Multi-Task Recurrent Convolutional Network for Rib Fracture Healing Period Prediction

Abstract

Rib fracture CT image sequences present challenges such as small fracture regions, subtle inter-class differences and complex temporal changes during healing. In this work, we propose a multi-task recurrent convolutional network (MRCN) for simultaneous rib fracture healing period prediction and fracture type classification. The method uses a shared feature encoder to extract semantic features from CT images and two task-specific branches to model healing status and fracture type information. In the prediction branch, a CE-LSTM module with contextual expression is introduced to capture temporal dependencies and contextual information across CT image sequences. In the classification branch, Bi-VA and CAM modules are used for multi-scale feature fusion and fracture-region enhancement. A joint optimisation strategy enables feature sharing and complementary learning between the two tasks. Experiments on the rib fracture dataset from Shanxi Bethune Hospital show that MRCN achieves 93.20% accuracy for healing period prediction and 87.90% accuracy for fracture type classification, showing better performance than the compared methods.