Lumbar intervertebral disc herniation interpretable diagnosis system based on cross-style consistency semi-supervised segmentation

Abstract

The invention discloses a lumbar disc herniation interpretable diagnosis system based on cross-style consistency semi-supervised segmentation. The method comprises the following steps: firstly, carrying out style diversification processing on a lumbar axial MRI image through Bessel transformation to generate an image pair for consistency learning; the aggressive student model processes the stylized image and carries out cooperative training through cross pseudo supervision, and the robust student model provides a reliable prediction target through smooth parameter updating; the trained model can realize automatic segmentation of an intervertebral disc and a posterior spinous process area. Based on the segmentation result, the system adopts a geometric rule algorithm to quantify the prominence degree, and automatically generates a grading result and a visual report according to a clinical MSU grading standard. According to the method, the dependence on labeled data is reduced, the interpretability of the model is improved, and efficient and accurate lumbar disc herniation auxiliary diagnosis with clinical guiding significance can be realized.

Description

Technical Field

[0001] This invention relates to the field of medical image processing, specifically to an interpretable diagnostic system for lumbar disc herniation based on cross-style consistency semi-supervised segmentation. Background Technology

[0002] Lumbar disc herniation is a common degenerative spinal disease. Epidemiological surveys show that due to changes in work patterns, lumbar disc herniation is increasingly affecting younger people. The proportion of such patients is rising year by year, seriously impacting the normal work and daily lives of contemporary young people. Diagnosis of lumbar disc herniation mainly relies on the visual observation and subjective judgment of radiologists and orthopedic surgeons using lumbar MRI images. In clinical practice, the classification of lumbar disc herniation is crucial yet extremely challenging for treatment—different types of disc herniation often require different clinical treatment plans. Specifically, relatively mild types such as bulging can usually be relieved or cured with conservative treatment; while severe lesions such as protrusion and extrusion require surgical intervention. Due to the high structural similarity between different types, inexperienced clinicians find it difficult to accurately classify them. Therefore, how to achieve accurate and effective classification of the type and early stage of disc herniation has become a critical clinical problem that urgently needs to be solved.

[0003] Currently, the technical solutions applied to the automatic diagnosis of lumbar disc herniation mainly follow the following two paradigms: The first paradigm is a semantic segmentation model based on fully supervised learning. This type of method requires a large number of pixel-level labels for training on the intervertebral disc and posterior spinous process regions, which are precisely annotated by experts. Although this type of method can achieve high accuracy on specific datasets, its limitations are very significant: (1) It is highly dependent on a large amount of high-quality labeled data. The annotation of medical images requires domain experts to invest a lot of time and effort, which is extremely costly, and the annotation quality is easily affected by subjective factors, which greatly limits the promotion and application of the model. (2) The model has poor generalization ability. Images acquired from different medical institutions, different MRI equipment (such as GE, Siemens, Philips, etc.), and different scanning parameters have significant style differences. When a fully supervised model trained on one dataset is directly applied to another dataset, its performance often drops sharply, making it difficult to meet the needs of actual clinical applications.

[0004] The second paradigm directly uses an end-to-end classification network to classify or grade whether a lumbar disc herniation is present. While this method reduces the reliance on pixel-level annotations, it has more fatal flaws: (1) Poor interpretability. Deep learning models are often considered black-box models, and the basis for their decisions is difficult to trace and verify. Clinical diagnosis requires extremely high reliability and transparency, and doctors are unlikely to trust a system that cannot provide diagnostic evidence. (2) Inability to provide localization information. This type of method can only provide a classification conclusion and cannot provide the specific location, extent, and direction of the disc herniation, which are crucial for clinical diagnosis, as segmentation models do.

[0005] Therefore, there is an urgent need in this field for a technical solution that can simultaneously address the following issues: achieve high-performance segmentation with limited labeled data; and provide clinically compliant, visualized diagnostic results, thereby enhancing the interpretability of the results and doctors' willingness to use them. Summary of the Invention

[0006] The purpose of this invention is to provide an interpretable diagnostic system for lumbar disc herniation based on cross-style consistency semi-supervised segmentation, in order to solve the problems of strong dependence on labeled data and insufficient interpretability of diagnostic results in existing technologies.

[0007] This invention is implemented as follows: An interpretable diagnostic system for lumbar disc herniation based on cross-style consistency semi-supervised segmentation is proposed. Its core lies in constructing a complete framework integrating stylized data augmentation, multi-model collaborative training, and interpretable rule-based post-processing algorithms. First, the system enhances the limited labeled data with stylistic diversification using the Bezier transform algorithm to improve the model's generalization ability. Then, a collaborative segmentation network consisting of two aggressive student models and one robust student model is used for training, employing cross-pseudo-supervision and dynamic exponential moving average mechanisms. This allows the model to fully utilize unlabeled data and achieve consistent and stable segmentation performance across stylistically diverse data. Finally, instead of simply outputting segmentation results, the system inputs the segmented disc and posterior spinous process masks into a geometric rule-based computation module. Based on the clinical MSU grading standard, quantitative analysis and automatic grading are performed, generating a visualized diagnostic report integrating the original image, segmentation contours, and grading conclusions.

[0008] The lumbar disc herniation interpretable diagnostic system based on cross-style consistency semi-supervised segmentation provided by this invention includes: The data collection module is used to acquire lumbar spine axial magnetic resonance imaging (MRI) datasets, collect corresponding pathological biopsy results and expert diagnostic reports, establish a precise correspondence between images and diagnostic reports, and divide the dataset into data with and without annotations. The data preprocessing module anonymizes the collected datasets, then divides them into training and test sets according to a certain ratio, and uses Bezier transformation to process each image in the training set to generate a pair of images with different styles but consistent semantic content, which are then fed into the segmentation model. The multi-student model collaborative segmentation module includes two aggressive student models and one robust student model. The aggressive student models take images with different styles after style transformation as input and perform collaborative optimization through a cross-pseudo-supervision strategy. The robust student model takes the original image as input, and its parameters do not participate in the model backpropagation. Instead, they are updated from the aggressive student models through a dynamic exponential moving average method. The geometric feature calculation module is used to calculate the protrusion degree quantification index of the intervertebral disc and posterior spinous process region mask output by the segmentation model using a geometric algorithm based on the MSU grading rules. The clinical grading and visualization module is used to automatically grade based on the calculation results of geometric features and according to the clinical MSU grading standard, and generate a visualized diagnostic report that integrates the original image, segmentation mask and grading information.

[0009] Furthermore, the interpretable diagnostic system for lumbar disc herniation based on cross-style consistency semi-supervised segmentation is characterized by the cross-pseudo-supervision strategy in the multi-student model collaborative segmentation module, which specifically comprises two steps. First, the prediction results generated by the first aggressive student model are used as pseudo-labels to supervise the training of the second aggressive student model. Simultaneously, the feature map output by the first aggressive student model is perturbed with noise and concatenated with the input image before being input into the second aggressive student model. Correspondingly, the prediction results of the second aggressive student model are also used as pseudo-labels to supervise the training of the first aggressive student model. The advantage lies in the addition of a feature interleaving learning mechanism to the classic cross-pseudo-supervision framework. By perturbing the feature map output by the first aggressive student model with noise and concatenating it with the input image before inputting it into the second aggressive student model, the second aggressive student model can learn the perturbed features from the first aggressive student model, driving both networks to learn a consistent feature representation space, thereby learning more robust and accurate features.

[0010] Furthermore, the interpretable diagnostic system for lumbar disc herniation based on cross-style consistency semi-supervised segmentation is characterized in that the parameter update of the robust student model adopts a dynamic exponential moving average, specifically as follows: The robust student model does not update its parameters through backpropagation; instead, it is obtained by exponentially averaging the parameters of the first aggressive student model. During training, the segmentation performance of the robust student model and the first aggressive student model on the training set is continuously compared. The model with the better performance is selected as the robust student model for the next training cycle, while the other model becomes the first aggressive student model. The advantage of this approach is that the robust student model, through exponential moving average, integrates the model parameters of the first aggressive student model throughout its training process, suppressing oscillations during model training. This results in generating more stable pseudo-labels to supervise the first and second aggressive student models, promoting smoother convergence. This dynamic model role switching eliminates the problem in the early stages of training where the robust student model's segmentation performance is lower than that of the first aggressive student model, thus failing to provide effective pseudo-labels for supervision.

[0011] Furthermore, the interpretable diagnostic system for lumbar disc herniation based on cross-style consistency semi-supervised segmentation is characterized in that the multi-student model collaborative segmentation module specifically comprises: The first and second radical student models use a cross-pseudo-supervision strategy on labeled and unlabeled training sets. The robust student model only generates pseudo-labels using the unlabeled training set and uses them to supervise the training of the first and second radical student models.

[0012] Furthermore, the lumbar disc herniation interpretable diagnostic system based on cross-style consistency semi-supervised segmentation is characterized in that the geometric feature calculation module obtains grading lines by finding and calculating the pixel coordinates of key detection points on the posterior spinous process segmentation mask, and calculates the relative position of the lumbar disc herniation point on the grading lines to obtain the grading result. The geometric feature calculation module includes: The reference point determination unit is used to determine the reference reference point based on the geometric characteristics of the posterior spinous process mask; The grading line determination unit is used to calculate the position of the grading line on the mask based on the reference point; The quantitative index calculation unit is used to calculate the relative geometric position of the intervertebral disc herniation area and the grading line to obtain the final grading result. The advantage of this geometric feature calculation module is that it uses a non-end-to-end calculation method to obtain the grading results of lumbar disc herniation. Its calculation process strictly follows the clinical MSU grading standard, which enhances the interpretability of the results.

[0013] Furthermore, the lumbar disc herniation interpretable diagnostic system based on cross-style consistency semi-supervised segmentation is characterized in that the grading strategy of the clinical grading and visualization module adopts the MSU grading standard, which divides the degree of herniation into four levels from 0 to 3, and divides the herniation area into three zones: A, B, and C.

[0014] Furthermore, the lumbar disc herniation interpretable diagnostic system based on cross-style consistency semi-supervised segmentation is characterized in that the Bezier transform algorithm used in the style transformation module can change the brightness, contrast, and texture style features of the image while maintaining the semantic integrity of the image content. Its advantages are... Furthermore, the lumbar disc herniation interpretable diagnostic system based on cross-style consistency semi-supervised segmentation is characterized in that the two radical student models and one robust student model use a structurally consistent U-Net segmentation network.

[0015] The present invention also provides the following technical solutions: An interpretable diagnostic system for lumbar disc herniation based on cross-style consistency semi-supervised segmentation includes the following steps: S1. Construct the above-mentioned interpretable diagnostic system for lumbar disc herniation based on cross-style consistency semi-supervised segmentation; S2. Obtain axial MRI image data of the lumbar spine, divide it into labeled data and unlabeled data, divide it into training set and validation set, and perform preliminary processing on the original images, including normalization and image size scaling, to improve data quality; S3. Apply the Bezier transform algorithm to the training set images to enhance style diversity, generating image pairs with consistent content but different styles, thus enriching the image diversity. S4. Using the multi-student model collaborative segmentation module, the image pairs with labeled masks after Bezier transformation are input into the first radical student model and the second radical student model respectively. A feature interleaving learning mechanism is added. The feature map output by the first radical student model is superimposed with noise and concatenated with the input image before being input into the second radical student model. The model parameters are optimized by the cross-entropy loss and Dice loss of the model output and the segmentation label. S5. Using the multi-student model collaborative segmentation module, the image pairs with unlabeled masks after Bezier transformation are input to the first aggressive student model and the second aggressive student model respectively. The original image is input to the robust student model. The feature map output by the first aggressive student model is superimposed with noise and concatenated with the input image before being input to the second aggressive student model. The first and second aggressive student models will use a cross pseudo-supervision strategy to update the model parameters. The pseudo-labels generated by the robust student model will supervise the first and second aggressive student models to update the model parameters. S6. Evaluate the segmentation performance of the trained model on the validation set, calculate key indicators such as Dice coefficient and 95% Hausdorff distance (HD95), and compare with existing methods to verify the superiority of the proposed segmentation method. S7. Based on the segmentation mask obtained from the segmentation model, a geometric algorithm based on hierarchical rules is used to calculate the quantitative index of intervertebral disc herniation grading, and the classification accuracy (ACC), recall, and other key indicators of the grading are calculated. The superiority of the method of the present invention is verified by comparing it with existing methods. S8. Automatically classify the quantitative results according to the clinical MSU grading standard, and generate a visual diagnostic report that integrates the original image, segmentation results and grading information.

[0016] Compared with the prior art, the advantages of the present invention are reflected in: 1. By introducing a multi-student model collaborative segmentation module, the system can more efficiently utilize a large number of unlabeled images for model training, significantly reducing the expensive labeling cost of medical images and greatly reducing the cost of model training; 2. Based on cross-pseudo-supervision, a dynamic exponential moving average mechanism and a feature perturbation mechanism are introduced to strengthen the feature perturbation and supervision constraints within the model. This allows the model to ignore shallow knowledge in the image, learn more abstract features, and improve segmentation accuracy. 3. Bezier transform is used for style transformation of images. Compared to traditional data augmentation methods, this method can enhance data without altering the semantics of the images. The generated image style pairs can increase image perturbation for the model. Combined with the strong mutual supervision characteristics in the proposed multi-student model collaborative segmentation module, this forces the model to learn content semantics rather than superficial style features, thereby improving the model's segmentation performance and robustness. 4. A geometric feature calculation module was designed, which uses a non-end-to-end calculation method to obtain the grading results of lumbar disc herniation. The calculation process strictly follows the clinical MSU grading standard to enhance the interpretability of the results. Attached Figure Description

[0017] Figure 1 This is the overall flowchart of the system of the present invention. Figure 2 Schematic diagram of Bessel transformation Figure 3 A schematic diagram of the structure of the multi-student model collaborative segmentation module. Figure 4 A diagram illustrating the dynamic exponential moving average strategy. Figure 5 A schematic diagram of the geometric algorithm based on hierarchical rules. Detailed Implementation

[0018] The present invention will be further described in detail below with reference to specific examples and accompanying drawings.

[0019] Figure 1This is a flowchart of the overall segmentation model in this invention, which shows in detail the process from generating style image pairs from the original image through Bezier transformation, to the multi-student model collaborative segmentation module, and finally to outputting the segmentation result.

[0020] Example An interpretable diagnostic system for lumbar disc herniation based on cross-style consistency semi-supervised segmentation includes the following steps: Step 1: Data Collection. Axial MRI images of the lumbar spine were collected from 726 patients in the orthopedic department of a certain hospital. Each patient's data included 15 slices from 5 intervertebral discs from L1 to S1.

[0021] Step 2: Image Labeling and Segmentation. Using image labeling software, three experienced orthopedic surgeons meticulously delineated the lumbar intervertebral disc and posterior spinous process regions on axial lumbar MRI images, obtaining pixel-level segmentation labels. Simultaneously, a majority voting method was used to label the degree of protrusion of each lumbar intervertebral disc. The data was divided into training and validation sets. All images in the validation set had hierarchical labels; all images in the validation set had pixel-level segmentation labels; and 20% of the images in the training set had pixel-level segmentation labels.

[0022] Step 3: Data Preprocessing. The data preprocessing module performs preprocessing operations on the images. Normalization scales the pixel values ​​to the [0,1] range, and the crop size is set to 224×224. A smooth, monotonic Bézier transform curve is used as the transform function to change the grayscale distribution of each image in the training set. The cubic Bézier transform curve is created by two endpoints (P0 and P3) and two independent intermediate points (P1 and P2), as shown below. Figure 2 As shown, the algorithm uses two sets of curves to convert the image into two different styles. In the first set of curves, control points P0 = (-1, -1) and P3 = (1, 1) are defined to obtain an increasing function. When the second coordinate (i.e., the ordinate) is inverted, another decreasing function is obtained. When P0 = P1 and P2 = P3, as shown in Figures (b) and (e), we can obtain a linearly increasing or decreasing function. Furthermore, when choosing P1 = (-a, a) and P2 = (a, -a) or P1 = (-a, -a) and P2 = (a, a), nonlinear functions as shown in Figures (c) and (d) or (f) and (g) are obtained. After establishing two sets of Bézier curves (one increasing and the other decreasing), the original image can be mapped to different grayscale distributions according to the curve functions, thereby obtaining two styles of images, as shown in the second row of the figure.

[0023] Step 4: Train the multi-student collaborative segmentation module using the training set. For example... Figure 3As shown, three U-Net networks with identical structures are used. After parameter initialization, they are labeled as the first radical student network, the second radical student network, and the robust student network, respectively, and the backpropagation of the robust student network is turned off.

[0024] For images with segmentation labels, after Bezier transformation, one image is randomly selected from Style 1 and fed into the first radical student network. The feature map and segmentation result output by the model are calculated. Another image is randomly selected from Style 2 and concatenated with the feature map output by the first radical student network with superimposed noise. The concatenated image is then fed into the second radical student network. The first and second radical student networks use the segmentation labels to calculate the loss and update the model.

[0025] For images without segmentation labels, after Bezier transformation, one image is randomly selected from Style 1 and fed into the first radical student network. The feature map and pseudo-label of the model output are obtained through calculation. Another image is randomly selected from Style 2 and concatenated with the feature map output by the first radical student network with superimposed noise. This concatenation is then fed into the second radical student network. The original image is input into the robust student model to obtain the feature map. The first and second radical student networks update the model using cross-pseudo-supervision and the pseudo-label of the robust student network. The robust student network updates the model from the parameters of the first radical student network using exponential moving average.

[0026] Step 5: Dynamic Collaboration EMA. For example... Figure 4 After every 5 rounds of model training iterations, the first aggressive student network and the robust student network will be compared in terms of the model's segmentation performance on the training set. The model with better performance will be selected as the robust student model in the next training cycle, while the other model will be selected as the first aggressive student model.

[0027] Step 6: Segmentation performance evaluation. Calculate the Dice coefficient and 95% Hausdorff distance (HD95) segmentation metric for all student models on the validation set. Select the optimal student model as the final segmentation model and obtain the corresponding segmentation results.

[0028] Step 7: Calculate the quantitative index of intervertebral disc herniation grading using a geometric algorithm based on grading rules.

[0029] The specific process is as follows: Figure 5 As shown, firstly, locate the posterior spinous process region (the region with a pixel value of 255 in the schematic diagram) in the segmentation mask and mark it as a rectangular region. Divide this region into left and right parts using the midpoint of its x-axis as the boundary. Find the 255-pixel point corresponding to the maximum y-axis value in each part, and mark it as points A and B. The line segment AB connecting the two points is the grading line for levels 1 and 2, as shown... Figure 5 As shown in (a) to 5(d).

[0030] Next, locate the lowest point P of the intervertebral foramen between the x-axis coordinates of points A and B. Draw a perpendicular line MP from point P to AB and calculate the length of MP. On MP, take point Q such that the length of MQ is 1 / 3 of MP. Draw a straight line CD parallel to AB through point Q. Points C and D are the intersection of the boundary lines of the posterior spinous process region. Line segment CD is the grading line for levels 2 and 3. Shift points A and B upwards by 1 / 3 of the length of MP to obtain points E and F. Line segment DE is the grading line for levels 0 and 1. Figure 5 As shown in (e) to 5(h).

[0031] Finally, calculate the partition lines. For example... Figure 5 As shown in (i) to 5(k), starting from point C on line segment CD, mark a point every 1 / 4 of the CD length, excluding the midpoint, as C, K1, K2, and D respectively. Draw perpendicular lines from these four points to CD, marking them as l1, l2, l3, and l4 as zoning lines. Since lumbar disc herniation usually protrudes into the intervertebral foramen, locating the lowest point of the lumbar disc region can determine the herniation point. For grading, if the y-axis coordinate of the herniation point is lower than EF, it can be classified as grade 1; lower than AB, it can be classified as grade 2; lower than CD, it can be classified as grade 3; if the most prominent point is not lower than EF, it is classified as grade 0. For zoning, if the herniation point is between lines l2 and l3, it is classified as zone A; if it is between lines l1 and l2 or lines l3 and l4, it is classified as zone B; the rest are zone C; if the most prominent point is grade 0, no zoning is calculated. In summary, the grading and zoning results can be obtained based on the relative position of the lowest point of the lumbar disc herniation on the grading and zoning lines.

[0032] Step 8: Evaluation of Classification Results. Evaluate the classification performance of the algorithm on the validation set, and calculate metrics such as average classification accuracy (ACC), precision (PRE), and recall.

Claims

1. An interpretable diagnostic system for lumbar disc herniation based on cross-style consistency semi-supervised segmentation, characterized in that, include: The data collection module is used to acquire lumbar spine axial magnetic resonance imaging (MRI) datasets, collect corresponding pathological biopsy results and expert diagnostic reports, establish a precise correspondence between images and diagnostic reports, and divide the dataset into data with and without annotations. The data preprocessing module anonymizes the collected datasets, then divides them into training and test sets according to a certain ratio, and uses Bezier transformation to process each image in the training set to generate a pair of images with different styles but consistent semantic content, which are then fed into the segmentation model. The multi-student model collaborative segmentation module includes two aggressive student models and one robust student model. The aggressive student models take images with different styles after style transformation as input and perform collaborative optimization through a cross-pseudo-supervision strategy. The robust student model takes the original image as input, and its parameters do not participate in the model backpropagation. Instead, they are updated from the aggressive student models through a dynamic exponential moving average method. The geometric feature calculation module is used to calculate the degree of protrusion quantification index of the intervertebral disc and posterior spinous process region mask output by the segmentation model using a geometric algorithm based on the Michigan State University (MSU) lumbar disc herniation grading rules (hereinafter referred to as "MSU grading standards"). The clinical grading and visualization module is used to automatically grade based on the calculation results of geometric features and according to the clinical MSU grading standard, and generate a visualized diagnostic report that integrates the original image, segmentation mask and grading information.

2. The interpretable diagnostic system for lumbar disc herniation based on cross-style consistency semi-supervised segmentation according to claim 1, characterized in that, The cross-pseudo-supervision strategy in the multi-student model collaborative segmentation module is specifically as follows: First, the prediction results generated by the first aggressive student model are used as pseudo-labels to supervise the training of the second aggressive student model. Simultaneously, a feature interleaving learning mechanism is incorporated: the feature map output by the first aggressive student model is noise-perturbed, concatenated with the input image, and then fed into the second aggressive student model. Correspondingly, the prediction results of the second aggressive student model are also used as pseudo-labels to supervise the training of the first aggressive student model, forming a symmetrical collaborative learning framework.

3. The interpretable diagnostic system for lumbar disc herniation based on cross-style consistency semi-supervised segmentation according to claim 1, characterized in that, The parameter updates of the robust student model employ a dynamic exponential moving average, specifically: The robust student model does not update its own parameters through backpropagation, but instead obtains them by exponentially averaging the parameters of the first aggressive student model. During training, the segmentation performance of the robust student model and the first aggressive student model on the training set is continuously compared. The model with better performance is selected as the robust student model in the next training cycle, while the other model is selected as the first aggressive student model.

4. The interpretable diagnostic system for lumbar disc herniation based on cross-style consistency semi-supervised segmentation according to claim 1, characterized in that, The multi-student model collaborative segmentation module specifically comprises: The first and second radical student models use a cross-pseudo-supervision strategy on labeled and unlabeled training sets. The robust student model only generates pseudo-labels using the unlabeled training set and uses them to supervise the training of the first and second radical student models.

5. The interpretable diagnostic system for lumbar disc herniation based on cross-style consistency semi-supervised segmentation according to claim 1, characterized in that, The geometric feature calculation module obtains grading lines by finding and calculating the pixel coordinates of key detection points on the posterior spinous process segmentation mask, and calculates the relative position of the lumbar intervertebral disc herniation point on the grading lines to obtain the grading result. The geometric feature calculation module includes: The reference point determination unit is used to determine the reference reference point based on the geometric characteristics of the posterior spinous process mask; The grading line determination unit is used to calculate the position of the grading line on the mask based on the reference point; The quantitative index calculation unit is used to calculate the relative geometric position of the intervertebral disc herniation area and the grading line to obtain the final grading result.

6. The interpretable diagnostic system for lumbar disc herniation based on cross-style consistency semi-supervised segmentation according to claim 1, characterized in that, The grading strategy of the clinical grading and visualization module adopts the MSU grading standard, which divides the degree of prominence into four levels from 0 to 3, and divides the prominence area into three zones: A, B, and C.

7. The interpretable diagnostic system for lumbar disc herniation based on cross-style consistency semi-supervised segmentation according to claim 1, characterized in that, The style transformation module uses the Bezier transform algorithm, which can change the brightness, contrast, and texture style features of the image while keeping the semantic content of the image unchanged.

8. The interpretable diagnostic system for lumbar disc herniation based on cross-style consistency semi-supervised segmentation according to claim 1, characterized in that, The two radical student models and one robust student model use a structurally consistent U-Net segmentation network.

9. An interpretable diagnostic system for lumbar disc herniation based on cross-style consistency semi-supervised segmentation, characterized by: Includes the following steps: S1. Construct an interpretable diagnostic system for lumbar disc herniation based on cross-style consistency semi-supervised segmentation as described in any one of claims 1 to 8; S2. Obtain axial MRI image data of the lumbar spine, divide it into labeled data and unlabeled data, and divide it into training set and validation set; S3. Apply the Bessel transform algorithm to the training set images to enhance style diversity and generate image pairs with different styles; S4. Using a multi-student model collaborative segmentation module, the model parameters are optimized through cross-pseudo-supervision and dynamic exponential moving average mechanisms to achieve automatic segmentation of the lumbar intervertebral disc region and the posterior spinous process region; S5. Based on the segmentation mask obtained from the segmentation model, a geometric algorithm based on the MSU grading rules is used to calculate the quantitative index of intervertebral disc herniation grading; S6. The quantitative results are automatically graded according to the clinical MSU grading standards, and a visual diagnostic report integrating the original image, segmentation results and grading information is generated.

Images