Intervertebral joint osteoarthritis image feature evaluation method based on deep learning

Abstract

The invention relates to the technical field of medical image auxiliary diagnosis, and provides an intervertebral joint osteoarthritis image feature evaluation method based on deep learning, which can synchronously identify five types of FJOA image features such as joint space stenosis, osteophyte, hypertrophy, subchondral bone erosion and subchondral cyst, and improves the comprehensiveness and efficiency of evaluation. The nnU-Net model is adopted to perform high-precision segmentation on an intervertebral joint region, so that the positioning precision is effectively improved; by introducing a shared feature extraction network based on ResNet-18 and five parallel classification sub-networks, joint modeling of multi-scale semantic information is realized, and the recognition accuracy and generalization ability of the model are improved; according to the method, a Grad-CAM technology is combined to generate an activation thermodynamic diagram, a model interpretation basis is superposed to an original image, visual display of an evaluation result is realized, interpretability and clinical applicability of the model are enhanced, and FJOA image features in axial lumbar vertebra CT images from two central data sets can be quantitatively evaluated so as to comprehensively quantify individual FJOA image features.

Description

Technical Field

[0001] This invention relates to the field of medical image-assisted diagnosis technology, and in particular to a method for evaluating imaging features of intervertebral joint osteoarthritis based on deep learning. Background Technology

[0002] Intervertebral joints (FJs), the only synovial joints in the human spine, are an important component of the vertebral "three-joint complex." These joints not only bear axial loads but also assist in spinal movement and limit excessive rotation and displacement of the spine. Notably, FJs are one of the leading causes of mechanical spinal pain, and interventional procedures related to FJ-related low back pain are the second most common in the United States. Similar to other synovial joints, degenerative changes in FJs primarily manifest as osteoarthritis. X-rays, CT, MRI, and SPECT scans are used to assess the severity of intervertebral joint osteoarthritis (FJOA), with CT scans being superior in displaying degenerative features of FJOA such as osteophytes and hypertrophy, thus improving consensus among physicians. Imaging features of FJOA include joint space narrowing (JSN), osteophytes, hypertrophy, subchondral bone erosion, and subchondral cysts, as described by Weishaopt et al. Previous studies have used an overall assessment of FJOA (grade 0-3) based on the severity of all these features, with higher grades indicating more severe FJOA degeneration. However, existing research has found that the relationship between the condensed FJOA grading scale and pain or prognosis remains unclear. This may be because each severity level in the FJOA grading scale includes multiple FJOA imaging features and their severity; for example, a grade 2 FJOA may only have moderate osteophytes, or it may simultaneously include JSN, moderate osteophytes, moderate hypertrophy, and mild subchondral bone erosion. This limitation severely impacts the clinical applicability of the FJOA grading system. Studies of individual imaging features have shown that osteophytes and hypertrophy may particularly lead to narrowing of the spinal canal, lateral recesses, and intervertebral foramina. This bony narrowing may compress surrounding structures and nerves, thus causing intractable pain. Paradoxically, osteophytes and hypertrophy are, to some extent, protective measures against spinal instability, potentially promoting spinal stability and even serving as predictive targets for early assessment of a positive response to FJOA interventional therapy. Therefore, accurately identifying and assessing each imaging feature of FJOA may be more valuable than simply relying on the condensed grading scale. However, most current studies lack reliable assessment methods. Intra-reader and inter-reader consistency in manually assessing the severity of various features is typically poor to moderate. This uncontrollable assessment variability not only affects the diagnosis and treatment of FJ-related low back pain, but also results in FJOA prevalence ranging from 3% to 76% across different studies. Until we can further explore the association between various imaging features of FJ degeneration and the clinical manifestations and treatment prognosis of FJ-related low back pain, reliable tools are needed to improve the accuracy and consistency of individual imaging feature assessments for FJOA.

[0003] Recently, deep learning (DL) based algorithms have demonstrated outstanding performance in medical image analysis, offering high accuracy, consistency, and impartiality. To our knowledge, existing literature has not yet explored DL methods capable of comprehensively quantifying individual FJOA image features.

[0004] The present invention aims to provide a two-stage multi-task DL model for quantitatively assessing the severity of FJOA imaging features in axial lumbar CT images from two central datasets, in order to comprehensively and reliably quantify individual FJOA imaging features. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of the prior art by providing a deep learning-based method for evaluating imaging features of intervertebral joint osteoarthritis, thereby resolving the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] This application provides a deep learning-based method for evaluating imaging features of intervertebral joint osteoarthritis, including the following steps:

[0008] S1. Collect lumbar CT images of patients with low back pain, and select image slices of complete intervertebral joint structures on both sides from all images for the L1 / L2 to L5 / S1 spinal segments to construct a dataset for the first stage of model training.

[0009] S2. Construct a one-stage model for automatic segmentation of the intervertebral joint region. Input the lumbar CT image to be processed into the one-stage model. The one-stage model uses a segmentation network based on nnU-Net to process the lumbar CT image, outputs a binary segmentation mask corresponding to the bilateral intervertebral joint regions, and crops out multiple intervertebral joint CT patches.

[0010] S3. Construct a two-stage model dataset for FJOA assessment using the intervertebral joint CT patch output from the one-stage model.

[0011] S4. Construct a two-stage multi-task deep learning model based on the ResNet-18 backbone network. Input the intervertebral joint CT patch output by the first-stage model into the second-stage model. In the second-stage model, extract multi-scale semantic features for each intervertebral joint CT patch as a shared representation and input them into five parallel fully connected classification networks. Each network outputs a different type of FJOA feature.

[0012] S5. Generate a gradient-weighted class activation map based on the classification results of each FJOA feature, and overlay it onto the original lumbar spine CT image to form a visual evaluation image.

[0013] Preferably, in step S1, annotation software is used to annotate the bilateral complete intervertebral joint structures in the image slices, and the annotated image slices are used to construct a dataset for the first-stage model training.

[0014] Preferably, in step S2, the dataset used for the first-stage model training is input into the nnU-Net segmentation network model for training. The hyperparameter tuning of the nnU-Net segmentation network model includes the following steps:

[0015] a1. Analyze the voxel spacing distribution of all lumbar CT images in the dataset of the first-stage model training, and determine the target voxel spacing based on the median or the tenth percentile.

[0016] a2. Resample all input images to the target voxel spacing to balance spatial resolution and training efficiency;

[0017] a3. Automatically configure image patch size, batch size, and network depth based on available GPU memory capacity;

[0018] a4. During the five-fold cross-validation process, keep all hyperparameters except the learning rate unchanged, divide the training data into five subsets, use one subset for validation in turn, and use the remaining subsets for training, in order to evaluate the performance of each hyperparameter combination.

[0019] a5. Under the premise of fixed hyperparameters, the learning rate decreasing strategy was adopted, and 1×10-3 was finally determined as the optimal value.

[0020] a6. Compare the performance of the three optimizers: stochastic gradient descent, Adam, and AdamW, and retain the optimizer that performs best on the task in this dataset.

[0021] Preferably, the binary segmentation mask output in S2 is processed through the following post-processing steps to generate regular intervertebral joint CT patches:

[0022] b1. Calculate the minimum row index i of all pixels in the mask that have a value of 1. min Maximum row index i max Minimum column index j min Maximum column index j max ;

[0023] b2. Construct a rectangular region between the minimum and maximum indices, set the pixel values ​​within this region to 1, and the rest to 0, to obtain a regular rectangular filled mask;

[0024] b3. Use the filling mask to crop the original CT image to obtain regular intervertebral joint CT patches, and adjust them uniformly to a preset size.

[0025] Preferably, in S3, two radiologists with experience in spinal imaging independently use annotation software to annotate the same intervertebral joint CT patch with image labels. When the two doctors' image labels are inconsistent, a third doctor makes the decision to generate a standard reference label. The annotated intervertebral joint CT patch is used to construct a two-stage model training dataset for FJOA evaluation.

[0026] Preferably, the evaluation criteria for the image labels include the following five categories of FJOA image features:

[0027] Joint space narrowing is defined as less than 2 mm and is classified as "none" or "present".

[0028] Bone spurs are graded as "none", "mild", "moderate" or "severe";

[0029] Hypertrophy is classified as "none", "mild", "moderate" or "severe";

[0030] Subchondral bone erosion is graded as "none", "mild" or "severe";

[0031] Subchondral cysts are classified as "absent" or "present".

[0032] Preferably, the two-stage multi-task deep learning model includes a shared feature extraction backbone network and five parallel fully connected sub-networks. The backbone network is a ResNet-18 structure, consisting of an initial convolutional layer and multiple residual modules. Each fully connected (FC) sub-network in the five sub-networks is responsible for predicting a specific FJOA image feature, and the classification task mapping is as follows:

[0033] Subnetwork 1 is used to predict joint space narrowing, with classifications including "present" or "absent";

[0034] Subnetwork 2 is used to predict osteophyte grading, with categories including "none", "mild", "moderate" or "severe";

[0035] Subnetwork 3 is used to predict the grading of hypertrophy, with categories including "none", "mild", "moderate" or "severe";

[0036] Subnetwork 4 was used to predict the grading of subchondral bone erosion, with categories including "none", "mild" or "severe".

[0037] Subnetwork 5 was used to predict subchondral cysts, categorized as either "present" or "absent".

[0038] Preferably, the multi-task classification model is jointly trained using a cross-entropy loss function, the form of which is:

[0039]

[0040] Where K is the total number of categories, y k The one-hot encoding of the real label (1 if it belongs to class k, 0 otherwise), p k The model predicts the probability that the pixel belongs to the k-th class (softmax output, ∑pk=1).

[0041] Preferably, step S5 specifically includes the following steps:

[0042] c1. Obtain the classification output of each fully connected classification network in the two-stage multi-task deep learning model, and select the feature map output of the last convolutional layer of each fully connected classification network as the basis for generating the gradient-weighted class activation map.

[0043] c2. Based on the classification results of each FJOA feature, calculate the inverse gradient weight of the corresponding category, and weight and accumulate the weight with the feature map to obtain the heatmap corresponding to each category.

[0044] c3. Interpolate and scale each heatmap to the same size as the original intervertebral joint CT image;

[0045] c4. Overlay each heatmap onto the corresponding original CT image patch to visualize the interpretation criteria of the FJOA features;

[0046] c5. After fusing multiple heat maps, overlay them onto the original lumbar spine CT image to output a complete visual evaluation result image.

[0047] Preferably, after step S5, the model further includes performance evaluation of the one-stage model and the two-stage model. The one-stage model uses the Dice score and its 95% confidence interval to evaluate the model's ability to detect FJs. The two-stage model uses comprehensive indicators including precision, accuracy, sensitivity, specificity, F1 score, confusion matrix, and ROC curve, and evaluates the consistency between the two-stage model's predictions and human evaluations through the Gwetκ coefficient and its 95% confidence interval.

[0048] This invention discloses a deep learning-based method for evaluating imaging features of intervertebral joint osteoarthritis, which has the following beneficial effects.

[0049] This invention can simultaneously identify five types of FJOA imaging features, including joint space narrowing, osteophytes, hypertrophy, subchondral bone erosion, and subchondral cysts, improving the comprehensiveness and efficiency of the assessment. The nnU-Net model is used for high-precision segmentation of the intervertebral joint region, effectively improving localization accuracy. By introducing a ResNet-18-based shared feature extraction network and five parallel classification sub-networks, multi-scale semantic information joint modeling is achieved, improving the model's recognition accuracy and generalization ability. Grad-CAM technology is used to generate activation heatmaps, overlaying the model's interpretation criteria onto the original image to visualize the assessment results, enhancing the model's interpretability and clinical applicability. Attached Figure Description

[0050] Figure 1 It is the bounding box of the articular surfaces (FJ) from the first lumbar vertebra to the first sacral vertebra;

[0051] Figure 2 This is a flowchart of the deep learning-based imaging feature assessment method for intervertebral joint osteoarthritis proposed in this application;

[0052] Figure 3 This is the performance of the deep learning (DL) model of this application on the internal test dataset;

[0053] Figure 4 This is the performance of the deep learning (DL) model of this application on an external test dataset;

[0054] Figure 5 This is a heatmap of the FJOA image feature assessment in this application.

[0055] Figure 6 This is a radar chart showing the accuracy of assessments by two spinal radiologists with and without the assistance of a deep learning (DL) model;

[0056] Figure 7 This invention is a flowchart showing the selection of patients from internal and external datasets. Detailed Implementation

[0057] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0058] In this document, the term "embodiment" means that a specific feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The term "embodiment" appearing in various places throughout the specification does not necessarily refer to the same embodiment, nor does it specifically limit its independence or connection with other embodiments. In principle, in this application, as long as there are no technical contradictions or conflicts, the technical features mentioned in each embodiment can be combined in any way to form corresponding implementable technical solutions.

[0059] Unless otherwise defined, the technical terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains; the use of related terms herein is merely for the purpose of describing particular embodiments and is not intended to limit this application.

[0060] In the description of this application, the term "and / or" is used to describe the logical relationship between objects, indicating that three relationships can exist. For example, A and / or B means: A exists, B exists, and A and B exist simultaneously. Additionally, the character " / " in this document generally indicates that the preceding and following objects have an "or" logical relationship.

[0061] Unless otherwise expressly specified or limited, the terms "installation," "connection," "linking," "fixing," and "setting," as used in the description of the embodiments of this application, should be interpreted broadly. For example, "connection" can be a fixed connection, a detachable connection, or an integral setting; it can be a mechanical connection, an electrical connection, or a communication connection; it can be a direct connection or an indirect connection through an intermediate medium; it can be the internal connection of two components or the interaction between two components. For those skilled in the art to which this application pertains, the specific meaning of the above terms in the embodiments of this application can be understood according to the specific circumstances.

[0062] Example

[0063] The intervertebral joints (FJs), as the only synovial joints of the spine, are an important component of the "three-joint complex," serving multiple functions including weight-bearing, motion guidance, and stability maintenance; they are a major source of mechanical spinal pain. Their degenerative changes manifest as osteoarthritis (FJOA), and CT, due to its superior ability to visualize bone structures, is the preferred imaging method for assessing the imaging features of FJOA. The imaging features of FJOA (based on the Weishuapt grading system) include joint space narrowing (JSN), osteophytes and facet hypertrophy, subchondral bone erosion, and subchondral cysts. Higher grades indicate more severe structural degeneration; however, the relationship between FJOA grade and pain or prognosis remains unclear. Studies have found that while features such as osteophytes and hypertrophy can compress neural structures (spinal canal / lateral recess / intervertebral foramen) and cause pain, they may also enhance spinal stability and predict a positive response to interventional treatment—highlighting the clinical value of individualized assessment of FJOA features compared to overall severity grading. Current reliance on manual assessment has significant limitations, with low inter-reader consistency (κ value typically 0.4–0.6) leading to contradictory conclusions in epidemiological and efficacy studies. Therefore, there is an urgent need to develop reliable tools to accurately quantify the characteristics of FJOA. While deep learning (DL) algorithms have demonstrated high accuracy and objectivity in medical image analysis, they have not yet been applied to systematic quantitative studies of individual FJOA image characteristics.

[0064] To address the problems existing in the current technology, such as Figure 2 As shown, a deep learning-based method for evaluating imaging features of intervertebral joint osteoarthritis is proposed. This method includes the following steps:

[0065] S1. Collect lumbar CT images of patients with low back pain, and select image slices of bilateral complete FJ structures from all images for the L1 / L2 to L5 / S1 spinal segments to construct a dataset for the first stage of model training.

[0066] S2. Construct a one-stage model for automatic segmentation of the FJ region. Input the lumbar CT image to be processed into the one-stage model. The one-stage model uses a segmentation network based on nnU-Net to process the lumbar CT image, outputs a binary segmentation mask corresponding to the bilateral FJ regions, and crops to obtain multiple FJCT patches.

[0067] S3. Construct a two-stage model training dataset for FJOA evaluation using the FJCT patch output by the one-stage model.

[0068] S4. Construct a two-stage multi-task deep learning model based on the ResNet-18 backbone network. Input the FJCT patch output by the first-stage model into the second-stage model. In the second-stage model, extract multi-scale semantic features for each FJCT patch as a shared representation and input them into five parallel fully connected classification networks. Each network outputs a different type of FJOA feature.

[0069] S5. Generate a gradient-weighted class activation map based on the classification results of each FJOA feature, and overlay it onto the original lumbar spine CT image to form a visual evaluation image.

[0070] The technical solution provided by this invention can simultaneously identify five types of FJOA imaging features, including joint space narrowing, osteophytes, hypertrophy, subchondral bone erosion, and subchondral cysts, improving the comprehensiveness and efficiency of the assessment. By introducing a ResNet-18-based shared feature extraction network and five parallel classification subnetworks, joint modeling of multi-scale semantic information is achieved, improving the model's recognition accuracy and generalization ability. The nnU-Net model is used for high-precision segmentation of the FJ region, effectively improving the localization accuracy. Combined with Grad-CAM technology to generate activation heatmaps, the model's interpretation criteria are superimposed on the original image, enabling a visual display of the assessment results and enhancing the model's interpretability and clinical applicability.

[0071] The above plan will be explained in detail below.

[0072] In one exemplary embodiment, such as Figure 2 As shown, a deep learning-based method for evaluating imaging features of intervertebral joint osteoarthritis is provided, including the following steps:

[0073] S1. Acquire lumbar spine CT images of patients with low back pain. From all images, select image slices of complete bilateral FJ structures for the L1 / L2 to L5 / S1 spinal segments to construct a dataset for the first stage of model training.

[0074] In this embodiment, the data were obtained from 1,400 patients who underwent lumbar spine CT scans for low back pain at an anonymous hospital as internal data. The study included all patients who underwent lumbar spine CT scans for low back pain. 268 patients with conditions such as spondylolisthesis, lumbar spondylolisthesis, lumbar fracture, previous lumbar spine surgery, or severe scoliosis were excluded.

[0075] In addition, data from 300 patients collected at an anonymous hospital were used as external data, with 72 patients excluded under the same exclusion criteria. Ultimately, the internal data included 1,132 patients, and the external data included 228 patients. It should be noted that the CT scanners used in this study differed. The internal dataset was primarily scanned using the uCT960+ scanner, which has a slice thickness of 1.0 mm, an inter-slice spacing of 0.5 mm, and a peak tube voltage of 120 kV. The external dataset was acquired using the Philips iCT256 scanner, which has a slice thickness of 0.9 mm, an inter-slice spacing of 0.7 mm, and a peak tube voltage of 120 kV. All lumbar spine CT images were stored in Digital Imaging and Communications in Medicine (DICOM) format.

[0076] Preferably, in this embodiment, annotation software is used in S1 to annotate the bilateral complete FJ structures in the image slices, and the annotated image slices are used to construct a dataset for the first-stage model training.

[0077] For example, in this embodiment, a spinal radiologist with 5 years of experience in spinal imaging (anonymous reviewer) selected representative image slices at the L1 / L2, L2 / L3, L3 / L4, L4 / L5, and L5 / S1 levels for each patient. Specifically, in this embodiment, one transverse image slice containing the complete bilateral fissure joint (FJ) structure and displaying all possible imaging features was manually selected for each segment of each patient to construct the dataset; thus providing 5,660 CT transverse image slices for the internal dataset and 984 CT transverse image slices for the external dataset. Subsequently, the anonymous reviewer systematically delineated the bilateral FJs in the CT transverse image slices using bounding boxes. Each bounding box matched the corresponding FJ size and completely surrounded the joint ( Figure 1 (The specific shape of the bounding box is shown). This results in a total of 11,320 FJ image slices in the internal dataset. The image slices with labeled boxes are then preprocessed, including resampling and normalization. Specifically, resampling involves calculating the median spacing of all 2D images in the training set along the X / Y / Z axes, and then resampling the images to this baseline spacing using trilinear interpolation. Normalization is then performed: first, outlier truncation is performed, using the 0.5% and 99.5% quantiles of the CT value distribution across the entire dataset as truncation thresholds. Based on the truncated data, the mean (μ) and standard deviation (σ) are calculated, and then... Standardization. After standardization, the dataset used for the first-stage model training was constructed. The internal dataset was randomly divided into a training set (n = 7,430), a validation set (n = 2,000), and a test set (n = 1,890). The external dataset consisted of 1,903 FJ image slices from 228 patients, collected using the same methods as the internal dataset. Figure 7 Patient selections from internal and external datasets are displayed.

[0078] It should be noted that in this embodiment, the internal dataset is responsible for training, validating, and testing the model performance. The external dataset is only used for testing. That is, there are two testing processes: one using the internal dataset and the other using the external dataset. In the field of medical deep learning, data from a single source may result in poor model generalization, while the performance of a model tested with external data is considered to be more generalizable and convincing due to factors such as different imaging equipment.

[0079] S2. Construct a one-stage model for automatic segmentation of the FJ region. Input the lumbar CT image to be processed into the one-stage model. The one-stage model uses a segmentation network based on nnU-Net to process the lumbar CT image, outputs a binary segmentation mask corresponding to the bilateral FJ regions, and crops multiple FJCT patches. It should be noted that the lumbar CT image to be processed is the whole lumbar CT image of a patient with low back pain.

[0080] Preferably, in this embodiment, in S2, the dataset from the first-stage model training is input into the nnU-Net segmentation network model for training. The hyperparameter tuning of the nnU-Net segmentation network model includes the following steps:

[0081] a1. Analyze the voxel spacing distribution of all lumbar CT images in the dataset of the first-stage model training, and determine the target voxel spacing based on the median or the tenth percentile.

[0082] a2. Resample all input images to the target voxel spacing to balance spatial resolution and training efficiency;

[0083] a3. Automatically configure image patch size, batch size, and network depth based on available GPU memory capacity;

[0084] a4. During the five-fold cross-validation process, all hyperparameters except the learning rate are kept constant. The training data is divided into five subsets, and one subset is used for validation in turn while the remaining subsets are used for training to evaluate the performance of each hyperparameter combination. It should be noted that other hyperparameters are adaptively adjusted based on the dataset features and available GPU memory for key hyperparameters (including image patch size, voxel spacing, batch size, and network depth).

[0085] a5. Under the premise of fixed hyperparameters, the learning rate decreasing strategy was adopted, and 1×10-3 was finally determined as the optimal value.

[0086] a6. Compare the performance of the three optimizers: stochastic gradient descent, Adam, and AdamW, and retain the optimizer that performs best on the task in this dataset.

[0087] In this embodiment, due to GPU memory limitations, we retained the image patch size (320×320) and batch size automatically inferred by nnU-Net. Furthermore, the number of training epochs was reduced from the default 1000 to 100.

[0088] Preferably, in this embodiment, the post-processing of image patches and prediction essentially involves filling irregular mask boundaries into regular rectangles. The core is to calculate the minimum bounding rectangle of the mask and then fill all pixels within that rectangle. This method ensures that the boundaries are strictly aligned with the coordinate axes, generating a regular rectangular mask. The binary segmentation mask output in S2 is then processed through the following post-processing steps to generate regular FJCT patches:

[0089] b1. Calculate the minimum row index i of all pixels in the mask that have a value of 1. min Maximum row index i max Minimum column index j min Maximum column index j max ;

[0090] The row range is:

[0091]

[0092] The column range is:

[0093]

[0094] b2. Construct a rectangular region between the minimum and maximum indices, set the pixel values ​​within this region to 1, and the rest to 0, to obtain a regular rectangular filled mask.

[0095] The filled mask is defined as follows:

[0096]

[0097] b3. Use a fill mask to crop the original CT image to obtain regular intervertebral joint image patches, and adjust them uniformly to the preset size.

[0098] In this embodiment, the obtained regular intervertebral joint image patch is preprocessed. The preprocessing includes unifying the image pixels to the 0-1 range, scaling the image block to a size of 320*320, and horizontally flipping the image block.

[0099] Z-score transformation was used to standardize the data of all extracted intervertebral joint image patches:

[0100]

[0101] The scaling algorithm is as follows:

[0102]

[0103] The horizontal flipping algorithm is as follows:

[0104] I flipped (x ′ ,y ′) = I resized (319-x ′ ,y ′ );

[0105] The final mapped coordinates are:

[0106]

[0107] S3. Construct a two-stage model training dataset for FJOA evaluation using the FJCT patch output by the one-stage model.

[0108] Preferably, in this embodiment, in S3, two radiologists with experience in spinal imaging independently use annotation software to annotate the same FJCT patch with image labels. When the two doctors' image labels are inconsistent, a third doctor makes the decision to generate a standard reference label. The annotated FJCT patch is then used to construct a two-stage model training dataset for FJOA evaluation.

[0109] In this exemplary embodiment, image evaluation was performed using the medical image annotation software Pair (version 2.7, Shenzhen, China). The reader received decision support through the example images provided by Weishaopt et al. It should be noted that (in 1999, Weishaopt et al. published a paper entitled "MR Imaging and CT in Osteoarthritis of the Lumbar Facet") The SCI paper "Joints" describes Weishaopt et al.'s definition of FJ degeneration features, including: 1. Normal joint space width defined as 2-4 mm; 2. Osteophytes graded as "none," "mild," "moderate," or "severe"; 3. Hypertrophy, graded as "none," "mild," "moderate," or "severe"; 4. Subchondral bone erosion, graded as "none," "mild," or "severe"; 5. Subchondral cysts, classified as "none" or "present." Since only the joint space has a clearly defined quantitative indicator of 2 mm, the severity of the other four features (mild, moderate, and severe) relies entirely on the physician's subjective judgment. Therefore, we provided example images from Weishaopt et al.'s article to our two assessing physicians for reference, suggesting that they should strive to maintain consistency with Weishaopt et al.'s assessment results when subjectively judging the severity of a feature. This is to avoid misjudging mild osteophytes as severe due to the physician's personal experience.

[0110] When the two radiologists agree, their rating is used as the ground truth. If there is disagreement, a third, anonymous spinal radiologist with eight years of experience re-evaluates the images and determines the ground truth. The Gwetκ coefficient is used to assess the inter-observer reliability of the two initial radiologists for each image feature.

[0111] It should be noted that the Gwetκ statistic is an improved method for assessing inter-rater reliability (IRR) of categorical data, proposed by statistician Kilem Gwet. The basic formula is:

[0112] First, predefine the evaluation results of two initial viewers for each feature, such as none = 0, present = 1, mild = 2, moderate = 3, severe = 4.

[0113] Then observational consistency (p) was calculated. a The specific formula is: Simply put, it means counting the number of cases where two assessors completely agree, for example: 42 out of 50 cases are consistent → p a =42 / 50=0.84.

[0114] Then, the corrected expected consistency (p) is calculated. e The specific formula is as follows: Simply put, it involves calculating the marginal probability of each category, such as the positive proportion π. + = Number of positive results from the first assessor + Number of positive results from the second assessor / 2 n Negative ratio: π - =1-π + Finally, the Gwetκ coefficient is calculated.

[0115] Preferably, in this embodiment, the evaluation criteria for image labels include the following five categories of FJOA image features:

[0116] Joint space narrowing is defined as less than 2 mm and is classified as "none" or "present".

[0117] Bone spurs are graded as "none", "mild", "moderate" or "severe";

[0118] Hypertrophy is classified as "none", "mild", "moderate" or "severe";

[0119] Subchondral bone erosion is graded as "none", "mild" or "severe";

[0120] Subchondral cysts are classified as "absent" or "present".

[0121] S4. Construct a two-stage multi-task deep learning model based on the ResNet-18 backbone network. Input the FJCT patch output by the first-stage model into the second-stage model. In the second-stage model, extract multi-scale semantic features for each FJCT patch as a shared representation and input them into five parallel fully connected classification networks. Each network outputs a different type of FJOA feature.

[0122] Preferably, in this embodiment, the two-stage multi-task deep learning model includes a shared feature extraction backbone network and five parallel fully connected sub-networks. The backbone network is a ResNet-18 structure, consisting of an initial convolutional layer and multiple residual modules. It should be noted that this model uses an improved ResNet-18 as the shared feature extraction backbone network, and its structure includes:

[0123] 1) Initial processing layer: 7×7 convolutional kernel (64 output channels, stride 2, padding 3) → batch normalization → ReLU activation → 3×3 max pooling (stride 2, padding 1);

[0124] 2) Four residual modules: Each module contains two basic residual blocks, where the residual block structure is [3×3 convolution → BN → ReLU → 3×3 convolution → BN] + identity mapping / projection shortcut connection → ReLU, specifically configured as follows: layer1 (2 residual blocks, 64 channels, no downsampling) → layer2 (first block downsampling: main path stride 2, shortcut 1×1 convolution stride 2; 128 channels) → layer3 (similar to downsampling, 256 channels) → layer4 (similar to downsampling, 512 channels);

[0125] 3) Feature aggregation layer: The 7×7 feature map is compressed into a 1×1×512 feature vector by global average pooling.

[0126] Based on this, the model connects five independent fully connected sub-networks in parallel. Each sub-network directly receives a 512-dimensional feature vector and maps it directly to the category space of the corresponding task through a single linear transformation layer (no hidden layers, no activation functions, and no dropout), forming an end-to-end multi-task learning architecture, with all parameters jointly optimized.

[0127] Each of the five subnetworks (FC) is responsible for predicting a specific FJOA image feature. The classification task mapping is as follows:

[0128] Subnetwork 1 is used to predict joint space narrowing, with classifications including "present" or "absent";

[0129] Subnetwork 2 is used to predict osteophyte grading, with categories including "none", "mild", "moderate" or "severe";

[0130] Subnetwork 3 is used to predict the grading of hypertrophy, with categories including "none", "mild", "moderate" or "severe";

[0131] Subnetwork 4 was used to predict the grading of subchondral bone erosion, with categories including "none", "mild" or "severe".

[0132] Subnetwork 5 was used to predict subchondral cysts, categorized as either "present" or "absent".

[0133] Preferably, in this embodiment, the multi-task classification model is jointly trained using the cross-entropy loss function, which has the following form:

[0134]

[0135] Where K is the total number of categories, y k The one-hot encoding of the real label (1 if it belongs to class k, 0 otherwise), p k The model predicts the probability that the pixel belongs to the k-th class (softmax output, ∑pk=1).

[0136] It should be noted that in this embodiment, the open-source deep learning framework PyTorch (V3.7) based on Python is used to construct a deep learning-based multi-task small joint diagnostic model. The two-stage model training dataset is divided into training and testing datasets in an 8:2 ratio, and the training data is used to construct the multi-task small joint diagnostic model. In this embodiment, the second-stage model training uses the same training rate and batch size, and all training processes are completed on an Ubuntu 18.04 system equipped with two NVIDIA TITAN RTX GPUs, undergoing 100 training cycles.

[0137] S5. Generate a gradient-weighted class activation map based on the classification results of each FJOA feature, and overlay it onto the original lumbar spine CT image to form a visual evaluation image.

[0138] Preferably, in this embodiment, S5 specifically includes the following steps:

[0139] c1. Obtain the classification output of each fully connected classification network in the two-stage multi-task deep learning model, and select the feature map output of the last convolutional layer of each fully connected classification network as the basis for generating the gradient-weighted class activation map.

[0140] c2. Based on the classification results of each FJOA feature, calculate the inverse gradient weight of the corresponding category, and weight and accumulate the weight with the feature map to obtain the heatmap corresponding to each category.

[0141] c3. Interpolate and scale each heatmap to the same size as the original FJCT image;

[0142] c4. Overlay each heatmap onto the corresponding original CT image patch to visualize the interpretation criteria of FJOA features;

[0143] c5. After fusing multiple heat maps, overlay them onto the original lumbar spine CT image to output a complete visual evaluation result image.

[0144] Preferably, in this embodiment, after S5, the performance evaluation of the one-stage model and the two-stage model is also included. The one-stage model uses the Dice score and its 95% confidence interval to evaluate the model's ability to detect FJ. The two-stage model uses comprehensive indicators including precision, accuracy, sensitivity, specificity, F1 score, confusion matrix and ROC curve, and evaluates the consistency between the two-stage model prediction and the human evaluation through the Gwetκ coefficient and its 95% confidence interval.

[0145] For example, in this embodiment, model performance was evaluated by anonymous readers and IT engineers specializing in image analysis (i.e., anonymous image analysis experts) using internal and external test datasets. In the first phase, the model's ability to detect FJOA was evaluated using the Dice score and its 95% confidence interval. In the second phase, the classification task (evaluating each feature of FJOA), we employed comprehensive metrics including precision, accuracy, sensitivity, specificity, F1 score, confusion matrix, and ROC curve. Inter-reader consistency and model reliability were evaluated using the Gwetκ coefficient and its 95% confidence interval. The consistency level is interpreted as follows: less than 0 indicates poor, 0-0.2 indicates slight, 0.2-0.4 indicates average, 0.4-0.6 indicates moderate, 0.6-0.8 indicates significant, and 0.8-1 indicates near perfect. To understand and visualize the decision-making process of the DL model in FJOA feature evaluation, we employed the Gradient Weighted Class Activation Map (Grad-CAM) method. This method generates heatmaps that locate highly activated regions in the image. Furthermore, paired t-tests (SPSS, version 27.0, IBM, Chicago, Illinois, USA) were used to analyze reader performance with and without DL assistance. We set a significance threshold of p < 0.05 and used an accuracy radar chart to visualize trends in reader evaluations.

[0146] In this embodiment, a total of 1,360 patients were included. The internal dataset included 1,132 patients (mean age 46.1 years ± 15.0 (standard deviation); 568 females; mean body mass index 24.5 kg / m²). 2 ±4.1). The external dataset contained 228 patients (mean age 51.5 years ± 16.8 years; 119 females; mean body mass index 23.6 kg / m²). 2±3.2). A total of 13,223 FJCT images were obtained from these two datasets. The patient screening process for both datasets is detailed in [link to documentation]. Figure 7 The patient demographic information is summarized in Table 1. Table 1 details the distribution of the true scores in the internal and external datasets.

[0147] Table 1. Demographic characteristics of the internal and external datasets.

[0148]

[0149]

[0150] Note: Values ​​are expressed as the number of participants, with percentages in parentheses. Mean data is presented as ± standard deviation; data in parentheses are percentages.

[0151] Across all datasets and image features, the level of inter-reader agreement between two reference radiologists in determining the true extent ranged from moderate to near perfect. The overall agreement Gwetκ value was 0.81 (95% CI: 0.79–0.83) for internal data and 0.71 (95% CI: 0.69–0.74) for external data. Agreement varied for specific features: subsynovial osteitis showed near perfect agreement, with 0.99 (95% CI: 0.99–1.00) for internal data and 0.86 (95% CI: 0.84–0.88) for external data. Osteophytes also showed near perfect agreement, with 0.85 (95% CI: 0.84–0.87) for internal data and 0.81 (95% CI: 0.79–0.83) for external data. The concordance rates for hypertrophy were 0.88 (internal, 95% CI: 0.86–0.90) and 0.52 (external, 95% CI: 0.49–0.55). The concordance rates for subchondral bone erosion were nearly perfect, with 0.90 (95% CI: 0.89–0.92) for internal data and 0.89 (95% CI: 0.88–0.91) for external data. The concordance rates for subchondral cysts were 0.58 (internal, 95% CI: 0.54–0.61) and 0.57 (external, 95% CI: 0.53–0.60), respectively.

[0152] After training on 9,430 axial CT images from the internal dataset, the nnU-Net model achieved a Dice score of 0.81 (95% confidence interval: 0.77–0.84) in the first stage, demonstrating its effectiveness. Subsequently, the model was used to segment the remaining 3,793 CT images from both the internal and external test datasets. To ensure the accuracy of the results, all images were reviewed by anonymous reviewers, who confirmed that the regions predicted by nnU-Net exhibited complete bilateral FJ structures.

[0153] To evaluate the performance of the DL model in feature classification, 1,890 FJ axial CT images from an internal test dataset were used. Figure 3 As shown, Figure 3 The following table shows the performance of the deep learning (DL) model on the internal test dataset. A, ROC curves for imaging feature classification of FJ osteoarthritis (FJOA), including joint space narrowing (JSN), osteophytes, hypertrophy, subarticular bone erosion (SUBERO), and subchondral cysts (SUBCYST). JSN has the highest area under the ROC curve (0.96). BF, Confusion matrix of the DL model's classification of FJOA imaging features shows a 1.1% (21 / 1890) difference between non-adjacent grades of osteophytes, 1.0% (19 / 1890) for hypertrophy, and 0.2% for subchondral bone erosion.

[0154] (4 / 1890). Results showed that the overall accuracy and AUC for JSN were 89.8% (1,697 / 1,890) and 0.96, respectively; for osteophytes, the overall accuracy and AUC were 79.6% (1,504 / 1,890) and 0.92, respectively; for hypertrophy, the accuracy and AUC were 65.5% (1,237 / 1,890) and 0.86, respectively; for subchondral bone erosion, the accuracy and AUC were 88% (1,663 / 1,890) and 0.93, respectively; and for subchondral cysts, the accuracy and AUC were 82.8% (1,565 / 1,890) and 0.91, respectively. Furthermore, the classification precision for each feature was evaluated. The classification precision for JSN was 85.6% (504 / 589) for missing features and 91.7% for present features. The classification accuracy for osteophytes was 91.6% (1,187 / 1,296); for absent osteophytes, 52.8% (255 / 483); and for mild osteophytes, 54%.

[0155] (54 / 100) for moderate, 72.8% (8 / 11) for severe. The classification precision for hypertrophy was 85.2% (386 / 453) for absent, 51.5% (456 / 886) for mild, 71.5% (392 / 548) for moderate, and 100% (3 / 3) for severe. The classification precision for subchondral bone erosion was 95.6% (1503 / 1573) for absent, 49.5% (150 / 303) for mild, and 71.4% (10 / 14) for severe. Finally, the classification precision for subchondral cysts was 88.8% (844 / 951) for absent and 76.8% (721 / 939) for present.

[0156] ROC curves for all features are as follows Figure 3 As shown in Figure A.

[0157] Detailed statistics for the five FJOA image features in the internal test dataset are as follows: Figure 3 As shown in BE. The analysis revealed the most significant differences between adjacent levels (e.g., mild vs. moderate).

[0158] Furthermore, the consistency of scores varied across different characteristics. Subchondral bone erosion showed the highest reliability, with a Gwetκ value of 0.86 (95% CI: 0.85–0.88). JSN followed closely behind, with a Gwetκ value of 0.82 (95% CI: 0.80–0.84). Overall, the DL model exhibited moderate to near-perfect reliability, as detailed in Table 2.

[0159] Table 2 Overview of the performance results of the multi-task deep learning model on the internal test dataset

[0160]

[0161]

[0162] Note: Precision is defined as the sum of true positives and false positives. The F1 score is defined as the harmonic mean of precision and recall. Sensitivity is defined as the sum of true positives and false negatives.

[0163] External validation testing used 1903 FJCT images from 228 patients. Figure 4 The following shows the performance of the deep learning (DL) model on an external test dataset. A, Receiver operating characteristic (ROC) curves for image feature classification of FJ osteoarthritis (FJOA), including joint space narrowing (JSN), osteophytes, hypertrophy, subarticular bone erosion (SUBERO), and subchondral cysts (SUBCYST). SUBERO has the highest area under the ROC curve (0.95). BF, Confusion matrix of the DL model's classification of FJOA image features shows that the difference between non-adjacent grades accounts for 1.6% in osteophytes.

[0164] (30 / 1903), accounting for 4.9% (94 / 1903) of hypertrophy and 0.7% of subchondral bone erosion.

[0165] (14 / 1903). The overall accuracy for feature classification and AUC were: JSN: 76.6%.

[0166] (1,457 / 1,903) and 0.90, osteophytes were 80.2% (1,527 / 1,903) and 0.94, and hypertrophy was 56%.

[0167] (1,066 / 1,903) and 0.81, subchondral bone erosion was 90% (1,705 / 1,903) and 0.95, and subchondral cysts were 89.8% (1,709 / 1,903) and 0.93. ROC curves for all characteristics are shown in [reference needed]. Figure 4 A.

[0168] Figure 4 BE provides detailed statistics on five FJOA imaging features evaluated in a centralized external test dataset. Subchondral bone erosion showed the highest reliability with a Gwetκ coefficient of 0.88 (95% CI: 0.86, 0.90), followed by subchondral cysts with a coefficient of 0.87 (95% CI: 0.86, 0.89). These results are summarized in Table 3.

[0169] Table 3. Overview of the performance results of the multi-task deep learning model on the external test dataset.

[0170]

[0171] Note: Precision is defined as the sum of true positives and false positives. The F1 score is defined as the harmonic mean of precision and recall. Sensitivity is defined as the sum of true positives and false negatives.

[0172] Figure 5 The original axial CT slices of FJ and its Grad-CAMs are shown. Figure 5 Axial CT images of the facet joint (AE) and FJOA; FJ, gradient-weighted classification activation map. These gradient-weighted classification activation maps are overlaid with the facet joint images to form a heatmap (red indicates high activation, blue indicates low activation). A, F, facet joint space narrowing (arrow); B, G, severe osteophytes (arrow); C, H, facet joint hypertrophy; D, I, subchondral bone erosion (SUBERO); E, J, subchondral cyst (SUBCYST). The heatmap shows the degree of importance the neural network places on abnormal regions during the evaluation process. Including JSN, (see...) Figure 5 A and 5F), osteophytes (see A and 5F) Figure 5 B and 5G), hypertrophy (see ...) Figure 5 C and 5H), subchondral bone erosion (see C and 5H), subchondral bone erosion ( Figure 5 D and 5I) and subchondral cysts (see D and 5I) and subchondral cysts ( Figure 5 E and 5J). It is worth noting that, in Figure 5 In A and 5F, the model particularly emphasized the localization of the articular surface center and considered the entire joint space when assessing articular cartilage defects. Abnormal areas were also accurately identified when assessing other imaging features. Furthermore, when displaying the same FJCT images ( Figure 5 When D and 5E), the model focuses on different areas depending on the assessment target (subchondral bone erosion and subchondral cysts), such as Figure 5 As indicated by the arrows in I and 5J, this observation suggests that the model's decision-making process mimics that of human experts by selectively focusing on pathological areas rather than simply relying on correlations between images.

[0173] Figure 6 This demonstrates the change in accuracy of the two readers' evaluations with and without the assistance of a deep learning model. Figure 6 Radar charts show the accuracy of assessments by two spinal radiologists with and without deep learning (DL) model assistance. Parts A and B show the accuracy trends of the two radiologists on the internal test dataset; Parts C and D show the accuracy trends on the external test dataset. With the assistance of the deep learning model, the accuracy of assessments of all facet joint osteoarthritis (FJOA) features by both doctors improved significantly (P<0.05). After adding the deep learning model, Reader 1 (anonymous reader) saw improvements in accuracy on multiple features. On the internal test dataset, the accuracy for synovitis improved from 60.6% (303 / 500) to 82% (410 / 500). Similarly, the accuracy for osteophytes improved from 33.8% (169 / 500) to 69.8% (349 / 500), for hypertrophy from 40.8% (204 / 500) to 59%, for subchondral bone erosion from 88.8% (444 / 500) to 90.8% (454 / 500), and for subchondral cysts from 70.2% (351 / 500) to 78.4% (392 / 500). Reader 2 (anonymous reader) also saw improved accuracy with the help of the deep learning model (p = 0.029). The accuracy for synovitis improved from 88.4% (442 / 500) to 90.4% (452 / 500), and for osteophytes from 52%.

[0174] The accuracy rate for diagnosing subchondral bone erosion improved from 42.2% (211 / 500) to 60.4% (302 / 500), for subchondral bone erosion from 78.4% (392 / 500) to 85.8% (429 / 500), and for subchondral cysts from 74.8% (374 / 500) to 81.8%.

[0175] (409 / 500). Both readers showed significant improvements in accuracy on external test datasets (p<0.001, p=0.001). These findings are summarized in Table 4. Furthermore, both readers reported a significant reduction in evaluation time after using the deep learning model, with the evaluation time per CT slice decreasing from approximately 40 seconds to approximately 15–20 seconds.

[0176] Table 4 Reader performance with and without DL model assistance.

[0177]

[0178] Note: Data represents the number of CT images of the face and joints. No duplicate images were used between evaluations. Unless otherwise stated, data in parentheses represent accuracy as a percentage. DL = Deep Learning.

[0179] Use a paired t-test for statistical analysis.

[0180] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0181] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0182] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for evaluating imaging features of intervertebral joint osteoarthritis based on deep learning, characterized in that, Includes the following steps: S1. Collect lumbar CT images of patients with low back pain, and select image slices of complete intervertebral joint structures on both sides from all images for the L1 / L2 to L5 / S1 spinal segments to construct a dataset for the first stage of model training. S2. Construct a one-stage model for automatic segmentation of the intervertebral joint region. Input the lumbar CT image to be processed into the one-stage model. The one-stage model uses a segmentation network based on nnU-Net to process the lumbar CT image, outputs a binary segmentation mask corresponding to the bilateral intervertebral joint regions, and crops out multiple intervertebral joint CT patches. S3. Construct a two-stage model dataset for FJOA assessment using the intervertebral joint CT patch output from the one-stage model. S4. Construct a two-stage multi-task deep learning model based on the ResNet-18 backbone network. Input the intervertebral joint CT patch output by the first-stage model into the second-stage model. In the second-stage model, extract multi-scale semantic features for each intervertebral joint CT patch as a shared representation and input them into five parallel fully connected classification networks. Each network outputs a different type of FJOA feature. S5. Generate a gradient-weighted class activation map based on the classification results of each FJOA feature, and overlay it onto the original lumbar spine CT image to form a visual evaluation image.

2. The method for evaluating imaging features of intervertebral joint osteoarthritis based on deep learning as described in claim 1, characterized in that, In step S1, annotation software is used to annotate the bilateral complete intervertebral joint structures in the image slices, and the annotated image slices are used to construct a dataset for the first-stage model training.

3. The method for evaluating imaging features of intervertebral joint osteoarthritis based on deep learning as described in claim 1, characterized in that, In step S2, the dataset used for the first-stage model training is input into the nnU-Net segmentation network model for training. The hyperparameter tuning of the nnU-Net segmentation network model includes the following steps: a1. Analyze the voxel spacing distribution of all lumbar CT images in the dataset of the first-stage model training, and determine the target voxel spacing based on the median or the tenth percentile. a2. Resample all input images to the target voxel spacing to balance spatial resolution and training efficiency; a3. Automatically configure image patch size, batch size, and network depth based on available GPU memory capacity; a4. During the five-fold cross-validation process, keep all hyperparameters except the learning rate unchanged, divide the training data into five subsets, use one subset for validation in turn, and use the remaining subsets for training, in order to evaluate the performance of each hyperparameter combination. a5. Under the premise of fixed hyperparameters, the learning rate decreasing strategy was adopted, and 1×10-3 was finally determined as the optimal value. a6. Compare the performance of the three optimizers: stochastic gradient descent, Adam, and AdamW, and retain the optimizer that performs best on the task in this dataset.

4. The method for evaluating imaging features of intervertebral joint osteoarthritis based on deep learning as described in claim 1, characterized in that, The binary segmentation mask output in S2 is processed through the following post-processing steps to generate regular intervertebral joint CT patches: b1. Calculate the minimum row index i of all pixels in the mask that have a value of 1. min Maximum row index i max Minimum column index j min Maximum column index j max ; b2. Construct a rectangular region between the minimum and maximum indices, set the pixel values ​​within this region to 1, and the rest to 0, to obtain a regular rectangular filled mask; b3. Use the filling mask to crop the original CT image to obtain regular intervertebral joint CT patches, and adjust them uniformly to a preset size.

5. The method for evaluating imaging features of intervertebral joint osteoarthritis based on deep learning as described in claim 1, characterized in that, In S3, two radiologists with experience in spinal imaging independently use annotation software to annotate the same intervertebral joint CT patch with image labels. When the two doctors' image labels are inconsistent, a third doctor makes the decision to generate a standard reference label. The annotated intervertebral joint CT patch is used to construct a two-stage model training dataset for FJOA evaluation.

6. The method for evaluating imaging features of intervertebral joint osteoarthritis based on deep learning as described in claim 5, characterized in that, The evaluation criteria for the image labels include the following five categories of FJOA image features: Joint space narrowing, defined as less than 2 mm, is classified as "none" or "present"; Bone spurs are graded as "none", "mild", "moderate" or "severe"; Hypertrophy is classified as "none", "mild", "moderate" or "severe"; Subchondral bone erosion is graded as "none", "mild" or "severe"; Subchondral cysts are classified as "absent" or "present".

7. The method for evaluating imaging features of intervertebral joint osteoarthritis based on deep learning as described in claim 1, characterized in that, The two-stage multi-task deep learning model includes a shared feature extraction backbone network and five parallel fully connected sub-networks. The backbone network is a ResNet-18 structure, consisting of an initial convolutional layer and multiple residual modules. Each of the five fully connected sub-networks is responsible for predicting a specific FJOA image feature, and the classification task mapping is as follows: Subnetwork 1 is used to predict joint space narrowing, with classifications including "present" or "absent"; Subnetwork 2 is used to predict osteophyte grading, with categories including "none", "mild", "moderate" or "severe"; Subnetwork 3 is used to predict the grading of hypertrophy, with categories including "none", "mild", "moderate" or "severe"; Subnetwork 4 is used to predict the grading of subchondral bone erosion, with categories including "none", "mild" or "severe"; Subnetwork 5 was used to predict subchondral cysts, with classifications including "present" or "absent".

8. The method for evaluating imaging features of intervertebral joint osteoarthritis based on deep learning as described in claim 7, characterized in that, The multi-task classification model is jointly trained using the cross-entropy loss function, which has the following form: Where K is the total number of categories, y k The one-hot encoding of the real label (1 if it belongs to class k, 0 otherwise), p k The model predicts the probability that the pixel belongs to the k-th class (softmax output, ∑pk=1).

9. The method for evaluating imaging features of intervertebral joint osteoarthritis based on deep learning as described in claim 1, characterized in that, S5 specifically includes the following steps: c1. Obtain the classification output of each fully connected classification network in the two-stage multi-task deep learning model, and select the feature map output of the last convolutional layer of each fully connected classification network as the basis for generating the gradient-weighted class activation map. c2. Based on the classification results of each FJOA feature, calculate the inverse gradient weight of the corresponding category, and weight and accumulate the weight with the feature map to obtain the heatmap corresponding to each category. c3. Interpolate and scale each heatmap to the same size as the original intervertebral joint CT image; c4. Overlay each heatmap onto the corresponding original CT image patch to visualize the interpretation criteria of the FJOA features; c5. After fusing multiple heat maps, overlay them onto the original lumbar spine CT image to output a complete visual evaluation result image.

10. The method for evaluating imaging features of intervertebral joint osteoarthritis based on deep learning as described in claim 1, characterized in that, The S5 section further includes performance evaluation of the one-stage and two-stage models. The one-stage model uses the Dice score and its 95% confidence interval to evaluate the model's ability to detect FJs. The two-stage model uses comprehensive indicators including precision, accuracy, sensitivity, specificity, F1 score, confusion matrix, and ROC curve, and evaluates the consistency between the two-stage model's predictions and human evaluations using the Gwetκ coefficient and its 95% confidence interval.

Images