Middle ear and bile adipoma automatic detection system based on CT image

Abstract

The invention discloses a CT image-based middle ear and bile adipoma automatic detection system, the system comprises four modules, a CT image diseased area automatic segmentation module can comprehensively analyze a CT image from the global and local through fusion of a multi-head self-attention mechanism, a convolution layer of a plurality of different convolution kernels and a depth separable convolution, and the CT image diseased area automatic segmentation module can automatically segment the diseased area of the CT image from the local to the global through the multi-head self-attention mechanism. Deeper characterization information of the CT image can be obtained by using the convolution residual network and 3D cavity convolution, and the middle ear and bile adipoma area can be better and automatically segmented by comparing a model segmentation result with segmented images of different sizes. And the middle ear and bile adipoma feature screening module screens the extracted features by using U test, Spearman correlation coefficient calculation and a Lasso model, and distinguishes the otitis media from the bile adipoma by using a multi-layer perceptron model. The method can automatically segment the middle ear and bile adipoma area and help a doctor to distinguish the otitis media from the bile adipoma.

Description

Technical Field

[0001] The invention relates to the technical field of medical image processing, and in particular to an automatic detection system for middle ear cholesteatoma based on CT images. Background Art

[0002] The topic of middle ear cholesteatoma has been the subject of much research and debate worldwide. Studies have found that 0.5% to 30% of the community suffers from chronic otitis media, affecting more than 20 million people worldwide. About 5 million people, or one in four patients, suffer from cholesteatoma, and the severity of the disease is mainly due to its impact on the ear, including pressure and infection. This system can help doctors better detect cholesteatoma.

[0003] In the prior art, some people use tympanic membrane photos to detect middle ear cholesteatoma, but the information that can be obtained by this method is less than that obtained through CT image data. At present, there are few automatic detection systems for middle ear cholesteatoma based on CT image data and they cannot fully utilize the CT image data information.

[0004] Therefore, there are few existing systems that use CT image data to detect middle ear cholesteatoma, and they cannot comprehensively analyze CT image data, cannot effectively distinguish otitis media from cholesteatoma, and cannot assist doctors in proposing effective treatment plans. Summary of the invention

[0005] To solve the above problems, the present invention provides an automatic detection system for middle ear cholesteatoma based on CT images to assist doctors in proposing effective treatment plans for middle ear cholesteatoma.

[0006] The present invention is an automatic detection system for middle ear cholesteatoma based on CT images, which includes a CT image diseased area automatic segmentation module, a CT image feature extraction module, a middle ear cholesteatoma feature screening module, and a middle ear cholesteatoma differentiation module, a total of 4 modules, which are implemented by the following technical solutions:

[0007] Automatic segmentation module of diseased areas in CT images: This module consists of an encoding submodule (Encoder), a pooling submodule, a deep feature information representation submodule (CRN-DCN), a decoding submodule (Decoder), and an output submodule.

[0008] The encoder module (Encoder) is composed of a multi-head self-attention mechanism and a local residual multi-convolution network (Multi-Head Attention - local residual Convolution Network) abbreviated as GA-LRCN and a depthwise separable convolution network (Depthwise Separable Convolution Network) abbreviated as DSCN; there are four encoder modules with the same structure, which are called the first encoder module, the second encoder module, the third encoder module, and the fourth encoder module in the order of appearance; the encoder module receives the input Encoder , and input Encoder Input to the multi-head self-attention mechanism to obtain the output of the multi-head self-attention mechanism, and convert the input Encoder The output is input into a convolutional network composed of multiple different convolution kernels to obtain the output of the local residual multi-convolutional network. The output of the multi-head self-attention mechanism is added to the corresponding elements of the output of the local residual multi-convolutional network and then subjected to a 3D convolution to obtain the output GA-LRCN The depthwise separable convolutional network takes input Encoder Input to 3D channel-by-channel convolution to get input DSCN_dep , input DSCN_dep Input 3D point-by-point convolution to get output DSCN . Output GA-LRCN and output DSCN Add together to get the output of the encoding submodule Encoder , 3D channel-by-channel convolution performs convolution operations independently on each input channel without mixing across channels; 3D point-by-point convolution uses a 1×1×1 3D convolution to perform a convolution operation on the output of the deep convolution layer in the channel dimension to achieve information integration and feature fusion between channels.

[0009] The pooling submodule consists of 3D maximum pooling and ReLu. The input of the pooling submodule is subjected to 3D maximum pooling and ReLu in turn to obtain the output of the pooling submodule.

[0010] The deep feature information representation submodule (CRN-DCN) consists of a convolutional residual network (CRN) and a 3D dilated convolution (DCN). The deep feature information representation submodule receives input CRN-DCN , input CRN-DCN Input 3D convolution, 3D batch normalization layer, ReLu, 3D convolution, 3D batch normalization layer, the result is the same as inputCRN-DCN Add the result output CRN1 , output CRN1 Input 3D hole convolution to get input DCN , input DCN Input 3D convolution, 3D batch normalization layer, ReLu, 3D convolution, 3D batch normalization layer in sequence to get the output result and compare it with the input DCN The result of the addition is then added to the input CRN-DCN Add together to get the output of the deep feature information representation submodule CRN-DCN .

[0011] The decoding submodule (Decoder) consists of 3D convolution, trilinear interpolation and ReLu. The input of the decoding submodule is sequentially subjected to 3D convolution, trilinear interpolation and ReLu to obtain the output of the decoding submodule. There are four decoding submodules with the same structure. They are called the first decoding submodule, the second decoding submodule, the third decoding submodule and the fourth decoding submodule in the order of appearance.

[0012] The output submodule consists of 3D convolution and softmax. The input of the output submodule is sequentially subjected to 3D convolution and softmax to obtain the output of the output submodule.

[0013] The automatic segmentation module of the diseased area of ​​the CT image has submodules connected in the following manner: the first encoding submodule, the first pooling submodule, the second encoding submodule, the second pooling submodule, the third encoding submodule, the third pooling submodule, the fourth encoding submodule, the fourth pooling submodule, the deep feature information representation submodule, and the first decoding submodule are connected in series in sequence, and then the output of the fourth encoding submodule is spliced ​​to the tail of the output of the first decoding submodule to obtain the output of the fourth encoding submodule and the output of the first decoding submodule; then the output of the fourth encoding submodule and the output of the first decoding submodule are spliced ​​into the second decoding submodule to obtain the output of the second decoding submodule; then the output of the third encoding submodule is spliced ​​to the tail of the output of the second decoding submodule to obtain the output of the third encoding submodule and the output of the second decoding submodule. The output of the third encoding submodule and the output of the second decoding submodule are then input into the third decoding submodule to obtain the output of the third decoding submodule; the output of the second encoding submodule is then spliced ​​to the tail of the output of the third decoding submodule to obtain the output of the second encoding submodule and the output of the third decoding submodule; the output of the second encoding submodule and the output of the third decoding submodule are then input into the fourth decoding submodule to obtain the output of the fourth decoding submodule; the output of the fourth encoding submodule is then spliced ​​to the tail of the output of the first decoding submodule to obtain the output of the fourth encoding submodule and the output of the first decoding submodule, and the output of the fourth encoding submodule and the output of the first decoding submodule are input into the output submodule to obtain the output result of the automatic segmentation module of the diseased area of ​​the CT image.

[0014] The segmentation result target_pre of the automatic segmentation module of the diseased area of ​​the CT image is compared with target1, target2, and target3. Specifically, the original segmented image target0 and target_pre are substituted into the formula loss4 to obtain loss4, and target1, target2, and target3 are respectively substituted into the target variable of the jac formula and target_pre to calculate jac, where target1, target2, and target3 are respectively expanded by 1, 2, and 3 pixel values ​​on the basis of the diseased area of ​​the CT image and converted into tensors and divided into training set data. The jac calculation results are recorded as loss1, loss2, and loss3, and the final loss is .

[0015] CT image feature extraction module: The diseased area of ​​the CT image automatically segmented by the module uses pyradiomics (a tool for extracting feature values ​​of medical imaging genomics) to extract the features of the diseased area of ​​each patient's CT image.

[0016] Middle ear cholesteatoma feature screening module: label and Perform U test and calculate p value. Feature deletion, get feature ; where label is the feature column of whether the patient is otitis media or cholesteatoma; then the Spearman correlation coefficient is used to calculate Correlation coefficient between features, only one feature with a correlation coefficient greater than 0.8 is retained, and then the ; Finally, Input the Lasso model (Lasso Regression Algorithm), reduce the coefficients of some unimportant features to zero, and select r features .

[0017] Middle ear cholesteatoma differentiation module: Input into the multi-layer perceptron model. And save the trained multi-layer perceptron model to distinguish otitis media from cholesteatoma.

[0018] The beneficial effects of the present invention are: the encoding submodule in the automatic segmentation module of the diseased area of ​​the CT image can analyze the global and local information of the data, and the introduction of the deep separable convolution can reduce the amount of data calculation. The residual connection in the deep feature information characterization submodule can effectively learn feature representation, and the 3D hole convolution can increase the receptive field, capture a wider context information, and help the network learn a more global and complex feature representation. The segmentation result target_pre obtained by the automatic segmentation module of the diseased area of ​​the CT image is compared with target0, target1, target2, and target3, which is conducive to improving the consistency between the segmentation result of the middle ear cholesteatoma under target1, target2, and target3 of the automatic segmentation module of the diseased area of ​​the CT image and target0, and improving the adaptability of the algorithm under different image quality conditions, and improving the segmentation accuracy. Then, the otitis media cholesteatoma can be further distinguished by the trained middle ear cholesteatoma distinction model, providing more reliable help for the segmentation of the diseased area of ​​otitis media and cholesteatoma and the distinction between otitis media and cholesteatoma. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 The overall flow chart of the present invention is Figure 2Module diagram for automatic segmentation of diseased areas in CT images Figure 3 Diagram of the feature screening module for middle ear cholesteatoma Figure 4 The encoding submodule diagram Figure 5 The submodule diagram for deep feature information representation Figure 6 Decoding submodule diagram DETAILED DESCRIPTION

[0020] The modules and implementation methods of the present invention are described below. According to the automatic detection system for middle ear cholesteatoma based on CT images proposed by the present invention, its implementation methods are as follows: CT image diseased area automatic segmentation module: The automatic segmentation module of the diseased area of ​​CT images consists of an encoding submodule (Encoder), a pooling submodule, a deep feature information representation submodule (CRN-DCN), a decoding submodule (Decoder), and an output submodule; First, convert each patient's CT image data and its corresponding diseased area image target into tensors data_ori and target0_ori, respectively. Add an extra dimension to the first and second dimensions of the tensor, so that the data dimension of data_ori is 1×n 0 ×s 0 ×h 0 × 0 (where n 0 is the number of channels, s 0 is the number of two-dimensional matrices, i.e., the number of CT image layers, h 0 is the length of the two-dimensional matrix, w 0 is the width of the two-dimensional matrix), the target0_ori data dimension is 1×2×s 0 ×h 0 × 0 , based on the target0_ori data, expand 1, 2, and 3 pixel values ​​respectively and convert them into tensors target1_ori, target2_ori, and target3_ori respectively. Then, the first 80% of the data_ori, target0_ori, target1_ori, target2_ori, and target3_ori data are used as training sets and named data, target0, target1, target2, and target3 respectively. The rest are used as test sets and named data_test, target0_test, target1_test, target2_test, and target3_test respectively; Then input the data into the first encoding submodule to get the output data c1 , and then data c1 Input to the first pooling submodule to get the output data p1 , data p1 Input to the second encoding submodule to get the output data c2 , and then data c2 Input to the second pooling submodule to get the output data p2 , data p2 Input to the third encoding submodule to get the output data c3 , and then data c3 Input to the third pooling submodule to get the output data p3 , data p3 Input to the fourth encoding submodule to get the output data c4 , and then data c4 Input to the fourth pooling submodule to get the output data p4 ; Then the data p4 Input to the deep feature information representation submodule to get the output data cd , data cd Input to the first decoding submodule to get the output data up1 , and then data c4 With data up1 According to the second dimension, we splice together to get data c4_up1 , data c4_up1 Input to the second decoding submodule to get the output data up2 , and then data c3 With data up2 According to the second dimension, we splice together to get data c3_up2 , data c3_up2 Input to the third decoding submodule to get the output data up3 , and then data c2 With data up3 According to the second dimension, we splice together to get data c2_up3 , data c2_up3 Input to the fourth decoding submodule to get the output data up4 , and then data c1 With data up4 According to the second dimension, we splice together to get the output data c1_up4 , data c1_up4 Input to the output submodule to get target_pre; Then calculate the loss4 of target_pre and target0; substitute target_pre, target1, target2, and target3 into the target variable of the jac formula and target_pre to calculate jac, and get loss1, loss2, and loss3, and then calculate , where the loss4 formula is: (1) The jac formula is: (2) (3) Where target refers to target1, target2, target3; Then train the model, where the number of training rounds is epoch (epoch is an integer greater than 1), the batch size is batch, the global random seed is 42, the initial learning rate is 0.01, and after every 10 rounds of training, the learning rate is decayed by a multiple of 0.1. The Adam algorithm is used to calculate the gradient of the loss to update the parameters of the automatic segmentation module of the diseased area of ​​the CT image. Finally, the trained automatic segmentation module of the diseased area of ​​the CT image is saved, and all patient CT images are automatically segmented; Among them, data c1 、data p1 、data c2 、data p2 、data c3 、data p3 、data c4 、data p4 、data cd 、data up1 、data c4_up1 、data up2 、data c3_up2 、data up3 、data c2_up3 、data up4 、data c1_up4 , target_pre data dimensions are: 1×n 0 ×s 0 ×h 0 × 0 , 1×n 1 ×s 0 ×h 0 × 0 , 1×n 1 ×s 1 ×h 1 ×1 , 1×2n 1 ×s 1 ×h 1 × 1 , 1×2n 1 ×s 2 ×h 2 × 2 , 1×4n 1 ×s 2 ×h 2 × 2 , 1×4n 1 ×s 3 ×h 3 × 3 , 1×8n 1 ×s 3 ×h 3 × 3 , 1×8n 1 ×s 4 ×h 4 × 4 , 1×8n 1 ×s 4 ×h 4 × 4 , 1×4n 1 ×s 3 ×h 3 × 3 , 1×12n 1 ×s 3 ×h 3 × 3 , 1×2n 1 ×s 2 ×h 2 × 2 , 1×6n 1 ×s 2 ×h 2 × 2 , 1×n 1 ×s 1 ×h 1 × 1 , 1×3n 1 ×s 1 ×h 1 × 1 , 1×2×s 0 ×h 0 × 0 , 1×(2+ n 1 )×s 0 ×h 0 × 0 , 1×2×s 0 ×h 0 × 0 , where n0 、n 1 is the number of channels n 0 ≠n 1 ;s 0 、s 1 、s 2 、s 3 、s 4 is the number of two-dimensional matrices s 0 ≠s 1 ≠s 2 ≠s≠s 4 ;h 0 、h 1 、h 2 、h 3 、h 4 is the length of the two-dimensional matrix h 0 ≠h 1 ≠h 2 、h 3 ≠h 4 ;w 0 、w 1 、w 2 、w 3 、w 4 is the width of the two-dimensional matrix w 0 ≠w 1 ≠w 2 ≠w 3 ≠w 4 ; The encoder module (Encoder) consists of a multi-head self-attention mechanism and a local residual multi-convolution network (Multi-HeadAttention - local residual Convolution Network) referred to as GA-LRCN and a depth-wise separable convolution network (Depthwise Separable Convolution Network) referred to as DSCN. Specifically: Assume that the input of the encoder module is input Encoder , input Encoder The outputs are respectively input into the multi-head self-attention mechanism, the local residual multi-convolutional network, and the depth-separable convolutional network to obtain output GA-LRCN 、output DSCN , output GA-LRCN and output DSCN Add together to get the output of the encoding submodule Encoder , where the multi-head self-attention mechanism and the local residual multi-convolutional network receive input EncoderAnd input them into the multi-head self-attention mechanism (Multi-Head Attention), the 3D convolution with convolution kernel size of a×a×a, b×b×b, c×c×c (a≠b≠c), and get the output input respectively. Attention 、input conv_a 、input conv_b 、input conv_c , input conv_a 、input conv_b 、input conv_c After being processed by the 3D batch normalization layer, the output input is obtained respectively. conv_a_bn 、input conv_b_bn 、input conv_c_bn , input conv_a_bn 、input conv_b_bn 、input conv_c_bn With input Encoder Add the corresponding elements to get input LRCN , input LRCN With input Attention Add the corresponding elements to get input LRCN _ Attention , input LRCN _ Attention Input 3D convolution to get output GA-LRCN , where the depthwise separable convolutional network receives input Encoder , input Encoder Input 3D channel-by-channel convolution to get output input DSCN_dep , input DSCN_dep Input 3D point-by-point convolution to get output output DSCN , where the 3D batch normalization layer standardizes the data of each feature channel; a×a×a, b×b×b, c×c×c are different convolution kernel sizes; 3D channel-by-channel convolution performs convolution operations independently on each input channel without mixing across channels; 3D point-by-point convolution uses a 1×1×1 3D convolution to perform convolution operations on the output of the deep convolution layer in the channel dimension to achieve information integration and feature fusion between channels; The pooling submodule consists of 3D maximum pooling and ReLu. Specifically: Assume that the input of the pooling submodule is input maxpool_relu , input maxpool_relu Input to 3D maximum pooling to get the result input maxpool , input maxpool After ReLu, the output submodule output is obtained relu ; The deep feature information representation submodule (CRN-DCN) consists of a convolutional residual network (CRN) and a 3D dilated convolution (DCN). Specifically: Assume that the input data is input CRN-DCN , input CRN-DCN Input convolution residual network to get output output CRN1 , output CRN1 Input to 3D hole convolution to get output input DCN , input DCN The output obtained after ReLu processing is input into the convolutional residual network to obtain the output CRN2 , output CRN2 With input CRN-DCN The corresponding elements are added to get the output of the deep feature information representation submodule. CRN-DCN , where the convolutional residual network inputs the input data into the 3D convolution to obtain the output of the first convolution, and then inputs the result obtained by the 3D batch normalization layer and ReLu processing into the 3D convolution and the 3D batch normalization layer to obtain the output, and adds the obtained output to the convolutional residual network input data to obtain the convolutional residual network output. The 3D hole convolution is similar to the 3D convolution calculation method, except that intervals are introduced between the convolution kernels; The decoding submodule (Decoder) consists of 3D convolution, trilinear interpolation and ReLu. Specifically: Assume that the input of the decoding submodule is input Decoder , input Decoder Input to 3D convolution to get the result input Decoder_conv , input Decoder_conv After trilinear interpolation, the data input is obtained Decoder_liner , input Decoder_liner After ReLu, the result output is obtained Decoder_ReLu ; The output submodule consists of 3D convolution and softmax. Specifically, assume that the input of the output submodule is input convsoft , input convsoft After inputting into the 3D convolution, the result is input conv After softmax, the output submodule output is obtained soft ; The 3D convolution extracts features by sliding the convolution kernel on the three dimensions of the 3D convolution input. The convolution kernel is a set of learnable parameters that perform convolution operations with the input of the 3D convolution. At each position, the convolution kernel is element-wise multiplied with the corresponding position of the 3D convolution input, and the results are added to generate the output data. The 3D max pooling layer finds the maximum value in each region of the input feature map and uses it as the value of the corresponding region of the output data; ReLu is defined as: (4) in is the input data, the function means input When the output is greater than 0, it is equal to the input; when the input is less than or equal to 0, the output is 0; Trilinear interpolation is to calculate the value of the target pixel by weighted average of the surrounding pixel values. In the three-dimensional case, interpolation is performed in three directions at the same time. Specifically, for the position of the target pixel, the nearest neighboring pixel is found in each direction, and a weighted average is performed based on their distance to calculate the interpolation result. Softmax converts the original output data into a probability distribution so that the probability of each category is between 0 and 1, and the sum of the probabilities of all categories is 1. The formula is: (5) in is the original output vector, Indicates the first elements, Represents an exponential function.

[0021] CT image feature extraction module: Using pyradiomics (a tool for extracting eigenvalues ​​of medical imaging omics), the diseased areas of the patient's CT images that were automatically segmented by the CT image diseased area automatic segmentation module were subjected to wavelet transform, LoG filtering, calculation of the square of each pixel value in the image, calculation of the square root of each pixel value in the image, calculation of the logarithm of each pixel value in the image, calculation of the exponent of each pixel value in the image, calculation of the gradient of each pixel value in the image, two-dimensional local binary pattern, and three-dimensional local binary pattern to obtain images after passing through various filters, and the Shape, First Order (first order), GLCM (gray-level co-occurrence matrix), GLSZM (gray-level size zone matrix), GLDM (gray-level difference matrix), NGTDM (gray-level gradient co-occurrence matrix) features, use the radiomics library in python to normalize the CT grayscale values, the normalized scale is 1000, set the resampling interpolation method to 'sitkNearestNeighbor', set the pixel spacing after resampling to [1,1,1], set the additional filling distance for LoG filtering to 10, set the minimum size of the region of interest to 1, perform verification, set the bin width of the grayscale value to 5, set the HU value in the CT image to be increased by 1000 as a whole, and divide the features of the diseased area of ​​the patient's CT image according to the principle of dividing the training set and the test set of the automatic segmentation module of the diseased area of ​​the CT image. Data is divided into training set and test set , and use The standard deviation and mean of are used to regularize the feature data extracted from the diseased areas of all patients’ CT images. The specific formula is: (4) in is the first feature column, when When standardizing for , when When standardizing for , For unregularized Mean, For unregularized Standard deviation.

[0022] Middle ear cholesteatoma feature screening module: The Mann-Whitney U test (abbreviated as: U test) was used to Conduct preliminary screening to obtain , specifically: calculate label and rank, then label and The feature data is rank transformed, and then the label and The sum of the ranks of the two sets of feature data is recorded as U1 and U2 respectively, and then the smaller U1 and U2 are selected as the test statistic U, and the mean (mean_U) and standard deviation (std_U) of the test statistic U are calculated. Among them, label is the feature column of whether the patient has otitis media or cholesteatoma; according to the properties of the standard normal distribution, the z score of the test statistic U is calculated as follows: (7) The cumulative probability of a given z-score is calculated using the normal distribution's cumulative distribution function (CDF), which is: (8) in, represents the cumulative probability of a given z-score. Multiply by 2 to get the p-value and delete if the p-value is less than 0.05 Features. Label is the feature column of whether the patient is otitis media or cholesteatoma; rank refers to the ranking assigned to each data after sorting the sample data; rank transformation refers to converting the original sample data into the corresponding rank. Using the Spearman correlation coefficient, Rescreen to obtain the filtered feature data , specifically: The Spearman correlation coefficient calculation formula is as follows: (9) is the Spearman correlation coefficient, is the sample size; The different features are sorted from small to large and assigned a ranking (starting from 1), and then calculated The difference between the corresponding rankings of different features is , only one feature with a feature correlation coefficient greater than 0.8 is retained to obtain the filtered feature data ; Then the regularization parameters of the Lasso model (Lasso Regression Algorithm) are traversed. The data is re-divided using five-fold cross validation, and the divided data is used as the input of the Lasso model. Then the Lasso model results under different regularization parameters are obtained. Then the MSE (the average of the square of the difference between the actual value and the predicted value of all data) of the divided training data under the same regularization parameter is calculated, and the MSE under different regularization parameters is compared. The regularization parameter that minimizes the MSE is selected as the regularization parameter of the Lasso model, and then Substitute it into the Lasso model with the regularization parameter determined, and filter out the regression coefficients greater than 0 through the Lasso model results. Features , and in Select the above r features as The final feature ,in , , , , is a positive integer greater than or equal to 1.

[0023] Middle ear cholesteatoma differentiation module: Input into the multi-layer perceptron model for training, where the parameters of the multi-layer perceptron model are: three hidden layers, each hidden layer corresponds to 3*hide MLP 、2*hide MLP 、hide MLP neurons; the maximum number of training rounds is epoch MLP ; Set the global random seed to globalseed; Set the parameters of the stochastic gradient descent (SGD) algorithm to update the model. MLP , epoch MLP is a positive integer, globalseed is an integer.

[0024] Then save the trained multi-layer perceptron model, and finally distinguish otitis media from cholesteatoma. Further, you can build a server and deploy this system to the web page. You can also deploy it directly to a computer for use. Or deploy it to a hardware device.

[0025] Parameter table of this patent:

[0026]

[0027]

[0028]

Claims

1. An automatic detection system for middle ear cholesteatoma based on CT images, including a CT image diseased area automatic segmentation module, a CT image feature extraction module, a middle ear cholesteatoma feature screening module, and a middle ear cholesteatoma differentiation module; The automatic segmentation module of the diseased area of ​​the CT image is composed of an encoding submodule (Encoder), a pooling submodule, a deep feature information representation submodule (CRN-DCN), a decoding submodule (Decoder), and an output submodule; The encoder module (Encoder) is composed of a multi-head self-attention mechanism and a local residual multi-convolution network (Multi-HeadAttention - local residual Convolution Network) abbreviated as GA-LRCN and a depthwise separable convolution network (Depthwise Separable Convolution The invention relates to a novel deep neural network (CNN) which is composed of a multi-head self-attention mechanism and a local residual multi-convolutional network, abbreviated as DSCN. There are four encoding modules with the same structure, which are called the first encoding submodule, the second encoding submodule, the third encoding submodule, and the fourth encoding submodule in the order of appearance. The overall advantage of the design scheme of the encoding submodule is that it can analyze the global and local information of the data, among which the advantage of the multi-head self-attention mechanism and the local residual multi-convolutional network is that it can effectively extract the global and local features of the data, and the advantage of the deep separable convolutional network is that it can further extract the local features of the data and reduce the amount of data calculation. The multi-head self-attention mechanism and the local residual multi-convolutional network are composed of a multi-head self-attention mechanism and three 3D convolutions with different convolution kernel sizes in parallel. The specific parameters are: the multi-head self-attention mechanism and the three 3D convolutions with different convolution kernel sizes receive the input of the encoding submodule, and the convolution kernels of the three convolution kernel sizes are respectively a×a×a, b×b×b, and c×c×c (where a, b, c are integers greater than or equal to 1) are connected in parallel and their final results are combined. The result is added together with the input of the encoding submodule, and the result obtained by the addition is added to the output of the multi-head self-attention mechanism to obtain the output of the multi-head self-attention mechanism and the local residual multi-convolutional network; the deep separable convolutional network is composed of two 3D convolutions with different convolution kernel sizes connected in series, specifically: the input of the deep separable convolutional network is independently convolved on each input channel, and the result obtained by performing a convolution operation with a convolution kernel size of 1×1×1 on the channel dimension to obtain the output of the deep separable convolutional network; the multi-head self-attention mechanism and the local residual multi-convolutional network are connected in parallel with the deep separable convolutional network, and the multi-head self-attention mechanism and the local residual multi-convolutional network and the deep separable convolutional network receive the input of the encoding submodule, and respectively obtain the output of the multi-head self-attention mechanism, the output of the local residual multi-convolutional network and the output of the deep separable convolutional network; then the output of the multi-head self-attention mechanism, the output of the local residual multi-convolutional network and the output of the deep separable convolutional network are added to obtain the output of the encoding submodule; The pooling submodule is composed of 3D maximum pooling and ReLu. The input of the pooling submodule is sequentially subjected to 3D maximum pooling and ReLu to obtain the output of the pooling submodule; The deep feature information representation submodule (CRN-DCN) is composed of a convolutional residual network (CRN) and a 3D dilated convolution (DCN); the specific connection method of the convolutional residual network is as follows: receiving the input of the convolutional residual network, and performing 3D convolution, 3D batch normalization layer, ReLu, 3D convolution, and 3D batch normalization layer on the input of the convolutional residual network in sequence to obtain the output result, and adding the obtained output result to the input of the convolutional residual network to obtain the output of the convolutional residual network; the specific connection method of the deep feature information representation submodule is as follows: the convolutional residual network is connected in series with the 3D dilated convolution and then connected in series with the convolutional residual network. Specifically, the first convolutional residual network receives the input of the deep feature information representation submodule to obtain the output of the first convolutional residual network, and the output of the first convolutional residual network is sequentially subjected to 3D dilated convolution and ReLu to obtain the first The output of the convolution residual network, 3D dilated convolution, and ReLu are combined, and the output of the first convolution residual network, 3D dilated convolution, and ReLu are taken as inputs, and input into the second convolution residual network to obtain the output of the second convolution residual network. Finally, the output of the second convolution residual network is added to the input of the deep feature information representation submodule to obtain the output of the deep feature information representation submodule. The technical advantage of the deep feature information representation submodule is that it can better extract data features and increase the receptive field. The source of the technical advantage is that the designed convolution residual network can effectively learn feature representation through residual connection, and 3D dilated convolution can increase the network's perception range, capture broader contextual information, and help the network learn more global and complex feature representations. The decoding submodule (Decoder) is composed of 3D convolution, trilinear interpolation and ReLu. The input of the decoding submodule is subjected to 3D convolution, trilinear interpolation and ReLu in sequence to obtain the output of the decoding submodule. There are four decoding submodules with the same structure, which are called the first decoding submodule, the second decoding submodule, the third decoding submodule and the fourth decoding submodule in the order of appearance. The output submodule consists of 3D convolution and softmax. The input of the output submodule is sequentially subjected to 3D convolution and softmax to obtain the output of the decoding submodule. The automatic segmentation module of the diseased area of ​​the CT image has submodules connected in the following manner: the first encoding submodule, the first pooling submodule, the second encoding submodule, the second pooling submodule, the third encoding submodule, the third pooling submodule, the fourth encoding submodule, the fourth pooling submodule, the deep feature information representation submodule, and the first decoding submodule are connected in series in sequence, and then the output of the fourth encoding submodule is spliced ​​to the tail of the output of the first decoding submodule to obtain the output of the fourth encoding submodule and the output of the first decoding submodule; then the output of the fourth encoding submodule and the output of the first decoding submodule are spliced ​​to the second decoding submodule to obtain the output of the second decoding submodule; then the output of the third encoding submodule is spliced ​​to the tail of the output of the second decoding submodule to obtain the output of the third encoding submodule and the output of the second decoding submodule. The output of the third encoding submodule and the output of the second decoding submodule are then input into the third decoding submodule to obtain the output of the third decoding submodule; the output of the second encoding submodule is then spliced ​​to the tail of the output of the third decoding submodule to obtain the output of the second encoding submodule and the output of the third decoding submodule; the output of the second encoding submodule and the output of the third decoding submodule are then input into the fourth decoding submodule to obtain the output of the fourth decoding submodule; the output of the fourth encoding submodule is then spliced ​​to the tail of the output of the first decoding submodule to obtain the output of the fourth encoding submodule and the output of the first decoding submodule, and the output of the fourth encoding submodule and the output of the first decoding submodule are input into the output submodule to obtain the output result of the automatic segmentation module of the diseased area of ​​the CT image; The automatic segmentation module of the diseased area of ​​the CT image is characterized by The loss function of the training network contains unique technical features: when training the network, the output target_pre of the automatic segmentation module of the diseased area of ​​the CT image is compared with target1, target2, and target3 to calculate the loss, wherein target1, target2, and target3 are respectively expanded by 1, 2, and 3 pixel values ​​on the basis of the diseased area of ​​the CT image and converted into tensors and divided into training set data; the advantage of this technical feature is that according to experimental observations, expanding the diseased area of ​​the CT image by 1, 2, and 3 pixel values ​​does not affect the analysis results of the outcome events. Therefore, in the deep learning algorithm, using the expansion of the diseased area of ​​the CT image as the basis for calculating the loss function can better improve the adaptability of the algorithm under different image quality conditions, improve the segmentation accuracy, and optimize the model training; specifically, the original segmented image target0 and target_pre are substituted into the loss4 formula to obtain loss4; The loss4 formula is: （1） Target1, target2, and target3 are respectively introduced into the target variable of the jac formula and target_pre to calculate jac. The obtained jac calculation results are recorded as loss1, loss2, and loss3. The final loss is ; The jac formula is: （2） （3） Where target refers to target1, target2, target3; CT image feature extraction module: extract the features of the diseased area of ​​each patient's CT image and name it feature, where feature contains all the features of the diseased area of ​​the patient's CT image. Each individual feature data in the features of the diseased area of ​​the patient's CT image is , , n is an integer greater than or equal to 1, according to the principle of dividing the training set and test set of the automatic segmentation module of the diseased area of ​​CT images, Data is divided into training set and test set ; Middle ear cholesteatoma feature screening module: First, perform a U test on the data to remove features with a p-value greater than 0.05; then use the Spearman correlation coefficient to calculate the correlation between features, and only retain one feature with a feature correlation coefficient greater than 0.8; finally, use the regularization parameter of the Lasso model (Lasso Regression Algorithm) to screen out the final features; Middle ear cholesteatoma differentiation module: The features of the diseased area of ​​the screened patient CT image are input into the multi-layer perceptron model for training, and the trained multi-layer perceptron model is saved, where the multi-layer perceptron model is a feedforward neural network model with multiple hidden layers and nonlinear activation functions.

2. The automatic detection system for middle ear cholesteatoma based on CT images as claimed in claim 1, wherein the CT image feature extraction module is implemented as follows: according to the diseased area of ​​the patient's CT image segmented by the automatic segmentation module of the diseased area of ​​the CT image, the features of the diseased area of ​​each patient's CT image are extracted; then, according to the principle of dividing the training set and the test set of the automatic segmentation module of the diseased area of ​​the CT image, each individual feature data in the features of the diseased area of ​​the patient's CT image is extracted. Also divided into training set and test set , and use The standard deviation and mean of Regularization is performed, and the specific formula is: （4） in is the first feature column, when When standardizing for , when When standardizing for , For unregularized Mean, Unregularized The standard deviation of .

3. The automatic detection system for middle ear cholesteatoma based on CT images as claimed in claim 1, characterized in that It includes a variety of feature screening techniques, specifically: first, the feature data of each separate training set in the features of the diseased area of ​​the patient's CT image is screened according to whether the patient is otitis media or cholesteatoma Perform U test ; Then use the Spearman correlation coefficient to calculate The correlation between ; Then use the Lasso model to The coefficients of some unimportant features are reduced to zero, and the features with zero coefficients are deleted to obtain the features ; The above three feature screening technologies act on the feature screening module of middle ear cholesteatoma; The advantages are that the U test and Lasso model can identify features that have a significant impact on the detection of middle ear cholesteatoma; the Spearman correlation coefficient can be used to calculate the correlation between features to reduce the impact of multicollinearity; and the combination of multiple feature selection methods is conducive to screening out representative features.

Images