An automated diagnostic system for axial spondyloarthritis

Abstract

The present invention discloses an automated diagnostic system for axial spondyloarthritis, which includes a data acquisition module, a data processing module, a model building module, and a model interpretation module; the model building module performs deep learning on MRI images, and constructs a joint prediction model in combination with clinical variables to obtain prediction results for axial spondyloarthritis. The present invention improves the diagnostic accuracy and efficiency of axial spondyloarthritis by combining deep learning and machine learning methods; multi-center retrospective and prospective data are used to ensure the comprehensiveness and representativeness of the data, and the analysis of MRI images by deep learning models can more accurately identify disease-related imaging features, while the machine learning model combined with clinical variables further optimizes the prediction performance. The interpretability analysis of the model enables doctors to understand and trust the prediction results of the model, which promotes the application and promotion of artificial intelligence in medical imaging diagnosis.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image diagnosis, and in particular to an automated diagnosis system for axial spinal arthritis. Background Art

[0002] Machine learning and deep learning are two rapidly developing branches of artificial intelligence (AI). These technologies are disrupting many industries, including medical diagnosis. Through automated machine learning models, computer systems are able to learn patterns from large amounts of complex, multidimensional data, in many cases matching or exceeding the recognition capabilities of human experts. In the field of medical imaging, the widespread availability of high-resolution imaging technologies such as MRI and CT has provided doctors with a wealth of diagnostic information. However, processing and interpreting such large amounts of data is a time-consuming and highly skilled task, especially in the case of complex diseases involving delicate lesions.

[0003] Existing research has proposed an innovative diagnostic model that combines radiomic features of MRI images with clinical risk factors to create an effective scalar prediction tool. However, existing technologies suffer from reliance on expert knowledge for feature selection and extraction, high computational complexity, high sensitivity to image quality, and poor model generalization. Summary of the Invention

[0004] (1) Technical problems solved

[0005] In order to solve the above technical problems, the present invention provides an automated diagnosis system for axial spinal arthritis.

[0006] (2) Technical solution

[0007] In order to solve the above-mentioned technical problems and achieve the purpose of the invention, the present invention is implemented through the following technical solutions:

[0008] An automated diagnostic system for axial spinal arthritis includes a data acquisition module, a data processing module, a model building module, and a model interpretation module, wherein:

[0009] The data acquisition module is used to collect clinical data and MRI images of consecutive patients with chronic low back pain;

[0010] Data processing module, which is used for image annotation and incomplete clinical data processing;

[0011] The model construction module includes an MRI image deep learning model construction and training verification submodule, a clinical risk factor determination submodule and a machine learning joint model construction submodule; the MRI image deep learning model construction and training verification submodule is used for model construction and training verification, and the MRI image deep learning model includes three independent single-sequence deep learning models, which are constructed based on three MRI sequences respectively. Each single-sequence deep learning model generates a prediction score reflecting the probability of axial spondyloarthritis, and the outputs of the single-sequence models are further combined to construct an integrated deep learning model; the clinical risk factor determination submodule analyzes multiple clinical variables and selects clinical variables with statistical differences as independent clinical risk factors; the machine learning joint model construction submodule is used to construct a model that combines clinical risk factors and MRI image information. The input of the joint model is independent clinical risk factors and deep learning scores derived from the MRI-based model, and outputs the prediction results of axial spondyloarthritis;

[0012] The model interpretation module is used to interpret deep learning models of MRI images based on gradient-weighted class activation mapping, and to interpret joint machine learning models based on Shapley additive interpretation.

[0013] Furthermore, the MRI image scanning orientation includes oblique axial and oblique coronal positions.

[0014] Furthermore, the three MRI sequences include T1-weighted imaging, T2-weighted imaging, and fluid-sensitive fat suppression sequence.

[0015] Furthermore, the incomplete clinical data processing uses a multiple imputation method to estimate missing values, including multiple imputation and Z-score standardization.

[0016] Furthermore, the integrated deep learning model includes three dual-sequence deep learning models and one triple-sequence deep learning model. The three dual-sequence deep learning models correspond to pairwise combinations of T1-weighted imaging, T2-weighted imaging, and fluid-sensitive fat suppression sequences, and the triple-sequence deep learning model corresponds to T1-weighted imaging, T2-weighted imaging, and fluid-sensitive fat suppression sequences.

[0017] Furthermore, the deep learning model backbone uses the ResNet-50 architecture; the standard 2D convolutional layers and pooling layers in the ResNet block are replaced with their corresponding 3D versions.

[0018] Furthermore, the clinical risk factor determination submodule also includes using univariate logistic regression analysis to study multiple clinical variables, determining clinical variables with statistical differences as candidate clinical variables, and further performing multivariate logistic regression analysis on these candidate clinical variables to select clinical variables with statistical differences as independent clinical risk factors.

[0019] Furthermore, the multiple clinical variables include age, gender, disease duration, CRP, ESR and HLA-B27.

[0020] Furthermore, the model interpretation of the MRI image deep learning model includes generating a heat map.

[0021] Furthermore, the interpretation of the joint machine learning model includes applying Shapley additive explanatory value to clarify the contribution of each variable, wherein the variables include clinical variables and MRI-based deep learning scores, to explain the prediction results and provide a measure of feature importance.

[0022] (3) Beneficial effects

[0023] Compared with the prior art, the present invention has the following beneficial effects:

[0024] The present invention improves the diagnostic accuracy and efficiency of axSpA by combining deep learning and machine learning methods. Specifically, the use of multi-center retrospective and prospective data ensures the comprehensiveness and representativeness of the data. The analysis of MRI images through deep learning models can more accurately identify disease-related imaging features, while the machine learning model combined with clinical variables further optimizes the predictive performance. The implementation of this method not only enhances the ability to identify the disease early, but also provides clinicians with more reliable decision support. In addition, the interpretability analysis of the model (such as Grad-CAM and SHAP technology) enables doctors to understand and trust the model's prediction results, promoting the application and promotion of artificial intelligence in medical imaging diagnosis. Overall, this embodiment improves the accuracy of disease diagnosis through innovative methods and provides patients with better medical services. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0026] Figure 1 is a schematic structural diagram of an automated diagnostic system for axial spondyloarthritis according to an embodiment of the present application;

[0027] Figure 2is a schematic diagram of a data acquisition module according to an embodiment of the present application;

[0028] Figure 3 is a schematic diagram of a data processing module according to an embodiment of the present application;

[0029] Figure 4 is a schematic diagram of a model construction module according to an embodiment of the present application;

[0030] Figure 5 This is a schematic diagram of model interpretation of an MRI image deep learning model according to an embodiment of the present application;

[0031] Figure 6 is an explanation diagram of the importance of the machine learning joint model according to an embodiment of the present application;

[0032] Figure 7 This is a result interface output diagram according to an embodiment of the present application;

[0033] Figure 8 This is a diagram showing the performance of the deep learning model for MRI according to an embodiment of the present application;

[0034] Figure 9 This is a diagram showing the performance effect of the joint model according to an embodiment of the present application;

[0035] Figure 10 This is a performance effect diagram of a hierarchical analysis of the performance of a machine learning joint model according to an embodiment of the present application;

[0036] Figure 11 It is a radar chart comparing the performance of the joint model according to the embodiment of the present application and the ASAS classification standard. DETAILED DESCRIPTION

[0037] The embodiments of the present disclosure are described in detail below with reference to the accompanying drawings.

[0038] The following describes the embodiments of the present disclosure through specific examples, and those skilled in the art can easily understand other advantages and effects of the present disclosure from the contents disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of the present disclosure, rather than all of the embodiments. The present disclosure can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present disclosure. It should be noted that, in the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. Based on the embodiments in the present disclosure, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present disclosure.

[0039] It should also be noted that the illustrations provided in the following embodiments are only schematic illustrations of the basic concept of the present disclosure. The illustrations only show components related to the present disclosure and are not drawn according to the number, shape and size of components in actual implementation. In actual implementation, the type, quantity and proportion of each component can be changed at will, and the component layout type may also be more complicated.

[0040] See also Figure 1 , an automated diagnostic system for axial spinal arthritis, comprising a data acquisition module, a data processing module, a model building module and a model interpretation module, wherein,

[0041] The data acquisition module is used to collect clinical data and MRI images of consecutive patients with chronic low back pain; clinical characteristics of clinical data, including age, gender, C-reactive protein, erythrocyte sedimentation rate, etc., are obtained from clinical records and laboratory test results; MRI images are obtained from magnetic resonance imaging; optionally, patient information can be obtained from different hospital centers; the structural diagram is shown as follows Figure 2 As shown;

[0042] Inclusion criteria were as follows: (A) a diagnosis of axSpA or non-axSpA with chronic low back pain by a rheumatologist at initial visit or after at least one year of follow-up; (B) MRI and laboratory tests performed within two weeks before starting treatment; and (C) complete clinical information available in the electronic medical record. Exclusion criteria included: (A) a concurrent diagnosis of axSpA and non-axSpA; (B) prior treatment with biologic antirheumatic drugs, including tumor necrosis factor inhibitors, CD20 monoclonal antibodies, and interleukin antagonists, before MRI; and (C) poor MRI image quality due to artifacts, missing sequences, or incomplete sacroiliac joint sections.

[0043] Clinical data included demographic characteristics such as age and sex, disease duration, and various spondyloarthritis (SpA) features: inflammatory back pain, arthritis, tendinitis, anterior uveitis, dactylitis, psoriasis, Crohn's disease or ulcerative colitis, good response to nonsteroidal anti-inflammatory drugs (NSAIDs), family history of SpA, HLA-B27 status, C-reactive protein level (CRP), and erythrocyte sedimentation rate (ESR). All clinical data were optionally extracted from electronic health records.

[0044] MRI scans were performed in oblique axial and coronal planes. MRI examinations were performed using 1.5T and 3.0T scanners from different manufacturers. Three MRI sequences were used: T1-weighted imaging (T1WI), T2-weighted imaging (T2WI), and a fluid-sensitive fat-suppressed (FS) sequence.

[0045] Data processing module, which is used for image annotation and incomplete clinical data processing; the structural diagram is as follows Figure 3 As shown;

[0046] The incomplete clinical data processing adopted a multiple imputation method to estimate missing values ​​to ensure that subsequent analysis used a complete data set, which included multiple imputation and Z-score standardization.

[0047] First, multiple imputation (MI) was used. This method is a statistical technique for addressing missing data and is particularly well-suited for restoring completeness in clinical studies. By inferring and filling missing values ​​for the target variable using other known clinical variables, it can reduce analytical bias caused by missing values. The open-source Python package MiceForest was used to perform multiple imputation of missing clinical information. This method uses other clinical variables (such as age, gender, and disease duration) to predict missing values ​​for target variables (such as CRP and ESR), allowing for simultaneous processing of continuous and categorical variables. Subsequently, clinical variables were Z-score standardized to enhance the comparability and validity of different features in subsequent analyses and modeling. Z-score standardization converts the data to a distribution with a mean of 0 and a standard deviation of 1 by subtracting the feature's mean from each data point and dividing the result by the feature's standard deviation. This process eliminates the influence of different dimensions and scales on model performance, thereby improving model convergence speed and predictive accuracy.

[0048] Image annotation involves manually delineating and reviewing the MRI images using bounding boxes. Bounding boxes encompass the ilium, sacrum, and bilateral sacroiliac joints. To improve annotation consistency, the bounding box is optionally defined using the largest slice of the ilium as the standard slice and then uniformly extended to other sacroiliac joint MRI slices. This approach not only provides a stable reference framework for research but also effectively forms a three-dimensional bounding box, enhancing the spatial representation of the data.

[0049] All annotated images were preprocessed to ensure consistency and reliability. This process included applying N4 bias field correction to mitigate intensity inhomogeneities; resampling the images to achieve an in-plane pixel size of 1.0 mm × 1.0 mm; and, after resampling, randomly cropping the images to bounding boxes of 12 × 256 × 512 dimensions and uniformly normalizing them to the range [-1, 1], a process known as signal normalization. Furthermore, to increase data diversity and enhance model robustness, a horizontal flipping procedure was implemented. This operation, performed on each randomly cropped image, effectively improved the model's adaptability to images with different orientations. Overall, these preprocessing steps ensured the quality and consistency of the MRI images for subsequent analysis and deep learning model training.

[0050] Model building blocks such as Figure 4 As shown, it includes the construction and training verification submodule of the MRI image deep learning model, the clinical risk factor determination submodule, and the machine learning joint model construction submodule;

[0051] The MRI image deep learning model construction and training verification submodule is used for model construction and training verification;

[0052] The model is constructed based on MRI images. The deep learning (DL) model first constructed includes three independent single-sequence DL models, which are constructed based on three MRI sequences (T1WI, T2WI and FS). Each single-sequence DL model generates a prediction score reflecting the probability of axSpA.

[0053] Subsequently, the outputs of the single-sequence models were further combined to construct an ensemble DL model. This ensemble DL model includes three two-sequence DL models (T1WI+T2WI, T1WI+FS, and T2WI+FS) and one three-sequence DL model (T1WI+T2WI+FS). The final output of the ensemble DL model is obtained by summing and averaging the prediction scores of each individual model. This process effectively integrates information from multiple sources, thereby enhancing the ability to identify axSpA and increasing the practical value of the model in clinical applications.

[0054] Based on the ResNet-50 architecture's outstanding performance and robustness in medical image analysis across various datasets, the proposed DL model uses the ResNet-50 architecture as its backbone. The standard 2D convolutional and pooling layers in the ResNet blocks are replaced with their corresponding 3D versions. In the modified 3D-ResNet50 network structure, the input dimensions are set to (12, 256, 512), the initial convolutional layer has 1 input channel and 64 output channels, the kernel sizes are (3, 7, 7), and the strides are set to (1, 2, 2) with appropriate padding. The network includes multiple bottlenecks. The first set of three bottleneck blocks increases the number of input channels from 64 to 256 and downsamples them, while subsequent sets maintain 256 output channels. The second set of four bottleneck blocks increases the number of channels to 512 and downsamples them, while the third set of six bottleneck blocks further increases the number to 1024 channels. The final set of three bottleneck blocks increases the number of channels to 2048. The fully connected layer takes 2051 input features and outputs a Softmax layer with a feature of 1 for the final output. All convolutional layers within the bottleneck structure are normalized by a BatchNormalization layer and use ReLU activation functions where appropriate. The modified bottleneck structure contains multiple convolutional layers, where the first and last convolutional layers use a convolution kernel of (1,1,1); while intermediate convolution layers use a convolution kernel of (3,3,3). The stride and padding vary depending on whether or not the convolution is downsampled. The overall design is designed to better adapt to anisotropic volumetric data.

[0055] Model training and validation included subgroup analyses to assess the performance of the MRI-based models across all test sets. Each DL model underwent five-fold cross-validation on the training set for model validation and epoch selection. To maintain a balanced ratio of positive and negative samples across the entire patient dataset, each fold was stratified accordingly. To address class imbalance, a binary cross-entropy loss function was used, with class weights adjusted inversely with the frequency of each class. The optimal model parameters were selected based on achieving the highest area under the receiver operating characteristic curve (AUC) during validation. For testing, the outputs of the five-fold cross-validation models were averaged.

[0056] Optionally, each DL model is trained for 150 epochs using the Adam optimizer, with a learning rate of 0.0003, moving average coefficients β_1 of 0.9, β_2 of 0.999, weight decay coefficient of 0.01, and a mini-batch size of 8. The binary cross entropy loss function is used.

[0057] The clinical risk factor determination submodule is used to identify clinical risk factors. Univariate logistic regression analysis was performed on six clinical variables: age, sex, disease duration, CRP, ESR, and HLA-B27. Clinical variables with statistically significant differences (P < 0.05) were identified as candidate clinical variables. Subsequently, multivariate logistic regression analysis was performed on these candidate clinical variables, and clinical variables with statistically significant differences (P < 0.05) were identified as independent clinical risk factors.

[0058] The machine learning joint model construction submodule is used to construct a model that combines clinical risk factors and MRI image information. The joint model inputs are independent clinical risk factors and the DL score derived from the MRI-based model, and outputs a prediction result for axSpA. The machine learning separator of the joint model is one of logistic regression, naive Bayes, K-nearest neighbor, support vector machine, decision tree, and random forest.

[0059] The model interpretation module is used to interpret the deep learning model of MRI images based on gradient-weighted class activation mapping (Grad-CAM) and the joint machine learning model based on Shapley additive interpretation (SHAP).

[0060] Furthermore, model interpretation of deep learning models for MRI images includes generating heat maps, such as Figure 5 As shown in the figure, the first column is the original MRI image; the second example is the heat map generated by the deep learning model based on MRI; the third column is the image generated by overlapping the original image and the heat map. Red areas represent areas where the model pays high attention, and blue areas represent areas where the model pays low attention.

[0061] Interpretation of the joint machine learning model included applying Shapley additive explanations (SHAP) values ​​to elucidate the contribution of each variable, including age, sex, HLA-B27, and MRI-based DL scores, which facilitated the interpretation of predictions and provided a measure of feature importance.

[0062] like Figure 6 As shown in the figure, the vertical axis represents the input variables of the machine learning model, and the horizontal axis represents the degree of influence of each input variable on the model. The input variables on the vertical axis are sorted from top to bottom according to their influence on the model output. Red represents risk factors for axSpA, and blue represents protective factors for axSpA.

[0063] For the joint machine learning model, Shapley additive explanations (SHAP) values ​​were applied to clarify the contribution of each variable (age, sex, HLA-B27, and MRI-based DL score). This approach facilitates the interpretation of predictions and provides a measure of feature importance.

[0064] like Figure 7 The output of the results interface of the present invention is shown. The first column shows the original MRI image; the second column shows the heatmap generated by the MRI-based deep learning model; and the third column shows the image generated by overlaying the original image and the heatmap. Red areas represent areas of high model attention, while blue areas represent areas of low model attention. The optimal cutoff value of the MRI deep learning model of the present invention was 0.592, with a specificity of 0.75, balancing sensitivity (65.4%–74.0%) and specificity (51.9%–52.6%) across the external test set. The combined ML model used an optimal cutoff value of 0.637 based on the highest Youden index.

[0065] like Figure 8 The figure shows the performance of the MRI-based deep learning model. The AUC in the figure is the area under the receiver operating characteristic curve.

[0066] Among the deep learning models based on single-sequence and dual-sequence MRI, the T1-weighted image + fat-suppressed image model showed the highest AUC of 0.837 (95% CI: 0.778–0.891) on the internal test set ( Figure 8 a). It is worth noting that there was no statistical difference between the performance of the T1-weighted image + fat-suppressed image and T1-weighted image + T2WI + fat-suppressed image models (p > 0.05). Since the T1-weighted image + fat-suppressed image model performed well and T1-weighted image and fat-suppressed image are sequences recommended by the ASAS standard, we selected the T1-weighted image + fat-suppressed image model as the final MRI-based model. The AUC of this model on three external test sets (sets I, II, and III, respectively) were 0.636 (95% CI: 0.452-0.800), 0.724 (95% CI: 0.589-0.847), and 0.710 (95% CI: 0.594-0.822), respectively. ( Figure 8 b).

[0067] like Figure 9 The figure shows the performance effect of the machine learning joint model. Figure annotation: AUC, area under the receiver operating characteristic curve.

[0068] Univariate and multivariate analyses identified age, sex, and HLA-B27 status as independent clinical risk factors for distinguishing axSpA from non-axSpA. Among the best performing ML joint models on the internal test set, logistic regression, naive Bayes, and KNN-7 models emerged as the best performing models ( Figure 9a). Further exploration of different K values ​​revealed that the KNN-11 model achieved the highest AUC of 0.853 (95% CI: 0.792–0.910) among all KNN-based models on the internal test set ( Figure 9 b), and was therefore selected for further analysis. The AUCs of this model on the external test sets I, II, and III were 0.872 (95% CI: 0.768–0.953), 0.780 (95% CI: 0.658–0.895), and 0.912 (95% CI: 0.841–0.966), respectively ( Figure 9 c). Using the optimal cutoff value of 0.637 based on the highest Youden index, the model demonstrated adequate specificity (57.9%–81.5%) and sensitivity (80.0%–92.3%) on the external test set.

[0069] like Figure 10 Performance effects of a hierarchical analysis of the performance of machine learning joint models; Figure caption: AUC, area under the receiver operating characteristic curve.

[0070] The joint model is more accurate in terms of age ( Figure 10 a; AUC was 0.769–0.838), gender ( Figure 10 b; AUC 0.802–0.858) and disease duration ( Figure 10 c; AUC was 0.849–0.852). In addition, it showed good performance in patients with HLA-B27 negative and unknown HLA-B27 status ( Figure 10 d; AUC was 0.756). However, its performance was relatively weak in the HLA-B27 positive subgroup (AUC was 0.570).

[0071] like Figure 11 Performance comparison of the combined model and the ASAS classification criteria. Figure captions: AUC, area under the receiver operating characteristic curve; ASAS, International Spondyloarthritis Society.

[0072] Radar plots showed that the combined model outperformed the ASAS classification standard with higher AUC (0.858 vs. 0.650, p < 0.001), accuracy (78.7% vs. 56.9%, p < 0.001), sensitivity (87.8% vs. 42.4%, p < 0.001), and F1 score (84.8% vs. 57.1%).

[0073] While conventional ASAS criteria often delay the diagnosis of axSpA, our model improves upon this, achieving higher sensitivity (87.8% vs. 42.4%, p < 0.001) and accuracy (78.7% vs. 56.9%, p < 0.001). It effectively identifies axSpA that is often overlooked due to strict adherence to the ASAS. For example, a case presenting only two features of SpA (e.g., inflammatory back pain and HLA-B27 positivity) may be misclassified if no other SpA features or radiographic sacroiliitis are present according to the ASAS criteria. However, subtle sacroiliac joint lesions may be overlooked, leading to misclassification as non-axSpA. Our model provides the correct classification in this setting, potentially reducing misdiagnosis caused by clinician fatigue and strict adherence to the ASAS. Furthermore, clinicians may sometimes misinterpret sacroiliac joint lesions in patients with non-axSpA as axSpA due to the overlap in features between the two conditions. Our model minimizes these misinterpretations by incorporating additional information beyond MRI and HLA-B27 status, including age and sex. These two variables also contribute to the diagnosis of axSpA, as the disease is more common in younger male patients. By combining SHAP with Grad-CAM visualization, clinicians can more easily spot clinical variables or image regions highlighted by the model, emphasizing the importance of a careful and comprehensive diagnostic approach.

[0074] In this embodiment, by combining deep learning and machine learning methods, the diagnostic accuracy and efficiency of axSpA are improved. Specifically, multi-center retrospective and prospective data are used to ensure the comprehensiveness and representativeness of the data. The analysis of MRI images by deep learning models can more accurately identify disease-related imaging features, while the machine learning model combined with clinical variables further optimizes the predictive performance. The implementation of this method not only enhances the ability to identify the disease early, but also provides clinicians with more reliable decision support. In addition, the interpretability analysis of the model (such as Grad-CAM and SHAP technology) enables doctors to understand and trust the model's prediction results, promoting the application and promotion of artificial intelligence in medical imaging diagnosis. Overall, this embodiment improves the accuracy of disease diagnosis through innovative methods and provides patients with better medical services.

[0075] The embodiments described above are merely descriptions of preferred implementations of the present invention and are not intended to limit the scope of the present invention. Without departing from the spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should fall within the scope of protection determined by the claims of the present invention.

Claims

1. An automated diagnostic system for axial spondyloarthritis, characterized in that: It includes data acquisition module, data processing module, model building module and model interpretation module, among which, The data acquisition module is used to collect clinical data and MRI images of consecutive patients with chronic low back pain; Data processing module, which is used for image annotation and incomplete clinical data processing; The model construction module includes a submodule for constructing and training a deep learning model for MRI images, a submodule for determining clinical risk factors, and a submodule for constructing a joint machine learning model. The deep learning model for MRI images includes three independent single-sequence deep learning models, each constructed based on three MRI sequences. Each single-sequence deep learning model generates a prediction score reflecting the probability of axial spondyloarthropathy. The outputs of the single-sequence models are combined to construct an integrated deep learning model. The joint machine learning model construction submodule is used to construct a model that combines clinical risk factors and MRI image information. The input of the joint model is the independent clinical risk factors and the deep learning scores derived from the MRI-based model, and the output is the prediction result of axial spondyloarthropathy. The clinical risk factor determination submodule uses univariate logistic regression analysis to study multiple clinical variables, determines clinical variables with statistical differences as candidate clinical variables, performs multivariate logistic regression analysis on these candidate clinical variables, and selects clinical variables with statistical differences as independent clinical risk factors; The model interpretation module is used to interpret deep learning models of MRI images based on gradient-weighted class activation mapping and to interpret joint machine learning models based on Shapley additive interpretation.

2. The automated diagnostic system for axial spondyloarthritis according to claim 1, characterized in that: The MRI image scanning orientation includes oblique axial orientation and oblique coronal orientation.

3. The automated diagnostic system for axial spondyloarthritis according to claim 1, characterized in that: The three MRI sequences include T1-weighted imaging, T2-weighted imaging, and fluid-sensitive fat-suppressed sequence.

4. The automated diagnostic system for axial spondyloarthritis according to claim 1, characterized in that: The incomplete clinical data were processed using a multiple imputation method to estimate missing values, including multiple imputation and Z-score standardization.

5. The automated diagnostic system for axial spondyloarthritis according to claim 3, wherein: The integrated deep learning model includes three dual-sequence deep learning models and one triple-sequence deep learning model. The three dual-sequence deep learning models correspond to pairwise combinations of T1-weighted imaging, T2-weighted imaging, and fluid-sensitive fat suppression sequences, respectively. The triple-sequence deep learning model corresponds to T1-weighted imaging, T2-weighted imaging, and fluid-sensitive fat suppression sequences.

6. The automated diagnostic system for axial spondyloarthritis according to claim 5, characterized in that: The deep learning model uses the ResNet-50 architecture; the standard 2D convolutional and pooling layers in the ResNet block are replaced with their corresponding 3D versions.

7. The automated diagnostic system for axial spondyloarthritis according to claim 1, characterized in that: The multiple clinical variables included age, sex, disease duration, CRP, ESR, and HLA-B27.

8. The automated diagnostic system for axial spondyloarthritis according to claim 1, characterized in that: The model interpretation of the MRI image deep learning model includes generating a heat map.

9. The automated diagnostic system for axial spondyloarthritis according to claim 8, wherein: Interpretation of the joint machine learning model included applying Shapley's additive explanatory value to elucidate the contribution of each variable, including clinical variables and MRI-based deep learning scores, to explain the predictions and provide a measure of feature importance.

Images