CN114821176B - Viral encephalitis classification system for MR (magnetic resonance) images of children brain - Google Patents

Abstract

Translated fromChinese

一种儿童脑部MR图像病毒性脑炎分类系统，包括计算机存储器、计算机处理器以及存储在所述计算机存储器中并可在所述计算机处理器上执行的计算机程序，所述计算机存储器中存有训练好的分类模型，所述的分类模型采用改进的SE ResNet网络模型，共包含四个卷积部分，每个卷积部分由若干个子模块组构成，每个子模块组包含Inception子模块和SE Res子模块，最终通过全连接层获得最后的分类结果；所述计算机处理器执行所述计算机程序时实现以下步骤：将待分类的儿童脑部MR影像输入训练好的分类模型中，得到病毒性脑炎分类结果。利用本发明，可以提升模型对于不同维度特征的学习能力，大大提升儿童病毒性脑炎诊断的效率和准确率。

A system for classifying viral encephalitis in children's brain MR images, comprising a computer memory, a computer processor, and a computer program stored in the computer memory and executable on the computer processor, wherein the computer memory stores The trained classification model adopts the improved SE ResNet network model, which includes four convolution parts in total, each convolution part is composed of several sub-module groups, and each sub-module group includes Inception sub-module and SE Res The sub-module finally obtains the final classification result through the fully connected layer; when the computer processor executes the computer program, the following steps are implemented: input the MR image of the child's brain to be classified into the trained classification model, and obtain the viral brain inflammation classification results. By using the present invention, the learning ability of the model for features of different dimensions can be improved, and the efficiency and accuracy of the diagnosis of children's viral encephalitis can be greatly improved.

Description

Translated fromChinese

一种儿童脑部MR图像病毒性脑炎分类系统A classification system for viral encephalitis on children's brain MR images

技术领域technical field

本发明属于医学人工智能领域，尤其是涉及一种儿童脑部MR图像病毒性脑炎分类系统。The invention belongs to the field of medical artificial intelligence, in particular to a classification system for children's brain MR images of viral encephalitis.

背景技术Background technique

儿童脑炎是儿科一种较为常见的疾病。一般情况下，可通过临床症状、实验室检查，以及影像学和脑电图检测进行综合判断。若经确诊，患者需要在专业医生的指导下，进行针对性治疗。Childhood encephalitis is a relatively common disease in pediatrics. In general, a comprehensive judgment can be made through clinical symptoms, laboratory tests, and imaging and EEG tests. If diagnosed, patients need to be targeted for treatment under the guidance of professional doctors.

目前医生诊断主要通过临床症状、实验室检查(脑脊液检查)、影像学和脑电图检测等方法进行检查，然而临床症状不太准确；影像学和脑电图检测则只有重症时才能肉眼观察到病变区域；脑脊液检查较准确，但耗时较长，且需要抽取脑脊液，对儿童造成创伤和痛苦。At present, doctors diagnose mainly through clinical symptoms, laboratory tests (cerebrospinal fluid examination), imaging and EEG detection, etc. However, clinical symptoms are not very accurate; imaging and EEG detection can only be observed with the naked eye in severe cases Lesion area; cerebrospinal fluid examination is more accurate, but it takes a long time and needs to extract cerebrospinal fluid, which causes trauma and pain to children.

随着人工智能和深度学习的发展，在医学领域，很多研究者尝试使用智能算法对影像学数据脑电图数据进行自动识别。With the development of artificial intelligence and deep learning, in the medical field, many researchers try to use intelligent algorithms to automatically identify EEG data from imaging data.

如公开号为CN112561863A的中国专利文献公开了一种基于深度学习的粒细胞图片细粒度分类识别系统；包含定位模块和分类模块，其中定位模块利用Hourglass网络模型对输入的粒细胞图片进行特征提取，将粒细胞图片中的全部细胞分别进行定位，再将定位后的细胞裁剪出来，留下单个完整的细胞，并将全部裁剪出来的细胞进行尺寸归一化处理；分类模块采用构建的深度学习分类模型对定位模块定位出的粒细胞进行分类；辅助临床医生准确高效完成粒细胞分类识别计数任务，减小主观性带来的误差，减轻医生的工作量，辅助医生做出疾病判断；该系统能够有效解决非均衡数据下的细胞分类以及粒细胞间的细粒度分类，提升网络分类识别效果。For example, the Chinese patent document with the publication number CN112561863A discloses a fine-grained classification and recognition system for granulocyte pictures based on deep learning; it includes a positioning module and a classification module, wherein the positioning module uses the Hourglass network model to extract features from the input granulocyte pictures, Position all the cells in the granulocyte picture separately, then cut out the positioned cells, leaving a single complete cell, and normalize all the cut out cells in size; the classification module adopts the constructed deep learning classification The model classifies the granulocytes located by the positioning module; assists clinicians to accurately and efficiently complete the task of granulocyte classification, identification and counting, reduces the error caused by subjectivity, reduces the workload of doctors, and assists doctors in making disease judgments; the system can Effectively solve the cell classification under unbalanced data and the fine-grained classification between granulocytes, and improve the network classification recognition effect.

公开号为CN112132808A的中国专利文献公开了一种基于常态模型学习的乳腺X线图像病变检测方法和装置。所述方法包括从乳腺X线图像中分割出乳腺区域；提取图像块，进行亮度归一化处理；选取一部分正常区域图像块作为训练集，输入到双重深度卷积神经网络模型进行训练，得到常态模型；从训练集中选取若干正常区域图像块作为模板，输入到常态模型，得到模板图像的特征向量；将测试集输入到常态模型，得到测试图像的特征向量；将模板图像和测试图像的特征向量输入到最近邻分类器执行一类分类，得到测试结果。The Chinese patent document with publication number CN112132808A discloses a mammography image lesion detection method and device based on normal model learning. The method comprises the steps of segmenting a mammary gland area from a mammogram image; extracting image blocks, and performing brightness normalization processing; selecting a part of normal area image blocks as a training set, and inputting them into a double-depth convolutional neural network model for training to obtain normal model; select several normal area image blocks from the training set as templates, input them into the normal model, and obtain the feature vector of the template image; input the test set into the normal model, and obtain the feature vector of the test image; combine the template image and the feature vector of the test image Input to the nearest neighbor classifier to perform a class of classification, and get the test result.

然而，对于儿童病毒性脑炎而言，影像学数据的显示的特征不明显，使用常规的深度学习方法，很难对儿童是否患有病毒性脑炎进行准确的诊断。However, for children with viral encephalitis, the characteristics of imaging data are not obvious, and it is difficult to accurately diagnose whether children have viral encephalitis using conventional deep learning methods.

发明内容Contents of the invention

本发明提供了一种儿童脑部MR图像病毒性脑炎分类系统，可以在腰穿脑脊液检查、临床检查之外，只基于MR图像进行病毒性脑炎的诊断，具有较高的准确率。The invention provides a classification system for viral encephalitis in children's brain MR images, which can diagnose viral encephalitis based only on MR images in addition to lumbar puncture cerebrospinal fluid examination and clinical examination, and has a high accuracy rate.

一种儿童脑部MR图像病毒性脑炎分类系统，包括计算机存储器、计算机处理器以及存储在所述计算机存储器中并可在所述计算机处理器上执行的计算机程序，所述计算机存储器中存有训练好的分类模型，所述的分类模型采用改进的SE ResNet网络模型，共包含四个卷积部分，每个卷积部分由若干个子模块组构成，每个子模块组包含Inception子模块和SE Res子模块，最终通过全连接层获得最后的分类结果；A kind of viral encephalitis classification system of children's brain MR image, comprises computer memory, computer processor and is stored in described computer memory and can execute the computer program on described computer processor, stored in the described computer memory The trained classification model, the classification model adopts the improved SE ResNet network model, including four convolution parts, each convolution part is composed of several sub-module groups, and each sub-module group contains Inception sub-module and SE ResNet The sub-module finally obtains the final classification result through the fully connected layer;

其中，Inception子模块通过不同尺度的卷积来提升对不同大小的特征的学习能力；SE Res子模块包含SE和Res两部分，SE部分通过压缩-扩张通道数来提升模型对于有效特征的学习能力，Res部分通过跳跃连接对SE部分的输入特征矩阵X和输出特征矩阵

进行拼接，提升模型对于不同维度特征的学习能力；Among them, the Inception sub-module improves the learning ability of features of different sizes through convolution of different scales; the SE Res sub-module includes two parts, SE and Res, and the SE part improves the learning ability of the model for effective features by compressing and expanding the number of channels , the Res part is connected to the input feature matrix X and the output feature matrix of the SE part through a skip connection

Carry out splicing to improve the learning ability of the model for different dimensional features;

所述计算机处理器执行所述计算机程序时实现以下步骤：When the computer processor executes the computer program, the following steps are implemented:

将待分类的儿童脑部MR影像输入训练好的分类模型中，得到病毒性脑炎分类结果。Input the MR images of children's brains to be classified into the trained classification model to obtain the classification results of viral encephalitis.

进一步地，所述Inception子模块的结构如下：在获得上一层的数据输入X后，进入到多核卷积层L_incep，此层设计三个不同尺寸的卷积核以及一个池化核，即C_incep＝[C₁,C₂,C₃,P₁]；其中，C₁,C₂,C₃的卷积核大小分别是1*1、3*3、5*5，P₁的核大小为3*3；通过这些卷积核得到四个不同的特征

F＝[F₁,F₂,F₃,F₄]，然后将各特征进行拼接，最终得到多核卷积层L_incep的输出O＝Concat(F₁,F₂,F₃,F₄)。Further, the structure of the Inception sub-module is as follows: After obtaining the data input X of the previous layer, it enters the multi-core convolution layer_Lincep , and this layer designs three convolution kernels of different sizes and a pooling kernel, namely C_incep ＝[C₁ ,C₂ ,C₃ ,P₁ ]; Among them, the convolution kernel sizes of C₁ ,C₂ ,C₃ are 1*1, 3*3,5*5 respectively, and the kernel of P₁ The size is 3*3; four different features are obtained through these convolution kernels

F=[F₁ , F₂ , F₃ , F₄ ], and then concatenate the features to finally obtain the output O=Concat(F₁ , F₂ , F₃ , F₄ ) of the multi-core convolutional layer_Lincep .

SE Res子模块中，SE部分的结构如下：获取上一层的的输入数据X(c*w*h)，其中c,w,h分别代表特征矩阵的通道数、宽、高；首先用1*1大小的全局池化层进行池化，获得池化后的特征矩阵F₁＝GobelPool(X)，其大小为c*w*h；然后用1/16*c的通道数进行全连接卷积，获得F₂＝FC(F₁)，其大小为c/16*w*h；然后用c通道数的卷积层进行卷积，获得F₃＝FC(F₂)，其大小为c*1*1；然后用sigmoid激活函数进行归一化操作，将权重归一化到0-1之间，获得F₄＝Sigmoid(F₃)，其大小为c*1*1；最后用此结果对输入数据进行加权，获得

其大小为c*w*h。In the SE Res sub-module, the structure of the SE part is as follows: Get the input data X(c*w*h) of the upper layer, where c, w, and h represent the channel number, width, and height of the feature matrix respectively; first use 1 A global pooling layer of size *1 performs pooling to obtain the pooled feature matrix F₁ =GobelPool(X), whose size is c*w*h; then use 1/16*c channel number to perform fully connected volume product, and obtain F₂ =FC(F₁ ), whose size is c/16*w*h; then perform convolution with a convolutional layer with the number of c channels, and obtain F₃ =FC(F₂ ), whose size is c *1*1; then use the sigmoid activation function to perform normalization operations, normalize the weights to between 0-1, and obtain F₄ =Sigmoid(F₃ ), whose size is c*1*1; finally use this The results are weighted on the input data to obtain

Its size is c*w*h.

Res部分的结构如下：特征矩阵X(c*w*h)通过SE部分后得到

两者通道数都是c，通过跳跃连接将两个特征矩阵进行拼接，从而得到新的特征矩阵

其大小为2c*w*h。The structure of the Res part is as follows: The feature matrix X(c*w*h) is obtained after passing through the SE part

The number of channels for both is c, and the two feature matrices are spliced by skip connections to obtain a new feature matrix

Its size is 2c*w*h.

所述分类模型的训练过程如下：The training process of the classification model is as follows:

(1)收集患有病毒性脑炎和正常儿童患者的T1W序列MR影像数据，对影像数据进行预处理；(1) Collect T1W sequence MR image data of children with viral encephalitis and normal children, and preprocess the image data;

(2)将预处理后的影像数据划分为训练集、验证集和测试集；(2) Divide the preprocessed image data into training set, verification set and test set;

(3)将训练集送入到构建的分类模型中进行训练，利用验证集对分类模型的性能进行评估，根据评估的效果对模型的超参数进行调整，通过反复训练、验证，最终得到性能达标的分类模型。(3) Send the training set into the constructed classification model for training, use the verification set to evaluate the performance of the classification model, adjust the hyperparameters of the model according to the evaluation effect, and finally obtain the performance standard through repeated training and verification classification model.

步骤(1)中，所述的预处理包括对影像进行缩放，选取最大切片数作为标准，未达到此切片数的数据通过复制首尾切片进行补充，使各案例的输入数据保持一致；同时对影像进行尺度的归一化，并采用高斯滤波器滤除噪声。In step (1), the preprocessing includes zooming the image, selecting the maximum number of slices as a standard, and supplementing the data that does not reach the number of slices by copying the first and last slices, so that the input data of each case is consistent; Normalize the scale and use a Gaussian filter to filter out noise.

步骤(2)中，将预处理后的影像数据按7:1:2划分为训练集、验证集和测试集。In step (2), the preprocessed image data is divided into training set, verification set and test set according to 7:1:2.

步骤(3)中，采用监督训练方法对分类模型进行训练。In step (3), the classification model is trained using a supervised training method.

与现有技术相比，本发明具有以下有益效果：Compared with the prior art, the present invention has the following beneficial effects:

1、本发明创新性地提出了利用儿童脑部MR图像对儿童是否患有病毒性脑炎进行判断，不再需要腰穿脑脊液检查，因而也无需进行手术，减少了儿童患者的痛苦，也极大地提升了诊断的效率。1. The present invention innovatively proposes to use children's brain MR images to judge whether children have viral encephalitis, no need for lumbar puncture and cerebrospinal fluid examination, and therefore no need for surgery, which reduces the pain of children and greatly reduces the pain of children. Greatly improved the efficiency of diagnosis.

2、本发明中的分类模型采用改进的SE ResNet网络模型，在Inception网络模型的基础上加入了SE Res模块，Inception网络模型通过不同尺度的卷积来提升对不同大小的特征的学习能力，SE Res模块首先通过压缩-扩张通道数来提升模型对于有效特征的学习能力，然后通过跳跃连接提升模型对于不同维度特征的学习能力；大大提升了诊断的效率和准确率。2. The classification model in the present invention adopts the improved SE ResNet network model, and the SE Res module is added on the basis of the Inception network model. The Inception network model improves the learning ability of features of different sizes through convolution of different scales. SE The Res module first improves the learning ability of the model for effective features by compressing and expanding the number of channels, and then improves the learning ability of the model for features of different dimensions through skip connections; greatly improving the efficiency and accuracy of diagnosis.

附图说明Description of drawings

图1为本发明一种儿童脑部MR图像病毒性脑炎分类系统的实施流程图；Fig. 1 is the implementation flowchart of a kind of children's brain MR image viral encephalitis classification system of the present invention;

图2为本发明中分类模型的整体结构图；Fig. 2 is the overall structural diagram of classification model among the present invention;

图3为分类模型中每个卷积部分的结构图；Figure 3 is a structural diagram of each convolutional part in the classification model;

图4为本发明分类模型中Inception子模块的网络结构示意图；Fig. 4 is a schematic diagram of the network structure of the Inception submodule in the classification model of the present invention;

图5为本发明分类模型中SE Res子模块的网络结构图。Fig. 5 is a network structure diagram of the SE Res sub-module in the classification model of the present invention.

具体实施方式Detailed ways

下面结合附图和实施例对本发明做进一步详细描述，需要指出的是，以下所述实施例旨在便于对本发明的理解，而对其不起任何限定作用。The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be noted that the following embodiments are intended to facilitate the understanding of the present invention, but do not limit it in any way.

一种儿童脑部MR图像病毒性脑炎分类系统，包括计算机存储器、计算机处理器以及存储在计算机存储器中并可在计算机处理器上执行的计算机程序，计算机存储器中存有训练好的分类模型。A classification system for viral encephalitis in children's brain MR images, comprising a computer memory, a computer processor and a computer program stored in the computer memory and executable on the computer processor, and a trained classification model is stored in the computer memory.

如图1所示，整个系统的实施流程如下：As shown in Figure 1, the implementation process of the entire system is as follows:

1、图像预处理1. Image preprocessing

收集患有病毒性脑炎和正常儿童T1W期MR影像数据，对影像进行缩放，因为每个案例扫描的MR切片数不一致，本方法选取最大切片数作为标准，未达到此切片数的数据通过复制首尾切片进行补充，使各案例的输入数据保持一致，也包括对影像进行尺度的归一化，采用高斯滤波器滤除噪声。Collect the T1W MR image data of children with viral encephalitis and normal children, and scale the images. Because the number of MR slices scanned in each case is inconsistent, this method selects the maximum number of slices as the standard, and the data that does not reach this number of slices are copied The first and last slices are supplemented to make the input data of each case consistent, which also includes normalizing the scale of the image, and using a Gaussian filter to filter out noise.

2、数据分组2. Data grouping

将70％的数据集作为训练集，10％的数据集作为验证集，20％的数据集作为测试集。70% of the data set is used as a training set, 10% of the data set is used as a validation set, and 20% of the data set is used as a test set.

3、模型构建3. Model construction

构建分类模型，分类模型采用改进的SE ResNet网络模型，此网络是在Inception网络模型的基础上加入了SE Res模块。Inception网络模型通过不同尺度的卷积来提升对不同大小的特征的学习能力，SE Res模块首先通过压缩-扩张通道数来提升模型对于有效特征的学习能力，然后通过跳跃连接提升模型对于不同维度特征的学习能力。Construct the classification model, the classification model adopts the improved SE ResNet network model, this network is added the SE Res module on the basis of the Inception network model. The Inception network model improves the learning ability of features of different sizes through convolution of different scales. The SE Res module first improves the learning ability of the model for effective features by compressing and expanding the number of channels, and then improves the model for features of different dimensions through skip connections. learning ability.

如图2和图3所示，模型共包含四个卷积块，卷积块1-卷积块4，每个卷积块由若干个子模块组构成，每个子模块组包含Inception子模块和SE Res子模块。As shown in Figure 2 and Figure 3, the model contains a total of four convolution blocks, convolution block 1-convolution block 4, each convolution block consists of several sub-module groups, each sub-module group contains Inception sub-modules and SE Res submodule.

Inception子模块的结构如图4所示，首先在通过Inception子模块获取不同尺度的特征。在获得上一层的数据输入X后，进入到多核卷积层L_incep，此层设计三个不同尺寸的卷积核以及一个池化核，即C_incep＝[C₁,C₂,C₃,P₁]，其中，C₁,C₂,C₃的卷积核大小分别是1*1、3*3、5*5，P₁的核大小为3*3。通过这些卷积核可以得到四个不同的特征

F＝[F₁,F₂,F₃,F₄]，然后将各特征进行拼接，最终得到多核卷积层L_incep的输出O＝Concat(F₁,F₂,F₃,F₄)。The structure of the Inception sub-module is shown in Figure 4. First, the features of different scales are obtained through the Inception sub-module. After obtaining the data input X of the previous layer, enter the multi-core convolution layer_Lincep , this layer designs three convolution kernels of different sizes and a pooling kernel, that is, C_incep = [C₁ ,C₂ ,C₃ ,P₁ ], where the convolution kernel sizes of C₁ , C₂ , and C₃ are 1*1, 3*3, and 5*5 respectively, and the kernel size of P₁ is 3*3. Four different features can be obtained through these convolution kernels

F=[F₁ , F₂ , F₃ , F₄ ], and then concatenate the features to finally obtain the output O=Concat(F₁ , F₂ , F₃ , F₄ ) of the multi-core convolutional layer_Lincep .

因为Inception子模块最终获取的特征数较多，为了提升计算速度和模型精度，本方法采用SE Res子模块对特征进行加权筛选。Because the Inception sub-module finally obtains a large number of features, in order to improve the calculation speed and model accuracy, this method uses the SE Res sub-module to perform weighted screening of features.

SE Res子模块的结构如图5所示，SE Res子模块包含SE和Res两部分。SE部分中，获取上一层的输入数据X(c*w*h)，其中c,w,h分别代表特征矩阵的通道数、宽、高。首先用1*1大小的全局池化层进行池化，获得池化后的特征矩阵F₁＝GobelPool(X)，其大小为c*w*h，然后用1/16*c的通道数进行全连接卷积，获得F₂＝FC(F₁)，其大小为c/16*w*h，然后用c通道数的卷积层进行卷积，获得F₃＝FC(F₂)，其大小为c*1*1，然后用sigmoid激活函数进行归一化操作，将权重归一化到0-1之间，获得F₄＝Sigmoid(F₃)，其大小为c*1*1，最后用此结果对输入数据进行加权，获得

其大小为c*w*h。通过SE部分对各通道进行了加权，进而筛选出对结果有相关性的通道。Res部分中，特征矩阵X(c*w*h)通过SE模块后可得到

两者通道数都是c，将两个特征矩阵进行拼接，从而得到新的特征矩阵

其大小为2c*w*h。The structure of the SE Res sub-module is shown in Figure 5. The SE Res sub-module includes two parts, SE and Res. In the SE part, the input data X(c*w*h) of the previous layer is obtained, where c, w, and h represent the number of channels, width, and height of the feature matrix, respectively. First use the global pooling layer ofsize 1*1 for pooling, obtain the pooled feature matrix F₁ =GobelPool(X), its size is c*w*h, and then use the number of channels of 1/16*c to perform pooling Fully connected convolution, obtain F₂ =FC(F₁ ), whose size is c/16*w*h, and then perform convolution with a convolutional layer with the number of c channels, obtain F₃ =FC(F₂ ), which The size is c*1*1, and then the sigmoid activation function is used for normalization operation, and the weight is normalized to between 0-1 to obtain F₄ =Sigmoid(F₃ ), whose size is c*1*1, Finally, this result is used to weight the input data to obtain

Its size is c*w*h. Each channel is weighted by the SE part, and then the channel that is relevant to the result is screened out. In the Res part, the feature matrix X(c*w*h) can be obtained after passing through the SE module

The number of channels of both is c, and the two feature matrices are spliced to obtain a new feature matrix

Its size is 2c*w*h.

最终模型通过全连接卷积层，得到输入案例是否为病毒性脑炎的概率。The final model obtains the probability of whether the input case is viral encephalitis through a fully connected convolutional layer.

4、模型训练和分类测试4. Model training and classification testing

分割模型训练时，将训练集送入到分类模型中；验证集对模型的超参数进行调整，使用优化器更新参数，对网络进行优化，对学习率进行自动调参，得到训练完成的分类网络；测试集用来估计学习过程完成之后的模型的泛化能力。When splitting model training, the training set is sent to the classification model; the verification set adjusts the hyperparameters of the model, uses the optimizer to update the parameters, optimizes the network, and automatically adjusts the learning rate to obtain the trained classification network ; The test set is used to estimate the generalization ability of the model after the learning process is completed.

5、评估阶段5. Evaluation stage

在测试集上，对模型的分类效果进行评估：对分类任务的评估，需要计算每一类的精确率(Precision)和召回率(Recall)。每一类的精确率用正确分类到本类的案例(TruePositive,TP)除以所有分类到本类的案例数(TP+FP)，当不属于本类的案例被模型分到本类时，计数为假阳性(False Positive，FP)。每一类的召回率正确分类到本类的案例(TruePositive,TP)除以本类的真实案例数(TP+TN)，当属于本类的案例被模型分到其它类时，计数为假阴性(True Negative，TN)。最终智能脑炎诊断分类模型在测试集上的分类性能AUC进行评估，AU曲线是以假正率(FP_rate)和假负率(TP_rate)为轴的ROC(ReceiverOperating Characteristic)曲线下面的面积。On the test set, evaluate the classification effect of the model: for the evaluation of the classification task, it is necessary to calculate the precision rate (Precision) and recall rate (Recall) of each category. The accuracy rate of each category is divided by the cases correctly classified into this category (TruePositive, TP) by the number of cases classified into this category (TP+FP). When the cases that do not belong to this category are classified into this category by the model, The count is False Positive (FP). The recall rate of each category is correctly classified into the case of this category (TruePositive, TP) divided by the number of real cases of this category (TP+TN). When the cases belonging to this category are classified into other categories by the model, the count is false negative (True Negative, TN). The classification performance AUC of the final intelligent encephalitis diagnostic classification model on the test set is evaluated. The AU curve is the area under the ROC (Receiver Operating Characteristic) curve with the false positive rate (FP_rate) and false negative rate (TP_rate) as the axes.

以上所述的实施例对本发明的技术方案和有益效果进行了详细说明，应理解的是以上所述仅为本发明的具体实施例，并不用于限制本发明，凡在本发明的原则范围内所做的任何修改、补充和等同替换，均应包含在本发明的保护范围之内。The embodiments described above have described the technical solutions and beneficial effects of the present invention in detail. It should be understood that the above descriptions are only specific embodiments of the present invention, and are not intended to limit the present invention. All within the scope of the principles of the present invention Any modifications, supplements and equivalent replacements should be included within the protection scope of the present invention.

Claims (8)

Translated fromChinese

1.一种儿童脑部MR图像病毒性脑炎分类系统，包括计算机存储器、计算机处理器以及存储在所述计算机存储器中并可在所述计算机处理器上执行的计算机程序，其特征在于：1. a children's brain MR image viral encephalitis classification system, comprising computer memory, computer processor and being stored in the computer memory and the computer program that can be carried out on the described computer processor, it is characterized in that:所述计算机存储器中存有训练好的分类模型，所述的分类模型采用改进的SE ResNet网络模型，共包含四个卷积部分，每个卷积部分由若干个子模块组构成，每个子模块组包含Inception子模块和SE Res子模块，最终通过全连接层获得最后的分类结果；每个子模块由一个Inception子模块连接一个SE Res子模块再连接一个Inception子模块再连接一个SERes子模块再连接一个Inception子模块再连接一个SE Res子模块而构成；There is a trained classification model in the computer memory, and the classification model adopts an improved SE ResNet network model, which includes four convolution parts, each convolution part is composed of several sub-module groups, and each sub-module group Contains Inception sub-module and SE Res sub-module, and finally obtains the final classification result through the fully connected layer; each sub-module is connected by an Inception sub-module to a SE Res sub-module, then to an Inception sub-module, to a SERes sub-module, and to a The Inception sub-module is connected to a SE Res sub-module to form;其中，Inception子模块通过不同尺度的卷积来提升对不同大小的特征的学习能力；SERes子模块包含SE和Res两部分，SE部分通过压缩-扩张通道数来提升模型对于有效特征的学习能力，Res部分通过跳跃连接对SE部分的输入特征矩阵X和输出特征矩阵

进行拼接，提升模型对于不同维度特征的学习能力；Among them, the Inception sub-module improves the learning ability of features of different sizes through convolution of different scales; the SERes sub-module includes two parts, SE and Res, and the SE part improves the learning ability of the model for effective features by compressing and expanding the number of channels. The Res part is connected to the input feature matrix X and the output feature matrix of the SE part through a skip connection

Carry out splicing to improve the learning ability of the model for different dimensional features;所述计算机处理器执行所述计算机程序时实现以下步骤：When the computer processor executes the computer program, the following steps are implemented:将待分类的儿童脑部MR影像输入训练好的分类模型中，得到病毒性脑炎分类结果。Input the MR images of children's brains to be classified into the trained classification model to obtain the classification results of viral encephalitis.2.根据权利要求1所述的儿童脑部MR图像病毒性脑炎分类系统，其特征在于，所述Inception子模块的结构如下：2. children's brain MR image viral encephalitis classification system according to claim 1, is characterized in that, the structure of described Inception submodule is as follows:在获得上一层的数据输入X后，进入到多核卷积层L_incep，此层设计三个不同尺寸的卷积核以及一个池化核，即C_incep＝[C₁,C₂,C₃,P₁]；其中，C₁,C₂,C₃的卷积核大小分别是1*1、3*3、5*5，P₁的核大小为3*3；通过这些卷积核得到四个不同的特征

F＝[F₁,F₂,F₃,F₄]，然后将各特征进行拼接，最终得到多核卷积层L_incep的输出O＝Concat(F₁,F₂,F₃,F₄)。After obtaining the data input X of the previous layer, enter the multi-core convolution layer_Lincep , this layer designs three convolution kernels of different sizes and a pooling kernel, that is, C_incep = [C₁ ,C₂ ,C₃ ,P₁ ]; among them, the convolution kernel sizes of C₁ , C₂ , and C₃ are 1*1, 3*3, and 5*5 respectively, and the kernel size of P₁ is 3*3; through these convolution kernels, four distinct features

F=[F₁ , F₂ , F₃ , F₄ ], and then concatenate the features to finally obtain the output O=Concat(F₁ , F₂ , F₃ , F₄ ) of the multi-core convolutional layer_Lincep .3.根据权利要求1所述的儿童脑部MR图像病毒性脑炎分类系统，其特征在于，SE Res子模块中，SE部分的结构如下：3. the children's brain MR image viral encephalitis classification system according to claim 1, is characterized in that, in the SE Res submodule, the structure of the SE part is as follows:获取上一层的输入数据X(c*w*h)，其中c,w,h分别代表特征矩阵的通道数、宽、高；首先用1*1大小的全局池化层进行池化，获得池化后的特征矩阵F₁＝GobelPool(X)，其大小为c*w*h；然后用1/16*c的通道数进行全连接卷积，获得F₂＝FC(F₁)，其大小为c/16*w*h；然后用c通道数的卷积层进行卷积，获得F₃＝FC(F₂)，其大小为c*1*1；然后用sigmoid激活函数进行归一化操作，将权重归一化到0-1之间，获得F₄＝Sigmoid(F₃)，其大小为c*1*1；最后用此结果对输入数据进行加权，获得

其大小为c*w*h。Obtain the input data X(c*w*h) of the previous layer, where c, w, and h respectively represent the number of channels, width, and height of the feature matrix; first, pooling is performed with a global pooling layer of 1*1 size to obtain The pooled feature matrix F₁ =GobelPool(X), whose size is c*w*h; then use the channel number of 1/16*c to perform fully connected convolution to obtain F₂ =FC(F₁ ), which The size is c/16*w*h; then convolution is performed with a convolutional layer with the number of c channels to obtain F₃ =FC(F₂ ), whose size is c*1*1; and then normalized with the sigmoid activation function normalize the weights between 0-1 to obtain F₄ =Sigmoid(F₃ ), whose size is c*1*1; finally use this result to weight the input data to obtain

Its size is c*w*h.4.根据权利要求1所述的儿童脑部MR图像病毒性脑炎分类系统，其特征在于，SE Res子模块中，Res部分的结构如下：4. the children's brain MR image viral encephalitis classification system according to claim 1, is characterized in that, in the SE Res submodule, the structure of the Res part is as follows:特征矩阵X(c*w*h)通过SE部分后得到

两者通道数都是c，通过跳跃连接将两个特征矩阵进行拼接，从而得到新的特征矩阵

其大小为2c*w*h。The characteristic matrix X(c*w*h) is obtained after passing the SE part

The number of channels for both is c, and the two feature matrices are spliced by skip connections to obtain a new feature matrix

Its size is 2c*w*h.5.根据权利要求1所述的儿童脑部MR图像病毒性脑炎分类系统，其特征在于，所述分类模型的训练过程如下：5. children's brain MR image viral encephalitis classification system according to claim 1, is characterized in that, the training process of described classification model is as follows:(1)收集患有病毒性脑炎和正常儿童患者的T1W序列MR影像数据，对影像数据进行预处理；(1) Collect T1W sequence MR image data of children with viral encephalitis and normal children, and preprocess the image data;(2)将预处理后的影像数据划分为训练集、验证集和测试集；(2) Divide the preprocessed image data into training set, verification set and test set;(3)将训练集送入到构建的分类模型中进行训练，利用验证集对分类模型的性能进行评估，根据评估的效果对模型的超参数进行调整，通过反复训练、验证，最终得到性能达标的分类模型。(3) Send the training set into the constructed classification model for training, use the verification set to evaluate the performance of the classification model, adjust the hyperparameters of the model according to the evaluation effect, and finally get the performance up to standard through repeated training and verification classification model.6.根据权利要求5所述的儿童脑部MR图像病毒性脑炎分类系统，其特征在于，步骤(1)中，所述的预处理包括对影像进行缩放，选取最大切片数作为标准，未达到此切片数的数据通过复制首尾切片进行补充，使各案例的输入数据保持一致；同时对影像进行尺度的归一化，并采用高斯滤波器滤除噪声。6. children's brain MR image viral encephalitis classification system according to claim 5, is characterized in that, in step (1), described preprocessing comprises that image is zoomed, chooses maximum slice number as standard, does not The data that reaches this number of slices is supplemented by copying the first and last slices, so that the input data of each case is consistent; at the same time, the scale of the image is normalized, and the Gaussian filter is used to filter out the noise.7.根据权利要求5所述的儿童脑部MR图像病毒性脑炎分类系统，其特征在于，步骤(2)中，将预处理后的影像数据按7:1:2划分为训练集、验证集和测试集。7. children's brain MR image viral encephalitis classification system according to claim 5, is characterized in that, in step (2), the image data after pretreatment is divided into training set, verification by 7:1:2 set and test set.8.根据权利要求5所述的儿童脑部MR图像病毒性脑炎分类系统，其特征在于，步骤(3)中，采用监督训练方法对分类模型进行训练。8. children's brain MR image viral encephalitis classification system according to claim 5, is characterized in that, in step (3), adopts supervised training method to train classification model.

Images