Orthopedic joint replacement postoperative infection prediction method and system based on deep learning segmentation optimization
Through the combination of SE-U-Net and DenseNet201 models, the segmentation and classification of infected areas after orthopedic joint replacement surgery is optimized, and the problems of insufficient segmentation accuracy and high false negative rates in the existing technology are solved, and early accurate diagnosis and real-time early warning are achieved.
Patent Information
- Application Number
- CN202510506087.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-22
- Publication Date
- 2025-07-29
AI Technical Summary
The prior art has problems such as insufficient segmentation accuracy, low classification sensitivity and high false negative rate in the prediction of periprosthetic infection after orthopedic joint replacement, resulting in hysteresis diagnosis and highly invasive examination.
The SE-U-Net lesion segmentation module is used to enhance the edge feature weight of infected areas, and combined with the DenseNet201 classification model, the segmentation and classification of infected areas are optimized by dynamically adjusting the cross entropy loss weight and adaptive threshold.
It improves the accuracy of lesion boundary extraction, reduces the risk of false negatives, realizes early accurate diagnosis and real-time early warning, and reduces postoperative remedial treatment.
Smart Images
Figure CN120388738A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of medical image processing, and particularly to a method and system for predicting infection after orthopedic joint replacement based on deep learning segmentation optimization. Background Art
[0002] Currently, the prediction of periprosthetic joint infection (PJI) after orthopedic joint replacement mainly relies on artificial experience. When identifying early tiny lesions such as soft tissue edema and abnormal bone marrow signals, it is extremely susceptible to the subjective experience of the radiologist, and there are significant differences in the judgment results of different radiologists. Moreover, it is difficult to quantify the infection risk through artificial analysis, and real-time early warning cannot be provided before surgery, resulting in a relatively high proportion of postoperative remedial treatments and obvious lag.
[0003] In addition, traditional medical image segmentation methods, such as traditional U-Net, are prone to ignoring the weight of lesion edge features during skip connections and perform poorly in dealing with the situation where the contrast between the perijoint lesion area and normal tissues is low, resulting in blurred segmentation boundaries; conventional threshold segmentation methods have low sensitivity to tiny infection foci and high miss rate.
[0004] Furthermore, due to the fact that the proportion of infected samples in PJI clinical data is usually less than 3%, the problem of class imbalance in medical image classification models is serious. The directly trained model tends to predict negative results, and a fixed classification threshold such as 0.5 is likely to lead to a high false negative rate, delaying clinical intervention.
[0005] Although the incidence of PJI after artificial joint replacement is only 1%-3%, it is a serious complication. The traditional diagnosis mode based on clinical manifestations and laboratory tests has a lag of several weeks to several months. Arthrocentesis has the risk of invasion, and false negative problems are prone to occur due to low contrast in early imaging signs, making early and accurate diagnosis difficult. Although preoperative MRI contains rich information, traditional image analysis methods have problems such as low efficiency of manual image reading, insufficient ability to identify tiny lesions, and insufficient mining of the association of multi-modal image data. Summary of the Invention
[0006] In the first aspect of the present invention, in order to solve the above technical problems, a method for predicting infection after orthopedic joint replacement based on deep learning segmentation optimization is provided. The method includes: Obtaining preoperative multi-modal MRI image data of a patient and performing preprocessing; the multi-modal MRI images include coronal, sagittal, and transverse images of T1-weighted and T2-weighted. Input the preprocessed MRI images into the SE-U-Net lesion segmentation module to output the ROI mask of the infected area of the MRI images; wherein, the SE-U-Net lesion segmentation module embeds an SE module at the skip connection between the encoder and the decoder, and dynamically enhances the edge feature weights of the infected area through the channel attention mechanism; Overlay the ROI mask with the original MRI images and input them into the DenseNet201 classification model to output the infection probability value; wherein, based on the sample ratio of infection and non-infection in the training set, automatically adjust the cross-entropy loss weight; based on the ROC curve of the validation set, calculate the dynamic classification threshold corresponding to the maximum value of the Youden Index; generate the infection risk level according to the comparison result between the infection probability value and the threshold; Visually output the infection probability value, the risk level, and the ROI boundary heat map to assist clinical decision-making.
[0007] Further, the preprocessing of the multi-modal MRI image data includes: Extract the pixel array of the MRI image and apply window width and window level parameters to retain the contrast features of the original image; Independently perform standardization processing on each slice to complete the gray-scale normalization operation; Use the cubic interpolation method to uniformly scale the multi-modal MRI images to the specified pixel resolution.
[0008] Further, the SE-U-Net lesion segmentation module adopts a composite loss function of Dice Loss and edge weighted loss to optimize the boundary segmentation accuracy of the low-contrast infected area, and the composite loss function satisfies the expression:
[0009] In the formula, Take 0.7, Take 0.3, and the two are determined by cross-validation; is Dice Loss; is the edge weighted loss.
[0010] Further, the optimization mechanism of the skip connection includes: Perform global average pooling on the feature map output by the encoder to compress the feature map and generate a channel description vector; Generate a channel weight vector through a fully connected layer and an activation function; Multiply the original feature map and the channel weight vector channel by channel for recalibration of the edge features of the infected area.
[0011] Further, the dimension design of the fully connected layer of the SE module is: The first fully connected layer compresses the number of channels to 1 / 16 of the original channels; the second fully connected layer restores to the original number of channels; and their activation functions are ReLU and Sigmoid in sequence.
[0012] Furthermore, the method for determining the dynamic classification threshold includes: Calculating the sensitivity and false positive rate at different thresholds in the validation set; Taking the difference between the sensitivity and the false positive rate, and selecting the threshold corresponding to the maximum value of the Youden Index to reduce the false negative rate.
[0013] Furthermore, the DenseNet201 classification model further includes multi-view fusion of the multi-modal MRI images, and the fusion method includes: Registering the coronal, sagittal and transverse images to the same three-dimensional coordinate system through affine transformation; Extracting feature vectors from the ROI images of each view respectively, concatenating them and inputting them into a fully connected layer for dimensionality reduction, and outputting the comprehensive infection probability.
[0014] Furthermore, the SE-U-Net lesion segmentation module shares the GPU memory with the DenseNet201 classification model and is accelerated by TensorRT to shorten the inference time.
[0015] In the second aspect of the present invention, there is provided an orthopedic joint replacement postoperative infection prediction system based on deep learning segmentation optimization, including: A multi-modal input module; used to acquire and structurally store T1-weighted and T2-weighted coronal, sagittal and transverse MRI images, and store the DICOM format data according to the patient ID, scan type and view directory tree; A preprocessing module, used to perform gray normalization, resolution unification and contrast retention processing on the MRI images; An SE-U-Net lesion segmentation module, used to accurately segment the infected area and output an ROI mask; A DenseNet201 classification model, configured to extract key features of the infected area and output an infection probability value; and A visualization interaction module, configured to dynamically display the infection probability, risk level and lesion heat map.
[0016] Furthermore, the SE-U-Net lesion segmentation module further includes a dynamic classification decision unit. The dynamic classification decision unit calculates dynamic class weights, automatically adjusts the cross-entropy loss weight based on the ratio of infected samples to non-infected samples in the DenseNet201 classification model training set; and uses Youden Index optimal threshold selection to calculate the balance point of sensitivity and specificity based on the ROC curve in the DenseNet201 classification model validation set.
[0017] Compared with the prior art, the embodiments of the present invention have the following beneficial effects: Through the boundary optimization segmentation technology of SE-U-Net, the present invention directly solves the problem of insufficient accuracy in extracting lesion boundaries by traditional segmentation methods; at the same time, by adopting a dynamic weight compensation and adaptive threshold classification mechanism, it can focus on the characteristics of the infected area and maximize the alleviation of the data imbalance between sensitivity and specificity, reducing the risk of false negatives. BRIEF DESCRIPTION OF THE DRAWINGS
[0018] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the drawings in the following description are only some embodiments of the present invention, and those of ordinary skill in the art can also obtain other drawings based on these drawings without creative efforts.
[0019] Figure 1 It is a flowchart disclosed in the embodiments of the present invention; Figure 2 It is a processing flowchart of the MRI image disclosed in the embodiments of the present invention; Figure 3 It is a curve graph of prediction effect indicators after introducing a dynamic class weight compensation mechanism and embedding an SE module in the embodiments of the present invention; Figure 4 It is an interactive interface and core result display graph disclosed in the embodiments of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS
[0020] In order to enable those skilled in the art to better understand the solution of the present invention, the following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.
[0021] In view of the problems such as insufficient segmentation accuracy and low classification sensitivity in the prediction of PJI before orthopedic joint replacement, the present invention provides a method for predicting postoperative infection of orthopedic joint replacement based on deep learning segmentation optimization. Please refer to Figure 1 and Figure 2 , which mainly includes the following steps: S1. Obtain the preoperative multimodal MRI image data of the patient and perform preprocessing. Among them, the multimodal MRI images include coronal, sagittal, and transverse images of T1-weighted and T2-weighted.
[0022] In this embodiment, the multimodal MRI images are stored according to a hierarchical directory tree. The root directory is the patient ID, and the subdirectories are successively the infection identifier (infected / non-infected), the scan type (T1 or T2), the perspective (coronal, sagittal, transverse), and the bottom layer is the DICOM file named according to the slice order to ensure the structured reading of multimodal data.
[0023] The preprocessing of the multimodal MRI image data includes: S11. Extract the pixel array of the MRI image and apply window width and window level parameters to retain the contrast features of the original image, so as to achieve DICOM parsing and standardization.
[0024] S12. Independently perform standardization processing on each slice to complete the gray normalization operation.
[0025] S13. In terms of spatial sampling, use the cubic interpolation method to uniformly scale the multimodal MRI images to the specified pixel resolution to eliminate the resolution differences between devices.
[0026] S2. Input the preprocessed MRI image into the SE-U-Net lesion segmentation module to output the ROI mask of the infected area of the MRI image. Among them, the SE-U-Net lesion segmentation module embeds the SE module at the skip connection between the encoder and the decoder, and dynamically enhances the edge feature weights of the infected area through the channel attention mechanism.
[0027] In a further embodiment of the present invention, the hierarchical design of the encoder structure of the SE-U-Net lesion segmentation module is as follows: the encoder includes 4 levels, each level consists of two 3×3 convolutional layers, the stride is 1, the padding mode is "same", and the output feature map size remains unchanged. The activation function uses ReLU.
[0028] Its downsampling design is as follows: at the end of each level, the feature map size is halved through 2×2 max pooling, and the number of channels is doubled, such as 64 channels in the first layer → 128 channels in the second layer.
[0029] In this embodiment, the optimization mechanism of the skip connection includes: S21. Perform global average pooling on the feature map output by the encoder to compress the feature map and generate a channel description vector.
[0030] S22. Generate a channel weight vector through a fully connected layer and an activation function.
[0031] S23. Multiply the original feature map and the channel weight vector channel by channel for recalibration of the edge features of the infected area.
[0032] Among them, the dimension design of the fully connected layer of the SE module is that the first fully connected layer compresses the number of channels to 1 / 16 of the original channels; the second fully connected layer restores to the original number of channels; its activation functions are ReLU and Sigmoid in sequence.
[0033] Those skilled in the art further explain that the SE-U-Net lesion segmentation module integrates the Squeeze-and-Excitation (SE) module on the basis of the traditional U-Net to enhance the correlation between feature channels, thereby improving the segmentation effect of the region of interest (ROI) of the lesion. The core goal of its SE module is to weight the features in the channel dimension so that the model can focus on the "useful channels" and ignore redundant information.
[0034] The technical principle of the SE module is: compress each channel into a scalar; learn the importance weights of each channel through two fully connected layers; then apply the importance weights to the original feature map to strengthen the channel attention mechanism.
[0035] Specifically: The Squeeze stage of the SE module structure: Perform global average pooling on the feature map output by the encoder to generate a channel description vector.
[0036] The Excitation stage of the SE module structure: The first fully connected layer compresses the number of channels to , the activation function is ReLU; the second fully connected layer restores to the original number of channels , the activation function is Sigmoid, generating a channel weight vector , which satisfies the expression:
[0037] In the formula, , are the weights of the fully connected layer; is the ReLU function; is the Sigmoid function.
[0038] Among them, feature recalibration is to multiply the original feature map and the weight vector channel by channel to obtain the calibrated feature map , 。
[0039] In a further solution of this embodiment, the hierarchical design of the decoder structure of the SE-U-Net lesion segmentation module is as follows: in its upsampling, the decoder doubles the size of the feature map through bilinear interpolation at each level, and then connects two 3×3 convolutional layers. Its skip connection is to splice the calibrated feature map of the corresponding level of the encoder with the upsampling result of the decoder to form a fused feature tensor and input it into the subsequent convolutional layer.
[0040] The SE-U-Net lesion segmentation module adopts a composite loss function of Dice Loss and edge-weighted loss to optimize the boundary segmentation accuracy of low-contrast infection regions. Among them, Dice Loss is used to measure the overlap degree between the predicted mask and the true label; the edge-weighted loss strengthens the sensitivity of the model to the edge of the infection region by calculating the gradient difference of the boundary pixels; its composite loss function is the weighted sum of the two:
[0041] In the formula, takes 0.7, takes 0.3, and the two are determined by cross-validation; is Dice Loss; is the edge-weighted loss.
[0042] The significance of the SE module for the ROI of medical images in this embodiment is as follows: due to the weak feature differences and easy noise interference of medical images, based on this, the SE module is introduced in all key layers, which can effectively strengthen the channel-level attention mechanism, improve the attention of the model to the "lesion-related channels", and thus enhance the accuracy of lesion segmentation.
[0043] S3. Superimpose the ROI mask and the original MRI image and input them into the DenseNet201 classification model to output the infection probability value; among them, based on the sample ratio of infection and non-infection in the training set, automatically adjust the cross-entropy loss weight; based on the ROC curve of the validation set, calculate the dynamic classification threshold corresponding to the maximum value of the Youden Index; generate the infection risk level according to the comparison result between the infection probability value and the threshold.
[0044] In this embodiment, the binary ROI mask output by the SE-U-Net lesion segmentation module is multiplied by the original image pixel by pixel to retain the characteristics of the infection region. Convert the single-channel mask image into a 3-channel RGB image to adapt to the input requirements of the pre-trained DenseNet201 (224×224×3).
[0045] The method for determining the dynamic classification threshold in this embodiment includes: calculating the sensitivity and false positive rate at different thresholds in the validation set; subtracting the false positive rate from the sensitivity, and selecting the threshold corresponding to the maximum Youden Index to reduce the false negative rate. Specifically, it includes the following methods: Regarding dynamic weight compensation: According to the quantity ratio of infected samples (positive class) to non-infected samples (negative class) in the training set, dynamically adjust the cross-entropy loss weight: , where , are the numbers of positive and negative class samples respectively; its cross-entropy loss satisfies the expression:
[0046] Regarding adaptive threshold classification: Draw an ROC characteristic curve on the independent validation set, traverse the threshold , with a step size of 0.01, and calculate the sensitivity and false positive rate at each threshold; then select the threshold corresponding to the maximum Youden Index: .
[0047] The area under the precision-recall curve (AUC) of this model is 0.9942, the optimal threshold is 0.1308, the recall rate of the infection category is 0.93, and the F1 score is 0.91. Such a high AUC score is of great significance for a clinical decision support system, because in a clinical decision support system, correctly identifying infected and non-infected cases with a high degree of certainty will have a significant impact on the treatment effect of patients. This threshold maximizes the sensitivity and specificity, achieving an ideal balance between the two. The high recall rate ensures that most PJI patients can be correctly labeled for further clinical examination and intervention. The high F1 score indicates that the model has achieved a good balance between correctly identifying infected cases (high recall rate) and ensuring the accuracy of its positive predictions (high precision). Please refer to Figure 3 .
[0048] S4. Visually output the infection probability value, risk level, and the heat map of the ROI boundary to assist clinical decision-making. Please refer to Figure 4 .
[0049] In this embodiment, the infection probability value is reserved to two decimal places; the risk level is divided into low (<0.3), medium (0.3 - 0.7), and high (>0.7); in the heat map of the ROI boundary, the ROI mask is superimposed on the original MRI image in semi-transparent red, and the color depth is positively correlated with the infection probability.
[0050] When the infection probability exceeds the dynamic threshold, the system automatically pushes optimization suggestions for the surgical plan, such as extending the antibiotic course or early revision surgery.
[0051] The DenseNet201 classification model also includes multi-view fusion of multi-modal MRI images, and its fusion method includes: Register the coronal, sagittal, and transverse images to the same three-dimensional coordinate system through affine transformation; then input the ROI masks of each view into the DenseNet201 classification model respectively. After extracting the feature maps, generate a 1×1×C feature vector through global average pooling; splice the multi-view feature vectors into a high-dimensional vector and input it into the fully connected layer to generate the comprehensive infection probability.
[0052] In this embodiment, the SE-U-Net lesion segmentation module and the DenseNet201 classification model are coupled through the ROI mask, that is, the ROI mask generated by the SE-U-Net lesion segmentation module will be passed as input to the DenseNet201 classification model. This coupling method enables the two models to cooperate closely and form an overall processing flow. The DenseNet201 classification model can perform more targeted classification predictions based on these real-time generated ROI masks, avoiding unnecessary analysis of the entire image, thereby improving the accuracy and efficiency of the prediction. At the same time, the lightweight design makes the entire coupled model more compact and easy to deploy and run.
[0053] In addition, the SE-U-Net lesion segmentation module is mainly used for lesion segmentation, and the DenseNet201 classification model is used for subsequent classification prediction. In this embodiment, the SE-U-Net lesion segmentation module and the DenseNet201 classification model share the GPU memory, and realize parallel processing of segmentation and classification through an asynchronous pipeline, which is beneficial to reducing the overall calculation time and cost; and is accelerated by TensorRT to shorten the inference time.
[0054] The present invention also protects an orthopedic joint replacement postoperative infection prediction system based on deep learning segmentation optimization using the above method, which mainly includes a multi-modal input module, a preprocessing module, an SE-U-Net lesion segmentation module, a DenseNet201 classification model, and a visualization interaction module, wherein: The multi-modal input module is used to acquire and structurally store the coronal, sagittal, and transverse MRI images of T1-weighted and T2-weighted, and store the DICOM format data according to the patient ID, scan type, and view directory tree.
[0055] The preprocessing module is used to perform gray normalization, resolution unification, and contrast retention processing on the MRI images.
[0056] The SE-U-Net lesion segmentation module is used to accurately segment the infected area and output the ROI mask.
[0057] The DenseNet201 classification model is configured to extract key features of the infected area and output the infection probability value.
[0058] The visualization interaction module is configured to dynamically display the infection probability, risk level, and lesion heat map.
[0059] In a further solution of this embodiment, the SE-U-Net lesion segmentation module further includes a dynamic classification decision unit. Among them, the dynamic classification decision unit uses dynamic class weight calculation to automatically adjust the cross-entropy loss weight based on the ratio of infected samples to non-infected samples in the DenseNet201 classification model training set; and uses Youden Index optimal threshold selection to calculate the balance point of sensitivity and specificity based on the ROC curve in the DenseNet201 classification model validation set.
[0060] Although the embodiments of the present invention have been shown and described, those of ordinary skill in the art can understand that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.
Claims
1. An orthopedic joint replacement postoperative infection prediction method based on deep learning segmentation optimization, characterized in that, The method includes: Obtaining preoperative multimodal MRI image data of the patient and performing preprocessing; the multimodal MRI images include coronal, sagittal, and transverse images of T1-weighted and T2-weighted; Inputting the preprocessed MRI images into the SE-U-Net lesion segmentation module to output the ROI mask of the infected area of the MRI images; wherein, the SE-U-Net lesion segmentation module embeds an SE module at the skip connection between the encoder and the decoder, and dynamically enhances the edge feature weights of the infected area through the channel attention mechanism; Overlaying the ROI mask with the original MRI images and inputting them into the DenseNet201 classification model to output the infection probability value; wherein, based on the sample ratio of infection and non-infection in the training set, the cross-entropy loss weight is automatically adjusted; based on the ROC curve of the validation set, the dynamic classification threshold corresponding to the maximum value of the Youden Index is calculated; according to the comparison result between the infection probability value and the threshold, the infection risk level is generated; Visually outputting the infection probability value, the risk level, and the ROI boundary heat map to assist clinical decision-making.
2. The method for predicting orthopedic joint replacement postoperative infection based on deep learning segmentation optimization according to claim 1, wherein The preprocessing of the multimodal MRI image data includes: Extracting the pixel array of the MRI image and applying window width and window level parameters to retain the contrast features of the original image; Independently performing normalization processing on each slice to complete the gray normalization operation; Using cubic interpolation to uniformly scale the multimodal MRI images to the specified pixel resolution.
3. The method for predicting orthopedic joint replacement postoperative infection based on deep learning segmentation optimization according to claim 1, wherein The SE-U-Net lesion segmentation module adopts a composite loss function of Dice Loss and edge weighted loss to optimize the boundary segmentation accuracy of low-contrast infected areas, and the composite loss function satisfies the expression: In the formula, takes 0.7, takes 0.3, and the two are determined by cross-validation; is the Dice Loss; is the edge weighted loss.
4. The method for predicting infection after orthopedic joint replacement based on deep learning segmentation optimization according to claim 1, wherein The optimization mechanism of the skip connection includes: Performing global average pooling on the feature map output by the encoder to compress the feature map and generate a channel description vector; Generating a channel weight vector through a fully connected layer and an activation function; Multiplying the original feature map and the channel weight vector channel by channel for recalibration of the edge features of the infected area.
5. The method for predicting orthopedic joint replacement postoperative infection based on deep learning segmentation optimization according to any one of claims 1-4, characterized in that, The dimension design of the fully connected layer of the SE module is: The first fully connected layer compresses the number of channels to 1 / 16 of the original channels; the second fully connected layer restores to the original number of channels; its activation functions are ReLU and Sigmoid in sequence.
6. The method for predicting postoperative infection of orthopedic joint replacement based on deep learning segmentation optimization according to claim 1, wherein The method for determining the dynamic classification threshold includes: Calculating the sensitivity and false positive rate at different thresholds in the validation set; Taking the difference between the sensitivity and the false positive rate, and selecting the threshold corresponding to the maximum value of the Youden Index to reduce the false negative rate.
7. The method for predicting orthopedic joint replacement postoperative infection based on deep learning segmentation optimization according to claim 1, wherein The DenseNet201 classification model also includes multi-view fusion of the multimodal MRI images, and its fusion method includes: Registering the coronal, sagittal, and transverse images to the same three-dimensional coordinate system through affine transformation; Respectively extracting feature vectors from the ROI images of each view, splicing them and inputting them into a fully connected layer for dimensionality reduction, and outputting the comprehensive infection probability.
8. The method for predicting orthopedic joint replacement postoperative infection based on deep learning segmentation optimization according to any one of claims 1-7, characterized in that, The SE-U-Net lesion segmentation module shares GPU memory with the DenseNet201 classification model and is accelerated by TensorRT to shorten the inference time.
9. An orthopedic joint replacement postoperative infection prediction system based on deep learning segmentation optimization, characterized in that, It includes: A multimodal input module; It is used to obtain and structurally store coronal, sagittal, and transverse MRI images of T1-weighted and T2-weighted, and store DICOM format data according to the patient ID, scan type, and perspective directory tree. A preprocessing module, which is used to perform gray normalization, resolution unification, and contrast retention processing on the MRI images. An SE-U-Net lesion segmentation module, which is used to accurately segment the infected area and output an ROI mask. A DenseNet201 classification model, configured to extract key features of the infected area and output an infection probability value. And A visualization interaction module, configured to dynamically display the infection probability, risk level, and lesion heat map.
10. The orthopedic joint replacement postoperative infection prediction system based on deep learning segmentation optimization according to claim 9, wherein, The SE-U-Net lesion segmentation module further includes a dynamic classification decision unit. The dynamic classification decision unit uses dynamic class weight calculation to automatically adjust the cross-entropy loss weight based on the ratio of infected samples to non-infected samples in the DenseNet201 classification model training set; and uses Youden Index optimal threshold selection to calculate the balance point of sensitivity and specificity based on the ROC curve in the DenseNet201 classification model validation set.