Knee osteoarthritis auxiliary diagnosis method based on semi-supervised learning and double knowledge distillation

Abstract

The invention discloses a knee osteoarthritis auxiliary diagnosis method based on semi-supervised learning and double knowledge distillation, and the method comprises the steps: carrying out personalized pixel shuffling: dynamically recognizing a knee osteoarthritis diagnosis key region and a privacy sensitive region according to a KL grading standard and an OARSI guide, adjusting a privacy strategy through visual feature entropy based on KL grading, and carrying out personalized pixel shuffling; high-intensity pixel disturbance is applied to the knee osteoarthritis diagnosis key area, the edge gradient is reserved, and low-intensity disturbance is adopted for the non-KL grading key area; attention-guided distillation: embedding a pre-trained bone geometric morphology convolution kernel, strengthening osteophyte edge feature response, and applying 1.5 times of attention weight to a KL3-4-level bony contact area; and consistency pseudo-tag optimization: after the teacher and the bridging network generate pseudo-tags, screening and retaining the pseudo-tags meeting the consistency condition for student model training to obtain a classification auxiliary result of knee osteoarthritis. According to the invention, the problems of scarcity of medical image data and privacy protection are solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent analysis and processing of medical images, and in particular to a knee osteoarthritis auxiliary diagnosis method based on semi-supervised learning and double knowledge distillation, which is particularly suitable for solving the difficulties in KOA clinical diagnosis, the scarcity of labeled data, and the privacy protection needs of patients. BACKGROUND

[0002] KOA is a common and highly disabling musculoskeletal disease that mainly affects the elderly population, but with the development of society, there are also more and more patients in the young population. The characteristic of KOA is the gradual degeneration of articular cartilage, which leads to pain, stiffness and limited mobility of the knee joint, thereby significantly affecting daily life and overall quality of life. As a chronic degenerative disease, KOA has become one of the main causes of disability worldwide, especially in an aging society, its burden is increasingly severe. The progression of the disease is usually accompanied by joint inflammation and cartilage damage, and in severe cases, it can lead to changes in joint morphology, further exacerbating dysfunction. Therefore, accurate diagnosis and treatment of the disease are important factors in restoring the health of patients.

[0003] However, resource-poor countries often face a shortage of medical resources, including limited professional doctors, equipment and treatment sites. The diagnosis and treatment of KOA require advanced medical facilities and equipment such as X-ray, MRI, CT and other imaging examinations, but in these countries, many areas may not be able to provide professional imaging examination services. Medical institutions often lack the professional knowledge of doctors, and the experience of the radiologists is insufficient, leading to difficulties in accurate diagnosis of KOA, and even misdiagnosis or missed diagnosis, many patients may not be able to obtain correct diagnosis and treatment in time.

[0004] In recent years, with the widespread application of deep learning and self-supervised artificial intelligence technology, especially in the field of medical image analysis and disease diagnosis, these technologies have shown great potential. Self-supervised artificial intelligence can perform complex tasks without human supervision, such as automatically detecting and classifying abnormal conditions in the knee joint area, assisting in improving the efficiency and accuracy of diagnosis. This not only improves the accessibility of medical services, but also enables remote medical care and diagnosis in resource-poor areas to be implemented and promoted. SUMMARY

[0005] The purpose of the present application is to provide a knee osteoarthritis auxiliary diagnosis method based on semi-supervised learning and double knowledge distillation, which combines self-supervised artificial intelligence technology, solves the problems of medical image data scarcity and privacy protection through self-optimization and real-time decision-making.

[0006] The technical solution adopted by the present application is:

[0007] The knee osteoarthritis auxiliary diagnosis method based on semi-supervised learning and double knowledge distillation comprises the following steps:

[0008] S1, personalized pixel shuffling based on knee osteoarthritis (KOA) pathological characteristics: according to the KL grading standard and OARSI guidelines, dynamically identify the key area and privacy sensitive area of knee osteoarthritis diagnosis, and adjust the privacy policy based on the visual feature entropy (VFE) of KL grading to retain the key features of diagnosis;

[0009] Further, the knee osteoarthritis diagnosis key area in S1 includes the femoral condyle osteophyte area with KL≥2 and osteophyte area≥10%, and the tibiofemoral joint space area with JSW≤3mm.

[0010] Further, the dynamic adjustment of the privacy policy based on the KL grading in S1 is specifically: high-intensity pixel disturbance is applied to the knee osteoarthritis diagnosis key area while retaining the edge gradient; low-intensity disturbance is used for non-KL grading key area to avoid interference with inflammatory features.

[0011] Specifically, the pixel disturbance intensity λ is dynamically adjusted based on the KL grading: when the femoral condyle edge osteophyte coverage area is detected to be≥15% (i.e. KL 3 level threshold), the pixel disturbance intensity λ is increased to 0.7; when the tibiofemoral joint space width is≤2.5mm (KL 4 level typical indication), median filtering (window size 5x5) is superimposed to eliminate patient identification features in the bony fusion area; low-intensity disturbance (λ=0.2) is used for the suprapatellar bursa synovial membrane hyperplasia area (ultrasound confirmed synovial membrane thickness≥4mm) to retain the texture features of pannus formation of synovial membrane and avoid interference with inflammatory features. In addition, by introducing personalized parameters, researchers or developers can customize the protection level of each area for different application scenarios.

[0012] S2, attention guided distillation based on knee osteoarthritis (KOA) pathological prior: embedding a pre-trained bone geometric morphological convolution kernel (OAI dataset osteophyte parameter initialization), strengthening the feature response of the osteophyte edge (KL≥2 level indication), and applying 1.5 times attention weight to the KL 3-4 level bony contact area;

[0013] Further, S2 specifically comprises the following steps:

[0014] S20, generating a spatial attention map: for a single-channel X-ray image of KOA, the absolute value of each spatial dimension feature map is calculated, and then average pooling (or maximum pooling) is performed at each pixel position to create a spatial attention map;

[0015] Further, the spatial attention map A space is calculated as follows:

[0016] A space =Pool(|I o |) (1);

[0017] wherein Pool represents the pooling method as average pooling. o ∈R H×W denotes the input image, H and W represent the height and width of the image, respectively.

[0018] Specifically, for the single-channel X-ray image of KOA, a spatial attention map can be generated by processing the spatial dimensions (height and width) of the image. Generally, this involves calculating the absolute value of each spatial dimension feature map, and then performing average pooling (or maximum pooling) at each pixel position to create a spatial attention map. For the input image I o ∈R H×W , wherein H and W represent the height and width of the image, respectively, and the spatial attention map A space is calculated as follows:

[0019] A space =Pool(|I o |) (1);

[0020] wherein Pool is the pooling method, and average pooling is used.

[0021] S21, normalization: using a temperature parameterized softmax function to normalize the spatial attention maps of the teacher and student.

[0022] Further, the specific expression of normalization is as follows:

[0023]

[0024] wherein T is a temperature hyperparameter used to control the smoothness of the softmax function; role represents the teacher or student, which means that role∈{teacher,student}; A teacher is the spatial attention map of the teacher model, and A student is the spatial attention map of the student model.

[0025] The temperature parameter T is dynamically adjusted according to the degree of patellofemoral joint alignment abnormality. When the patellar tilt angle (PTA) is ≥10° (calculated by Hough transform to detect the angle between the lower pole connecting line of the patella and the axis of the femoral trochlear groove), T is reduced from 1.0 to 0.5 to strengthen feature focusing; when the tibial plateau sagittal angle (PSA) is ≥7°, a 1.5 times weight coefficient is applied to the subchondral bone sclerosis area in the normalization process.

[0026] S22, calculate the total loss of attention-guided distillation: merge and average the normalized attention maps of the teacher model and the student model to generate the final attention mask M; the calculation expression of the average of the two normalized spatial attention maps is as follows:

[0027]

[0028] wherein, denotes the normalized spatial attention map of the teacher model; denotes the normalized spatial attention map of the student model;

[0029] S23, based on the spatial attention mask M, weight the pixel loss of different positions according to the importance of pixels to obtain the feature distillation loss; the expression of the feature distillation loss L mask is as follows:

[0030] L mask =∑ i,j M(i,j)·L student (i,j) (4);

[0031] wherein, L student (i,j) is the loss of the student model at position (i,j), and M(i,j) is the attention mask value at the position.

[0032] S24, calculate the spatial attention loss for encouraging the student model to simulate the teacher model, and the specific expression is as follows:

[0033] L atten =λ·‖A teacher -A student ‖2 (5);

[0034] wherein, λ is a hyperparameter for balancing the weight between the distillation loss and the attention loss.

[0035] S25, integrate the feature distillation loss and the spatial attention loss for encouraging the student model to simulate the teacher model to obtain the distillation loss L attendis under the guidance of attention, and the specific expression is as follows:

[0036] L attendis =L mask +L atten (6)。

[0037] Specifically, in traditional semi-supervised learning, the student model is trained by the pseudo-label generated by the teacher model, but the quality of the pseudo-label is often affected by the performance of the teacher model, especially when the teacher model is wrong or inaccurate, the quality of the pseudo-label will decrease significantly, thereby affecting the learning effect of the student model. By introducing an attention mechanism, attention-guided knowledge distillation can help the student model focus on the teacher model's learning in key areas or important features, rather than relying directly on the global output of the teacher model. For the KL grading requirements of KOA, the pre-trained Hessian-Affine Convolutional Kernel (OAI dataset bone osteophyte morphology parameter initialization) is embedded to strengthen the feature response of the osteophyte edge (KL≥2 core indicators); a 1.5 times attention weight is applied to the bone contact area of KL3-4 grade to guide the model to focus on the key area of cartilage degeneration; when JSW≤2.5mm (KL4 grade indicator), a gap narrow compensation factor γ=0.8 is introduced to optimize the accuracy of joint space measurement.

[0038] S3, consistent pseudo-label optimization according to ACR guidelines: after the teacher and the bridging network generate the pseudo-label, the pseudo-label that meets the consistency condition is screened and retained, which is used for student model training to output classification auxiliary results of knee osteoarthritis (KOA);

[0039] Further, the consistency condition is set according to the ACR guidelines, specifically: the KL grading difference of the teacher and the bridging network is ≤1 grade, the JSW deviation is ≤0.5mm, and the osteophyte area error is ≤15%.

[0040] Further, the consistency module in S3 includes the following steps:

[0041] S30, calculating the loss function of the bridging network; first, input the unlabeled dataset D W ={I W ,C W} into the teacher network to obtain the output classification summary information C T of the teacher network. Then, filter the pseudo-label using the threshold value. At the same time, input the same input dataset into the bridging network for training to generate the output C B . The pseudo-label generated by the teacher network can also guide the learning of the bridging network and optimize its performance. The loss function L bridge of the bridging network is as follows:

[0042]

[0043] S31, consistent pseudo-label generation: according to the classification information summary C T and C B , generate the final consistent pseudo-label Specifically, if the classification results of two images are the same (i.e., their class labels are the same), they are considered consistent and are taken as candidate pseudo-labels The core of this process is to extract more reliable pseudo-labels through consistency comparison, and the specific method is

[0044]

[0045] The consistency determination needs to meet the following conditions at the same time: the KL grading difference is less than or equal to 1 grade, the tibiofemoral joint space width (JSW) measurement deviation is less than or equal to 0.5mm, and the relative error of osteophyte coverage area is less than or equal to 15%.

[0046] S32, final pseudo-label generation and loss calculation: next, and merge to form the final pseudo-label to guide the training of the student network. The final unsupervised loss is the cross-entropy loss between the student model output C S and , as shown in the following formula:

[0047]

[0048] Finally, the total loss function is as follows:

[0049] L total = L sup + L unsup + L bridge + L attendis (10);

[0050] wherein, L bridge represents the loss function of the bridge network, L unsup represents the cross-entropy loss between the student model output classification summary information C S and the final pseudo-label ; L attendis represents the attention-guided distillation loss; L sup represents the supervised loss of the student model pre-training, which is calculated by cross-entropy loss.

[0051] Further, the final output includes: KL grading diagnosis result (0-4 grade); tibiofemoral joint space narrowing degree grading (normal / mild / moderate / severe); osteophyte area ratio (accurate to 0.1%); treatment suggestion generation (conservative treatment / arthroscopic cleaning / artificial joint replacement).

[0052] Specifically, in the semi-supervised teacher-student network, the pseudo labels with low confidence are usually filtered out by setting a high confidence threshold to reduce the negative impact of low-quality labels on the training of the student network. However, in the early stage of training, the teacher model may not have reached a high accuracy, which makes it risky to rely only on high-confidence pseudo labels to train the student model. To address the problem of KL grading pseudo label noise, a bridge network is introduced and combined with ACR guidelines to set consistency conditions: the KL grading difference between the teacher and the bridge network is less than or equal to 1 level, the JSW deviation is less than or equal to 0.5 mm, and the osteophyte area error is less than or equal to 15%. Only pseudo labels that meet the conditions are retained (error rate is reduced from 20% to 8%), which improves the reliability of KL grading.

[0053] The above technical solutions are adopted in the present application. Personalized pixel shuffling allows users to flexibly define the size and strength of the privacy protection area according to specific needs and task objectives. By introducing personalized parameters, the protection level of each area can be customized for different application scenarios. Attention-guided distillation helps the student model focus on the teacher model's learning in key areas or important features, rather than relying directly on the teacher model's global output. Attention guidance can help the student model concentrate on learning key areas or important information in the image, thereby improving the effectiveness of pseudo labels. The consistency module. In the early stage of training, the teacher model has low accuracy, and relying only on high-confidence pseudo labels is risky, which may generate false high-confidence pseudo labels and affect the performance of the final model. Therefore, a bridge network is introduced to improve the accuracy of pseudo labels by comparing them with the pseudo labels generated by the teacher network.

[0054] This invention employs the semi-supervised learning framework PADisitillation, using a teacher-bridge-student model double distillation technique to improve KOA medical image classification technology. This minimizes the need for large labeled datasets, enabling effective training with limited data. Furthermore, integration with telemedicine improves remote diagnosis, enhances the accuracy of KOA diagnoses, and protects patient privacy in resource-constrained environments. This invention also improves matching between datasets of varying quality through a personalized pixel shuffling method, ensuring data privacy. Finally, this invention uses an attention-guided distillation strategy, encouraging student models to focus on important regions or features learned by the teacher model, rather than relying on global pseudo-labels. The model framework constructed by the method of this invention performs excellently in classifying the severity of knee osteoarthritis. In multicenter validation on the OAI dataset: 89.7% of KL2-3 transitional cases were identified (84.5% at the FixMatch baseline), and the identification rate of KL4 bony fusion was 98.2%. The error in tibiofemoral joint space width was 0.43±0.11mm (physician measurement error 0.38±0.09mm, p=0.12), and the relative error in osteophyte coverage area was 7.3% (12.5% ​​in the traditional method). Even under the interference of metal artifacts and synovial calcification noise after joint replacement surgery, the ROC-AUC remained at 0.915±0.023. Attached Figure Description

[0055] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments;

[0056] Fig. 1 This is a schematic diagram illustrating the principle and flow of the auxiliary diagnostic method for knee osteoarthritis based on semi-supervised learning and dual knowledge distillation of the present invention.

[0057] Fig. 2 This is a schematic diagram illustrating the principle framework of the auxiliary diagnostic method for knee osteoarthritis based on semi-supervised learning and dual knowledge distillation of the present invention.

[0058] Fig. 3 This is a schematic diagram illustrating the specific process of the auxiliary diagnostic method for knee osteoarthritis based on semi-supervised learning and dual knowledge distillation according to the present invention. Detailed Implementation

[0059] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings.

[0060] This invention aims to address two core challenges in the diagnosis of knee osteoarthritis (KOA): the scarcity of medical image data and the protection of patient privacy.

[0061] like Figs. 1 to 3The application discloses a knee osteoarthritis auxiliary diagnosis method based on semi-supervised learning and double knowledge distillation, and the method comprises the following steps:

[0062] S1, personalized pixel shuffling based on knee osteoarthritis (KOA) pathological characteristics: according to the KL grading standard and the OARSI guideline, the key area and the privacy sensitive area for knee osteoarthritis diagnosis are dynamically identified, the privacy policy is adjusted based on the KL grading through the visual feature entropy (VFE), and the diagnostic key features are reserved;

[0063] Further, the knee osteoarthritis diagnosis key area in S1 includes the femoral condyle osteophyte area with KL≥2 and osteophyte area≥10%, and the tibiofemoral joint space area with JSW≤3mm.

[0064] Further, the dynamic adjustment of the privacy policy based on the KL grading in S1 is specifically as follows: high-intensity pixel disturbance is applied to the knee osteoarthritis diagnosis key area while the edge gradient is reserved; low-intensity disturbance is adopted for the non-KL grading key area to avoid interference with the inflammation features.

[0065] Specifically, the pixel disturbance intensity λ is dynamically adjusted based on the KL grading: when the femoral condyle edge osteophyte coverage area is detected to be≥15% (i.e. KL 3 level threshold), the pixel disturbance intensity λ is increased to 0.7; when the tibiofemoral joint space width is≤2.5mm (KL 4 level typical indication), the median filter (window size 5x5) is superimposed to eliminate the patient identification features in the bony fusion area; low-intensity disturbance (λ=0.2) is adopted for the suprapatellar bursa synovial membrane hyperplasia area (ultrasound confirmed synovial membrane thickness≥4mm) to reserve the texture features of the pannus generated by the synovial membrane, and to avoid interference with the inflammation features. In addition, by introducing the personalized parameters, researchers or developers can customize the protection level of each area for different application scenarios.

[0066] S2, attention guided distillation based on knee osteoarthritis (KOA) pathological prior: embedding a pre-trained bone geometric morphological convolution kernel (OAI data set osteophyte parameter initialization), strengthening the feature response of the osteophyte edge (KL≥2 level indication), and applying 1.5 times attention weight to the KL 3-4 level bony contact area;

[0067] Further, S2 specifically comprises the following steps:

[0068] S20, generating a spatial attention map: for the single-channel X-ray image of KOA, the spatial attention map can be generated by processing the spatial dimensions (height and width) of the image. Generally, this involves calculating the absolute value of each spatial dimension feature map, and then performing average pooling (or maximum pooling) at each pixel position to create a spatial attention map. For the input image I o ∈R H×Wwhere H and W represent the height and width of the image respectively, the spatial attention map A space is calculated as follows:

[0069] A space = Pool(|I o |) (1);

[0070] where Pool is the pooling method, and the average pooling is used.

[0071] S21, Normalization: The teacher and student spatial attention maps are normalized using a temperature parameterized softmax function, and the specific expression of normalization is as follows:

[0072]

[0073] where T is a temperature hyperparameter for controlling the smoothness of the softmax function; role represents the teacher or student, which means that role∈{teacher, student}; A teacher is the spatial attention map of the teacher model, and A student is the spatial attention map of the student model.

[0074] The temperature parameter T is dynamically adjusted according to the degree of patellofemoral joint alignment abnormality. When the patellar tilt angle (PTA) is ≥10° (calculated by Hough transform to detect the angle between the lower pole connecting line of the patella and the axis of the femoral trochlear groove), T is reduced from 1.0 to 0.5 to strengthen the feature focusing; when the tibial plateau sagittal angle (PSA) is ≥7°, a 1.5 times weight coefficient is applied to the subchondral bone sclerosis area in the normalization process.

[0075] S22, Calculate the total loss of attention guided distillation: merge and average the normalized attention maps of the teacher and student models to generate the final attention mask M; since the image is single-channel, the operation is relatively simple, and the calculation expression of averaging the two normalized spatial attention maps is as follows:

[0076]

[0077] where, represents the normalized spatial attention map of the teacher model; represents the normalized spatial attention map of the student model;

[0078] S23, based on the spatial attention mask M, the feature distillation loss is obtained by weighting the pixel loss of different positions according to the importance of the pixels; the expression of the feature distillation loss L mask is as follows:

[0079] L mask =∑ i,jM(i,j)·L student (i,j) (4);

[0080] wherein, L student (i,j) is the loss of the student model at position (i,j), and M(i,j) is the attention mask value at the position.

[0081] It is worth mentioning that the elements of the mask M represent the relative importance of each pixel, ranging from 0 to 1. When calculating the feature distillation loss, the spatial attention mask M is introduced, and the pixel loss at different positions is weighted according to the importance of the pixels. This means that the model will pay more attention to the areas with higher weights in the attention mask (i.e. important areas), and suppress the influence of other areas.

[0082] S24, calculate the spatial attention loss for encouraging the student model to simulate the teacher model, and the specific expression is as follows:

[0083] L atten = λ·‖A teacher -A student ‖2 (5);

[0084] wherein, λ is a hyperparameter for balancing the weight between the distillation loss and the attention loss.

[0085] S25, integrate the feature distillation loss and the spatial attention loss for encouraging the student model to simulate the teacher model to obtain the attention-guided distillation loss L attendis , and the specific expression is as follows:

[0086] L attendis = L mask + L atten (6)。

[0087] wherein, L student (i,j) is the loss of the student model at position (i,j), and M(i,j) is the attention mask value at the position.

[0088] Specifically, in this embodiment, the attention-guided distillation optimizes the student model by introducing a spatial attention mechanism. To address the problem that the quality of pseudo labels in semi-supervised learning depends on the performance of the teacher model, this method uses the spatial attention maps of the teacher and student models (generated by feature map pooling and normalization) to construct an attention mask M, which gives higher weights to key areas (such as joint edges) in loss calculation, and designs an attention loss L atten to force the student model to imitate the attention distribution of the teacher model. Through weighted focus and feature alignment, this method enables the student model to more accurately learn the core pathological features in medical images, effectively reduces the interference of low-quality pseudo labels, and improves the classification performance and model robustness.

[0089] In traditional semi-supervised learning, the student model is trained by the pseudo labels generated by the teacher model, but the quality of these pseudo labels is often affected by the performance of the teacher model. Especially in the case of incorrect or inaccurate teacher model, the quality of the pseudo labels will decrease significantly, which will affect the learning effect of the student model. By introducing attention mechanism, attention guided knowledge distillation can help the student model focus on the learning of the teacher model in key areas or important features, rather than directly relying on the global output of the teacher model. For the KL grading requirement of KOA, the pre-trained Hessian Affine Convolutional Kernel (OAI dataset bone osteophyte morphology parameter initialization) is embedded to strengthen the feature response of the osteophyte edge (KL≥2 core indicator); 1.5 times attention weight is applied to the bone contact area of KL3-4 grade to guide the model to focus on the key area of cartilage degeneration; when JSW≤2.5mm (KL4 grade indicator), a gap narrow compensation factor γ=0.8 is introduced to optimize the accuracy of joint space measurement.

[0090] S3, consistency pseudo label optimization consistent with ACR guidelines: after the teacher and the bridge network generate the pseudo label, the pseudo label that meets the consistency condition is screened and retained, which is used for student model training to output the classification auxiliary results of knee osteoarthritis (KOA);

[0091] Further, the consistency condition is set according to the ACR guidelines, specifically: the KL grading difference of the teacher and the bridge network is ≤1 grade, the JSW deviation is ≤0.5mm, and the osteophyte area error is ≤15%.

[0092] Further, the consistency module in S3 includes the following steps:

[0093] S30, calculate the loss function of the bridge network; first input the unlabeled dataset D W ={I W ,C W} into the teacher network to get the output C T . Then, filter the pseudo label using the threshold value. At the same time, input the same input dataset into the bridge network for training to generate the output C B . The pseudo label generated by the teacher network can also guide the learning of the bridge network and optimize its performance. The loss function L bridge of the bridge network is as follows:

[0094]

[0095] Where, represents the cross-entropy loss of the classification summary information C B of the bridge network and the pseudo label .

[0096] S31, Consistency pseudo-label generation: According to the classification information summary C T and C B , generate the final consistency pseudo-label Specifically, if the classification results of two images are the same (i.e., their class labels are the same), they are considered consistent and are used as candidate pseudo-labels The core of this process is to extract more reliable pseudo-labels through consistency comparison. The specific method is:

[0097]

[0098] The consistency determination needs to meet the following conditions simultaneously: the KL grading difference is less than or equal to one grade, the tibiofemoral joint space width (JSW) measurement deviation is less than or equal to 0.5 mm, and the relative error of osteophyte coverage area is less than or equal to 15%.

[0099] S32, final pseudo-label generation and loss calculation: Next, and merge to form the final pseudo-label to guide the training of the student network. The final unsupervised loss is the cross-entropy loss between the student model output C S and , as shown in the following formula:

[0100]

[0101] Finally, the total loss function is as follows:

[0102] L total = L sup + L unsup + L bridge + L attendis (10);

[0103] wherein, L bridge represents the loss function of the bridge network, L unsup represents the cross-entropy loss between the student model output classification summary information C S and the final pseudo-label ; L attendis represents the attention-guided distillation loss; L sup represents the supervised loss of the student model pre-training, which is calculated by the cross-entropy loss.

[0104] Further, the final output includes: KL grading diagnosis result (0-4 grades); tibiofemoral joint space narrowing degree grading (normal / mild / moderate / severe); osteophyte area ratio (accurate to 0.1%); treatment suggestion generation (conservative treatment / arthroscopic cleaning / artificial joint replacement).

[0105] That is, in this embodiment, the consistency module optimizes the reliability of pseudo labels by introducing a bridging network. To overcome the limitations of filtering pseudo labels with a high confidence threshold in traditional semi-supervised learning (e.g., low accuracy of the teacher model at the beginning of training leading to false high-confidence pseudo labels), the module simultaneously inputs unlabeled data into the teacher network and the bridging network, generating pseudo labels and C B , and selecting candidate pseudo labels with consistent classification results through consistency comparison and generating the final supervision signal combining the pseudo labels of the teacher network This double verification mechanism significantly reduces the interference of low-quality or false pseudo labels on the student model and optimizes the collaborative training of the bridging network and the student through loss functions L bridge and L unsup , improving the robustness and classification accuracy of the model in data-scarce scenarios.

[0106] Specifically, in semi-supervised teacher-student networks, pseudo labels are usually filtered by setting a high confidence threshold to eliminate those with low confidence. This strategy helps to reduce the negative impact of low-quality labels on student network training. However, at the initial stage of training, the teacher model may not have reached a high accuracy, which makes it risky to rely solely on high-confidence pseudo labels to train the student model. To address the KL grading pseudo label noise problem, a bridging network is introduced and combined with the ACR guidelines to set consistency conditions: the KL grading difference between the teacher and the bridging network ≤ 1 level, the JSW deviation ≤ 0.5 mm, and the osteophyte area error ≤ 15%, only retaining pseudo labels that meet the conditions (error rate from 20% to 8%), improving the reliability of KL grading.

[0107] The present application has the following beneficial effects compared with the prior art: 1. In the case of limited labeled data, the model performance is improved using a large amount of unlabeled data, and the dependence on expert labeling is reduced. The teacher-bridge-student model architecture is used, the teacher model generates pseudo-labels to guide the student model learning, and the bridge network is introduced to optimize the consistency of pseudo-labels, reducing the negative effects of low-quality pseudo-labels. 2. While protecting patient privacy, ensuring data availability, supporting cross-regional medical data sharing, and meeting privacy compliance requirements. A small amount of labeled data and a large amount of unlabeled data are used, the teacher model generates pseudo-labels to guide the student model learning, and the bridge network is introduced to optimize the consistency of pseudo-labels, reducing the negative effects of low-quality pseudo-labels. 3. Improve the model's ability to recognize key pathological features in medical images, and enhance classification accuracy and robustness. The spatial attention mechanism guides the student model to focus on key areas (such as cartilage degeneration features), avoiding reliance on global pseudo-labels, and improving classification accuracy. 4. Reduce the interference of low-quality or incorrect pseudo-labels, and improve the stability and performance of the model in data-scarce scenarios. The bridge network is introduced, and the pseudo-labels generated by the teacher network are compared for consistency, and more reliable pseudo-labels are selected. 5. On the OAI multi-center test set, the accuracy of KL 2-3 transition cases is improved from 84.5% (baseline) to 89.7%; JSW measurement error is 0.43±0.11mm (close to physician level 0.38±0.09mm); relative error of osteophyte area is 7.3% (traditional method 12.5%); only 10% of labeled data (500 cases) is required to achieve the performance of traditional fully supervised model (5000 cases), while supporting KOA image sharing and model training under cross-institutional privacy protection.

[0108] Obviously, the described embodiments are part of the embodiments of the present application, not all of the embodiments. The embodiments in the present application and the features in the embodiments can be combined with each other without conflict. The components of the embodiments of the present application described and shown in the drawings herein can be arranged and designed in various different configurations. Therefore, the detailed description of the embodiments of the present application is not intended to limit the scope of the claimed application, but only represents selected embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor are within the scope of protection of the present application.

Claims

1. A knee osteoarthritis auxiliary diagnosis method based on semi-supervised learning and double knowledge distillation, characterized in that: It comprises the steps of: S1, personalized pixel shuffling based on the pathological characteristics of knee osteoarthritis: according to the KL grading standard and OARSI guidelines, dynamically identify the key areas and privacy sensitive areas of knee osteoarthritis diagnosis, adjust the privacy policy based on the visual feature entropy, apply high intensity pixel disturbance to the key areas of knee osteoarthritis diagnosis and retain the edge gradient; for non-KL grading key area, low intensity disturbance is adopted to retain the diagnostic key features; S2, attention guided distillation based on the prior of knee osteoarthritis pathology: embedding pre-trained bone geometric morphological convolution kernel, strengthening the osteophyte edge feature response, applying 1.5 times attention weight to KL3-4 grade bone contact area; S3, consistency pseudo-label optimization in accordance with ACR guidelines: after the teacher and the bridging network generate pseudo-labels, the pseudo-labels that meet the consistency condition are selected and retained for student model training to obtain the classification auxiliary results of knee osteoarthritis.

2. The knee osteoarthritis auxiliary diagnosis method based on semi-supervised learning and double knowledge distillation according to claim 1, characterized in that: The key areas of knee osteoarthritis diagnosis in S1 include femoral condyle osteophyte area with KL≥2 and osteophyte area≥10%, and tibial-femoral joint space area with JSW≤3mm.

3. The knee osteoarthritis auxiliary diagnosis method based on semi-supervised learning and double knowledge distillation according to claim 1, characterized in that: In S1, the dynamic adjustment of privacy policy based on KL grading is as follows: when the femoral condyle edge osteophyte coverage area is detected to be≥15%, the pixel disturbance intensity λ is increased to 0.7; when the tibial-femoral joint space width is≤2.5mm, median filtering is superimposed to eliminate the patient identification features in the bone fusion area; for the suprapatellar bursa synovial membrane hyperplasia area, low intensity disturbance λ=0.2 is adopted to retain the texture features of pannus formation, and to avoid interference with the inflammation features.

4. The knee osteoarthritis auxiliary diagnosis method based on semi-supervised learning and double knowledge distillation according to claim 1, characterized in that: S2 specifically comprises the following steps: S20, generate spatial attention map: calculate the absolute value of each spatial dimension feature map of the single-channel X-ray image of KOA, and then perform average pooling at each pixel position to create a spatial attention map; S21, normalization: use a temperature parameterized softmax function to normalize the spatial attention maps of the teacher and student, and the specific expression of normalization is as follows: where T is a temperature hyper-parameter to control the smoothness of the softmax function; role represents the teacher or student, i.e., role e {teacher, student}; A teacher is the spatial attention map of the teacher model, A student is the spatial attention map of the student model; S22, calculate the total loss of attention guided distillation: combine and average the normalized attention maps of the teacher model and the student model to generate the final attention mask M; the calculation expression of the average of the two normalized spatial attention maps is as follows: wherein, denotes the normalized spatial attention map of the teacher model; denotes the normalized spatial attention map of the student model; S23, based on the spatial attention mask M, the pixel loss of different positions is weighted according to the importance of the pixels to obtain a feature distillation loss; the feature distillation loss L mask The expression is as follows: L mask =∑ i,j M(i,j)·L student (i,j) (4) where L student (i,j) is the loss of the student model at position (i,j), M(i,j) is the attention mask value at that position. S24, calculate the spatial attention loss used to encourage the student model to simulate the teacher model, and the specific expression is as follows: L atten = λ · ‖A teacher -A student ‖2 (5) Where λ is a hyperparameter used to balance the weight between the distillation loss and the attention loss; S25, integrating the feature distillation loss and the spatial attention loss for encouraging the student model to simulate the teacher model to obtain a distillation loss L under attention guidance attendis The specific expression is as follows: L attendis = L mask + L atten (6).

5. The knee osteoarthritis auxiliary diagnosis method based on semi-supervised learning and double knowledge distillation according to claim 4, characterized in that: spatial attention map A space The calculation method is as follows: A space = Pool(|I o |) (1) where Pool denotes the pooling method is average pooling; I o ∈R H×W denotes the input image, H and W represent the height and width of the image, respectively.

6. The knee osteoarthritis auxiliary diagnosis method based on semi-supervised learning and double knowledge distillation according to claim 1, characterized in that: The consistency condition in S3 is set according to the ACR guidelines, specifically: the KL grading difference between the teacher and the bridging network is≤1 grade, the JSW deviation is≤0.5mm, and the osteophyte area error is≤15%.

7. The knee osteoarthritis auxiliary diagnosis method based on semi-supervised learning and double knowledge distillation according to claim 1, characterized in that: The consistency module in S3 comprises the following steps: S30, calculate the loss function of the bridging network; obtain the classification summary information of the bridging network on the unlabeled data set, and combine the pseudo-labels filtered based on the same unlabeled data set classification summary information of the teacher network to perform cross-entropy loss calculation to obtain the loss function L of the bridging network bridge As follows: wherein L CE represents the cross-entropy loss; S31, consistency pseudo-label generation: C is summarized according to the classification information of the teacher network and the bridge network T and C B Perform consistency determination to generate final consistency pseudo-label The condition for consistency determination is: KL grading difference ≤ 1 level, tibiofemoral joint space width measurement deviation ≤ 0.5mm, relative error of osteophyte coverage area ≤ 15%; the calculation expression of the consistency pseudo-label is: S32, final pseudo label generation: combine the pseudo label of the teacher network and the consistency pseudo label to form a final pseudo label for guiding the training of the student network merge to form a final pseudo label for guiding the training of the student network S33、unsupervised loss: calculate student model output C S With The cross-entropy loss between and as unsupervised loss L unsup The calculation expression is as follows: S34, calculate the total loss, and the total loss function is as follows: L total = L sup + L unsup + L bridge + L attendis (10); wherein, L bridge represents the loss function of the bridging network, L unsup represents the cross-entropy loss between the student model output classification summary information C S and the final pseudo label ; L attendis represents the distillation loss under the attention guidance; L sup represents the supervised loss of the student model pre-training, which is calculated by the cross-entropy loss.

8. The knee osteoarthritis auxiliary diagnosis method based on semi-supervised learning and double knowledge distillation according to claim 7, characterized in that: S30 specifically comprises the following steps: S30-1, inputting the unlabeled data set D W = {I W , C W} into the teacher network to obtain the classification summary information C of the teacher network T S30-2, on the classified summary information C T Threshold filtering to obtain pseudo labels S30-3, input the same input data set into the bridge network for training, to obtain the classification summary information C of the bridge network B ; S30-4, pseudo labels generated based on the teacher network computing the loss function L of the bridge network bridge , specifically as follows:

9. The knee osteoarthritis auxiliary diagnosis method based on semi-supervised learning and double knowledge distillation according to claim 1, characterized in that: The final output classification auxiliary results include: KL grading diagnosis results, tibial-femoral joint space stenosis grading; osteophyte area ratio; treatment suggestion generation.

Images