 40% of the maximum grayscale value.
更进一步，所述步骤三中，影像中的人工标记会遮挡部分纹理，影响待分析影像的完整性，去除影像中人工标记并进行修复的过程为：Further, in the third step, the artificial mark in the image will block part of the texture and affect the integrity of the image to be analyzed. The process of removing the artificial mark in the image and repairing is as follows:
步骤3.1根据超声影像人工标记偏黄色的特点，即R,G通道数值远高于B通道数置，在原始影像中筛选疑似人工标记的像素区域，进行标签化得到疑似标记区域；Step 3.1 According to the characteristic that the manual marking of ultrasound images is yellowish, that is, the value of R and G channels is much higher than the number of B channels, screen the pixel area suspected of manual marking in the original image, and carry out labeling to obtain the suspected marked area;
步骤3.2在粗定位影像中将疑似标记区域的像素点数值置为0，去除人工标记；Step 3.2 Set the pixel value of the suspected marked area to 0 in the coarse positioning image, and remove the artificial mark;
步骤3.3对去除标记后的区域进行修复，假设待修复区域的边缘某点为p，其近邻区域内某已知像素点为q，q为p点提供的近似值，计算公式为：Step 3.3 Repair the area after removing the mark. Suppose a point on the edge of the area to be repaired is p, and a known pixel point in the adjacent area is q, and q is the approximate value provided by point p. The calculation formula is:
所述步骤五中，构建卷积神经网络的过程为：In the step 5, the process of constructing the convolutional neural network is:
步骤5.1输入一组大小为224*224*1的肾脏超声影像；Step 5.1 Input a set of kidney ultrasound images with a size of 224*224*1;
步骤5.2先经过3*3大小的卷积操作提取特征，再经过2*2大小的最大池化，重复两次；Step 5.2 First extract features through a 3*3 convolution operation, and then go through a 2*2 maximum pooling, repeat twice;
步骤5.3经过多尺度残差模块Block模块提取特征，该模块具有三个分支，分别为1*1，3*3，5*5大小的卷积核，分别提取后进行特征拼接，再经过3*3大小的卷积操作再次提取特征。Step 5.3 Extract features through the multi-scale residual module Block module. This module has three branches, which are convolution kernels of 1*1, 3*3, and 5*5 sizes. After extraction, feature splicing is performed, and then 3* A convolution operation of size 3 extracts features again.
步骤5.4重复三次步骤5.3，在经过7*7大小的全局平均池化；Step 5.4 Repeat step 5.3 three times, after the global average pooling of size 7*7;
步骤5.5经过全连接层fc1，dropout层，再经过全连接层fc2，输出二分类筛查结果。Step 5.5 Pass through the fully connected layer fc1, the dropout layer, and then through the fully connected layer fc2, and output the two-category screening results.
所述步骤五中，卷积神经网络网络架构主要由2个卷积层，2个最大池化层，3个多尺度残差卷积模块(Block模块)，1个全局池化层，2个全连接层及1个dropout层组成。每个卷积层后都对特征进行批标准化操作，提高训练速度，并经过ReLU激活函数，提升网络的非线性表达。如图3所示，每个多尺度残差卷积模块(Block模块)内包含2个卷积层，一条快捷支路。Block模块的输入输出特征大小保持不变。其中第1个卷积层包含3个不同卷积核尺度的分支，分别为1*1，3*3，5*5大小，它们对输入进行特征提取，最后通过通道堆叠进行特征拼接。第2个卷积层包含1个3*3大小的卷积核；快捷支路起点为输入，终点为第二个卷积层后的加操作。使得输入特征可以直接和第二个卷积层提取的特征进行数值相加。这样网络只需要计算相比原输入的残差，降低了训练难度。dropout层随机使部分神经元失活，在数据量不大的情况下可以缓解过拟合。本案例中比例设置为50％。最后一个全连接层的输出尺寸为1，对应筛查结果的二分类(患病或健康)。In the fifth step, the convolutional neural network architecture mainly consists of 2 convolutional layers, 2 maximum pooling layers, 3 multi-scale residual convolution modules (Block modules), 1 global pooling layer, 2 It consists of a fully connected layer and a dropout layer. After each convolutional layer, batch normalization is performed on the features to improve the training speed, and through the ReLU activation function, the nonlinear expression of the network is improved. As shown in Figure 3, each multi-scale residual convolution module (Block module) contains 2 convolution layers and a shortcut branch. The input and output feature sizes of the Block module remain unchanged. The first convolutional layer contains three branches with different convolution kernel scales, which are 1*1, 3*3, and 5*5. They extract features from the input, and finally perform feature splicing through channel stacking. The second convolutional layer contains a 3*3 convolution kernel; the starting point of the shortcut branch is the input, and the end point is the addition operation after the second convolutional layer. This allows the input features to be directly numerically added to the features extracted by the second convolutional layer. In this way, the network only needs to calculate the residual error compared to the original input, which reduces the difficulty of training. The dropout layer randomly deactivates some neurons, which can alleviate overfitting when the amount of data is not large. The ratio is set to 50% in this case. The output size of the last fully-connected layer is 1, corresponding to the binary classification (sick or healthy) of the screening results.
实例：本案例中使用的肾脏超声影像共2类，即患有慢性肾脏病或健康。由于采集的影像数据量有限，数据预处理后，通过粗定位区域的平移进行数据扩大，得到肾脏超声影像共828例样本，健康样本与慢性肾脏病样本相同数量。从样本中随机选取628个样本作为训练集，100个样本作为验证集，100个样本作为测试集。下面具体介绍影像去除人工标记，模型的训练和测试过程。Example: There are 2 types of renal ultrasound images used in this case, namely with chronic kidney disease or healthy. Due to the limited amount of image data collected, after data preprocessing, the data was expanded by the translation of the coarse positioning area, and a total of 828 samples of renal ultrasound images were obtained, and the number of healthy samples was the same as that of chronic kidney disease samples. 628 samples were randomly selected from the samples as the training set, 100 samples as the validation set, and 100 samples as the test set. The following is a detailed introduction to the removal of artificial labels from images, the training and testing process of the model.
步骤一，影像去除人工标记并修复。In step 1, the images are manually marked and repaired.
步骤1.1根据超声影像人工标记偏黄色的特点，即R,G通道数值远高于B通道数置，如RGB范围为[90,90,0]～[255,255,10]，在原始影像中筛选疑似人工标记的像素区域，进行标签化得到疑似标记区域；Step 1.1 According to the characteristic that the manual marking of ultrasound images is yellowish, that is, the value of R and G channels is much higher than the number of B channels, for example, the RGB range is [90,90,0]～[255,255,10], screen the original image for suspected The artificially marked pixel area is labeled to obtain the suspected marked area;
步骤1.2在粗定位影像中将疑似标记区域的像素点数值置为0，去除人工标记；Step 1.2 In the coarse positioning image, set the pixel value of the suspected marked area to 0, and remove the artificial mark;
步骤1.3对去除标记后的区域利用快速行进法进行修复，假设待修复区域的边缘某点为p，其近邻区域内某已知像素点为q，q为p点提供的近似值公式为：综合其有限小范围邻域内所有像素点的值，最终确定p点的修复值。Step 1.3 Use the fast marching method to repair the area after removing the mark. Suppose a point on the edge of the area to be repaired is p, a known pixel point in the neighboring area is q, and q is the approximate value formula provided by the p point: The repair value of point p is finally determined by synthesizing the values of all pixel points in its finite and small-scale neighborhood.
步骤二，神经网络的构建和训练，具体结构如图2所示。Step 2, the construction and training of the neural network, the specific structure is shown in Figure 2.
步骤2.1网络架构主要由2个卷积层，2个最大池化层，3个多尺度残差卷积模块(Block模块)，1个全局池化层，2个全连接层及1个dropout层组成。Step 2.1 The network architecture mainly consists of 2 convolution layers, 2 max pooling layers, 3 multi-scale residual convolution modules (Block modules), 1 global pooling layer, 2 fully connected layers and 1 dropout layer composition.
步骤2.2卷积层中卷积核大小均为3*3，滑动步长为1，padding为1。每个卷积层后都对特征进行批标准化操作，提高训练速度，并经过ReLU激活函数，提升网络的非线性表达。卷积层后的最大池化层将缩减特征图尺寸。Step 2.2 In the convolutional layer, the size of the convolution kernel is 3*3, the sliding step size is 1, and the padding is 1. After each convolutional layer, batch normalization is performed on the features to improve the training speed, and through the ReLU activation function, the nonlinear expression of the network is improved. The max pooling layer after the convolutional layer will reduce the feature map size.
步骤2.3卷积层中所有参数权重初始化为随机正交矩阵初始化，权重正则化方式为L2正则，偏置值初始化为0。Step 2.3 All parameter weights in the convolutional layer are initialized to random orthogonal matrix initialization, the weight regularization method is L2 regularization, and the bias value is initialized to 0.
步骤2.4全连接层中，权重初始化为随机正态分布，权重正则化方式为L2正则，偏置值初始化为0；dropout层随机失活比例设置为50％。Step 2.4 In the fully connected layer, the weight is initialized to a random normal distribution, the weight regularization method is L2 regularity, and the bias value is initialized to 0; the random deactivation ratio of the dropout layer is set to 50%.
步骤2.5本实例采用python语言编程，利用pytorch框架搭建网络。模型采用batch训练的方式，通过随机梯度下降法进行训练。训练集生成器和验证集生成器每个batch的样本数batch size均为16，训练集生成器每返回40次数据作为一个轮次(epoch)，一个轮次训练完成后，生成器会返回5次并计算验证集损失，损失函数为交叉熵损失函数。模型优化器为Adam，初始学习率为0.01，每10个轮次降低二分之一。模型最大训练轮次为100，验证与训练损失收敛后停止训练，并保存模型为作为最终训练结果。Step 2.5 This example uses python language programming and uses the pytorch framework to build a network. The model adopts the batch training method and is trained by the stochastic gradient descent method. The number of samples in each batch of the training set generator and the validation set generator is 16. The training set generator returns 40 times of data as an epoch. After an epoch of training is completed, the generator will return 5 Second and calculate the loss of the validation set, the loss function is the cross entropy loss function. The model optimizer is Adam, and the initial learning rate is 0.01, which is reduced by half every 10 epochs. The maximum number of training rounds of the model is 100. After the validation and training losses converge, the training is stopped, and the model is saved as the final training result.
步骤三，神经网络模型测试Step 3, Neural Network Model Test
载入模型，将预处理完毕的影像测试集样本输入模型分析，将筛查结果与其标签对比得到模型的筛查准确率。Load the model, input the preprocessed image test set samples into the model analysis, and compare the screening results with their labels to obtain the screening accuracy of the model.
经过上述步骤的操作，即可实现用于筛查超神影像慢性肾脏病的卷积神经网络的构建、训练与测试。After the above steps, the construction, training and testing of a convolutional neural network for screening chronic kidney disease with super-imaging can be realized.
以上所述的具体描述，对发明的目的、技术方案和有益效果进行了进一步详细说明，所应理解的是，以上所述仅为本发明的具体实施例，用于解释本发明，并不用于限定本发明的保护范围，凡在本发明的精神和原则之内，所做的任何修改、等同替换、改进等，均应包含在本发明的保护范围之内。The above-mentioned specific description further describes the purpose, technical solutions and beneficial effects of the invention in detail. It should be understood that the above-mentioned descriptions are only specific embodiments of the present invention, which are used to explain the present invention and are not intended to be used for The protection scope of the present invention is limited, and any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910160852.XACN109961838A (en) | 2019-03-04 | 2019-03-04 | A deep learning-based ultrasound imaging-assisted screening method for chronic kidney disease |
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910160852.XACN109961838A (en) | 2019-03-04 | 2019-03-04 | A deep learning-based ultrasound imaging-assisted screening method for chronic kidney disease |
| Publication Number | Publication Date |
|---|---|
| CN109961838Atrue CN109961838A (en) | 2019-07-02 |
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910160852.XAPendingCN109961838A (en) | 2019-03-04 | 2019-03-04 | A deep learning-based ultrasound imaging-assisted screening method for chronic kidney disease |
| Country | Link |
|---|---|
| CN (1) | CN109961838A (en) |
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110706793A (en)* | 2019-09-25 | 2020-01-17 | 天津大学 | Attention mechanism-based thyroid nodule semi-supervised segmentation method |
| CN110852396A (en)* | 2019-11-15 | 2020-02-28 | 苏州中科华影健康科技有限公司 | Sample data processing method for cervical image |
| CN111209916A (en)* | 2019-12-31 | 2020-05-29 | 中国科学技术大学 | Focus identification method and system and identification equipment |
| CN111260641A (en)* | 2020-01-21 | 2020-06-09 | 珠海威泓医疗科技有限公司 | A handheld ultrasound imaging system and method based on artificial intelligence |
| CN111724314A (en)* | 2020-05-08 | 2020-09-29 | 天津大学 | A method for detection and removal of special markers in medical images |
| CN112529835A (en)* | 2020-10-22 | 2021-03-19 | 浙江大学 | Image splicing tampering detection and positioning method based on source camera identification |
| CN113487628A (en)* | 2021-07-07 | 2021-10-08 | 广州市大道医疗科技有限公司 | Model training method, coronary vessel identification method, device, equipment and medium |
| CN118396963A (en)* | 2024-05-08 | 2024-07-26 | 首都医科大学附属北京积水潭医院 | Liver and spleen morphology analysis method based on CT image and deep learning |
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106056595A (en)* | 2015-11-30 | 2016-10-26 | 浙江德尚韵兴图像科技有限公司 | Method for automatically identifying whether thyroid nodule is benign or malignant based on deep convolutional neural network |
| CN107292333A (en)* | 2017-06-05 | 2017-10-24 | 浙江工业大学 | A kind of rapid image categorization method based on deep learning |
| US20180177461A1 (en)* | 2016-12-22 | 2018-06-28 | The Johns Hopkins University | Machine learning approach to beamforming |
| CN108257135A (en)* | 2018-02-01 | 2018-07-06 | 浙江德尚韵兴图像科技有限公司 | The assistant diagnosis system of medical image features is understood based on deep learning method |
| CN108520518A (en)* | 2018-04-10 | 2018-09-11 | 复旦大学附属肿瘤医院 | A method and device for ultrasonic image recognition of thyroid tumors |
| CN109063712A (en)* | 2018-06-22 | 2018-12-21 | 哈尔滨工业大学 | An intelligent diagnosis method and system for multi-model diffuse liver diseases based on ultrasound images |
| CN109242791A (en)* | 2018-08-22 | 2019-01-18 | 东北农业大学 | A kind of batch restorative procedure for damaged plant leaf blade |
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106056595A (en)* | 2015-11-30 | 2016-10-26 | 浙江德尚韵兴图像科技有限公司 | Method for automatically identifying whether thyroid nodule is benign or malignant based on deep convolutional neural network |
| US20180177461A1 (en)* | 2016-12-22 | 2018-06-28 | The Johns Hopkins University | Machine learning approach to beamforming |
| CN107292333A (en)* | 2017-06-05 | 2017-10-24 | 浙江工业大学 | A kind of rapid image categorization method based on deep learning |
| CN108257135A (en)* | 2018-02-01 | 2018-07-06 | 浙江德尚韵兴图像科技有限公司 | The assistant diagnosis system of medical image features is understood based on deep learning method |
| CN108520518A (en)* | 2018-04-10 | 2018-09-11 | 复旦大学附属肿瘤医院 | A method and device for ultrasonic image recognition of thyroid tumors |
| CN109063712A (en)* | 2018-06-22 | 2018-12-21 | 哈尔滨工业大学 | An intelligent diagnosis method and system for multi-model diffuse liver diseases based on ultrasound images |
| CN109242791A (en)* | 2018-08-22 | 2019-01-18 | 东北农业大学 | A kind of batch restorative procedure for damaged plant leaf blade |
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110706793A (en)* | 2019-09-25 | 2020-01-17 | 天津大学 | Attention mechanism-based thyroid nodule semi-supervised segmentation method |
| CN110852396A (en)* | 2019-11-15 | 2020-02-28 | 苏州中科华影健康科技有限公司 | Sample data processing method for cervical image |
| CN111209916A (en)* | 2019-12-31 | 2020-05-29 | 中国科学技术大学 | Focus identification method and system and identification equipment |
| CN111209916B (en)* | 2019-12-31 | 2024-01-23 | 中国科学技术大学 | Focus identification method and system and identification equipment |
| CN111260641A (en)* | 2020-01-21 | 2020-06-09 | 珠海威泓医疗科技有限公司 | A handheld ultrasound imaging system and method based on artificial intelligence |
| CN111724314A (en)* | 2020-05-08 | 2020-09-29 | 天津大学 | A method for detection and removal of special markers in medical images |
| CN112529835A (en)* | 2020-10-22 | 2021-03-19 | 浙江大学 | Image splicing tampering detection and positioning method based on source camera identification |
| CN113487628A (en)* | 2021-07-07 | 2021-10-08 | 广州市大道医疗科技有限公司 | Model training method, coronary vessel identification method, device, equipment and medium |
| CN113487628B (en)* | 2021-07-07 | 2024-02-23 | 广州市大道医疗科技有限公司 | Model training method, coronary vessel identification method, device, equipment and medium |
| CN118396963A (en)* | 2024-05-08 | 2024-07-26 | 首都医科大学附属北京积水潭医院 | Liver and spleen morphology analysis method based on CT image and deep learning |
| Publication | Publication Date | Title |
|---|---|---|
| CN109961838A (en) | A deep learning-based ultrasound imaging-assisted screening method for chronic kidney disease | |
| CN109886986B (en) | Dermatoscope image segmentation method based on multi-branch convolutional neural network | |
| CN112070772B (en) | Blood leukocyte image segmentation method based on UNet++ and ResNet | |
| CN108665456B (en) | Method and system for real-time marking of breast ultrasound lesion region based on artificial intelligence | |
| Liu et al. | A framework of wound segmentation based on deep convolutional networks | |
| CN110335231A (en) | A method for assisted screening of chronic kidney disease with ultrasound imaging based on texture features and depth features | |
| CN108615051A (en) | Diabetic retina image classification method based on deep learning and system | |
| CN110490892A (en) | A kind of Thyroid ultrasound image tubercle automatic positioning recognition methods based on USFaster R-CNN | |
| CN107274386A (en) | A kind of cervical cell liquid-based smear artificial intelligence aids in diagosis system | |
| CN113610859B (en) | Automatic thyroid nodule segmentation method based on ultrasonic image | |
| Bai et al. | Automatic segmentation of cervical region in colposcopic images using K-means | |
| CN111476794B (en) | Cervical pathological tissue segmentation method based on UNET | |
| CN109670489B (en) | Weakly supervised early age-related macular degeneration classification method based on multi-instance learning | |
| CN117078697B (en) | Fundus disease seed detection method based on cascade model fusion | |
| CN109886346A (en) | A Cardiac MRI Image Classification System | |
| CN114300099A (en) | Allolymphocyte typing method based on YOLOv5 and microscopic hyperspectral image | |
| CN118116576B (en) | Intelligent case analysis method and system based on deep learning | |
| CN117934489A (en) | Fundus hard exudate segmentation method based on residual error and pyramid segmentation attention | |
| CN113160261A (en) | Boundary enhancement convolution neural network for OCT image corneal layer segmentation | |
| CN112102332A (en) | Cancer WSI segmentation method based on local classification neural network | |
| CN116596952A (en) | A multi-level lesion detection optimization method for pathological slice image segmentation and detection | |
| CN112233085A (en) | A segmentation method of cervical cell image based on pixel prediction enhancement | |
| CN112183448B (en) | A three-level classification and multi-scale FCN image segmentation method for depodding soybeans | |
| CN117036288A (en) | Tumor subtype diagnosis method for full-slice pathological image | |
| CN116844696A (en) | A facial image and traditional Chinese medicine constitution identification method based on deep neural network |
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication | Application publication date:20190702 |