A Prediction Model of Rib Fracture Healing Period Based on Multi-task Learning

Rib fractures represent a common and clinically significant manifestation of thoracic trauma, typically diagnosed using chest computed tomography (CT) imaging. However, accurate fracture type classification and prediction of healing progression pose considerable challenges in both clinical practice and intelligent analysis systems. These difficulties stem from the inherent characteristics of rib fractures: their often diminutive size, highly variable morphologies, and low contrast relative to surrounding tissues. To address these interconnected challenges, this research presents a novel multi-task learning framework specifically designed for the joint classification of rib fracture types and the prediction of their healing stages. Our methodological approach incorporates two key and synergistic components. First, for the task of fracture type classification, we introduce a Three-Branch Hierarchical Multi-Scale Fusion Model. This architecture explicitly captures and integrates complementary feature information derived from diverse hierarchical levels and spatial scales within the CT data. By fusing coarse and fine-grained features contextually, the model achieves a significantly enhanced capability to discern subtle morphological differences inherent to various fracture patterns. Second, tackling the complexities of healing stage prediction – including limited sample sizes, class imbalance, and the need to model temporal dependencies inherent in longitudinal healing data – we develop a Depthwise Separable Convolution-enhanced Long Short-Term Memory (DSConv-xLSTM) Network. This innovative module leverages the computational efficiency and spatial feature extraction power of Depthwise Separable Convolutions, coupled with the temporal modeling strength of LSTM units. Consequently, the DSConv-LSTM effectively integrates critical spatial characteristics extracted from the input scans with long-range sequential dependencies present within healing timelines, enabling robust spatio-temporal feature analysis within a unified framework. Comprehensive experimental evaluation, utilizing a substantial dataset of clinical rib fracture CT scans with expert annotations for type and healing stage, validates the efficacy and stability of our proposed multi-task model. For fracture type classification, the model demonstrates outstanding performance, achieving both classification accuracy and recall rates exceeding 0.95. Notably, in the challenging task of predicting distinct rib fracture healing stages (early, intermediate, late), the model also exhibits high accuracy and robustness, achieving prediction accuracy and recall rates consistently above 0.91 across the different healing phases. These compelling results underscore the model’s strong capability to overcome the data limitations (small samples, imbalance) and the complex nature of spatio-temporal feature learning required for accurate healing stage assessment. In conclusion, this research successfully tackles critical bottlenecks in intelligent rib fracture analysis by introducing a sophisticated multi-task learning architecture. By effectively fusing multi-scale spatial features for nuanced type classification and integrating spatio-temporal information for longitudinal healing prediction via the DSConv-LSTM mechanism, the proposed framework delivers state-of-the-art performance on both tasks. Its demonstrated accuracy, recall, and stability highlight significant potential for enhancing diagnostic precision and supporting clinical decision-making in the management of traumatic rib injuries. Future work will focus on clinical deployment and integration within hospital workflows.