基于CT影像多任务深度学习模型预测股骨头坏死病变分割及塌陷风险OA

BACKGROUND:Previous studies have developed radiomics-based machine learning models for the diagnosis and staging of osteonecrosis of the femoral head.However,uncertainty in collapse prediction results may lead to unnecessary early intervention,and the lack of follow-up data from non-surgical patients has limited research on collapse prediction. OBJECTIVE:To develop a machine learning model based on CT images and establish an automated pipeline for predicting the risk of femoral head collapse. METHODS:Using the hip joint as the analysis unit,99 patients with Association Research Circulation Osseous(ARCO)stage Ⅰ and Ⅱ osteonecrosis of the femoral head(175 hips)were included.The data were divided into training and test sets at a ratio of 8:2.The proposed model employed a multi-task learning architecture with a shared encoder to simultaneously perform necrotic region segmentation and collapse risk prediction.Uncertainty modeling was introduced to quantify prediction uncertainty,enhancing model robustness and generalization.Model performance was evaluated using five-fold cross-validation. RESULTS AND CONCLUSION:In the task of femoral head collapse risk prediction,the multi-task model incorporating uncertainty modeling(MTL-Uncer)achieved an area under the curve(AUC)of 0.84±0.04 and an F1 score of 0.76±0.04,significantly outperforming the single-task model(AUC:0.71±0.07,F1:0.66±0.08).In the necrotic region segmentation task,the MTL-Uncer model achieved a Dice coefficient of 0.901±0.003 and an intersection over union of 0.819±0.004,outperforming the single-task model.Decision curve analysis showed that the MTL-Uncer model had significantly better clinical net benefit within the risk threshold range of 0.1-0.8.Overall,this study developed several deep learning models for osteonecrosis of the femoral head based on CT images,among which the MTL-Uncer model demonstrates the best predictive performance and holds promise as an effective tool for predicting femoral head collapse risk.