This study aims to develop a multi-label classification model based on pelvic X-rays for the diagnosis and assessment of Ankylosing Spondylitis (AS), addressing the challenges of limited expert resources in underdeveloped regions. We propose a deep learning-based multi-label classification model incorporating a prior attention mechanism. This model utilizes predefined bounding boxes to accurately locate key areas, such as the bilateral sacroiliac joints and hip joint, allowing the model to focus on relevant regions during inference. The performance of the model was validated on a multi-center external test set. Experimental results demonstrate that the model achieves significant improvements in diagnostic performance and in the assessment performance of the sacroiliac joints and hip joints. The optimal model’s diagnostic performance achieved an accuracy of 0.875 on a multi-center external test set, surpassing doctors with less experience (0.798) and performing comparably to more experienced human experts (0.880). The accuracy for the assessment of left and right sacroiliac joints and left and right hip joints reached 0.964, 0.970, 0.827, and 0.863, respectively. The proposed multi-label classification model with a prior attention mechanism offers a promising, cost-effective tool for AS diagnosis and assessment, especially in resource-limited settings.

Background: Lumbar disc herniation (LDH) is a common spinal disorder leading to chronic low back pain and radiating leg pain. Minimally invasive surgery is the preferred treatment, but postoperative recovery varies significantly, making accurate preoperative prognosis prediction a challenge. Methods: This study developed and prospectively validated a multimodal risk-stratification framework integrating structured clinical variables, conventional imaging parameters, CT/MRI radiomics, and high-dimensional image-derived features. A total of 773 patients were analyzed, including 665 patients in the retrospective development cohort and 108 patients in an independent prospective validation cohort. The retrospective cohort was divided into a training cohort ( n = 532) and an internal validation cohort ( n = 133). Feature selection, parameter optimization, and classification-threshold determination were performed exclusively within the training cohort, and the finalized framework was subsequently evaluated in the internal and prospective validation cohorts. Results: The integrated multimodal approach achieved an ROC-AUC of 0.937 (95% CI, 0.900–0.975) and a PR-AUC of 0.849 in the internal validation cohort, and an ROC-AUC of 0.976 (95% CI, 0.951–1.000) and a PR-AUC of 0.941 in the prospective validation cohort. Its discrimination was comparable to that achieved using structured clinical and conventional imaging information alone and was higher than that of isolated quantitative imaging approaches. At the prespecified training-derived classification threshold of 0.318, sensitivity, specificity, and balanced accuracy in the prospective cohort were 85.3%, 95.9%, and 90.6%, respectively. Calibration analysis demonstrated residual probability miscalibration despite strong discrimination, although training-derived logistic recalibration improved the Brier score. Conclusion: Multimodal integration of clinical and quantitative imaging information demonstrated strong discrimination for 12-month postoperative outcome stratification in both internal and prospective validation cohorts. However, the incremental discriminative value beyond structured clinical and conventional imaging information was limited, and residual miscalibration remained in the prospective cohort. Further multicenter validation, standardized quantitative imaging workflows, and calibration assessment are warranted before clinical translation.

Ankylosing spondylitis (AS) is an immune-mediated disease with an unknown etiology, posing challenges in effective treatment. This study aims to investigat