TSNet:A Multi-modal Deep Learning Framework for Subtyping Appendicitis: Integrating Ultrasound Images, Handcrafted Features, and Clinical Data

Purpose Preoperative differentiation of Acute Appendicitis (AA), Chronic Appendicitis (CA), and the clinically challenging Acute Exacerbation of Chronic Appendicitis (AEC) remains a significant diagnostic dilemma. We developed a multi-modal deep learning framework integrating ultrasound (US) images with domain-specific features to enhance diagnostic precision. Methods This retrospective study included 605 pathology-confirmed patients (392 AA, 150 CA, 63 AEC). The dataset comprised preoperative US images and a 19-dimensional feature vector encompassing clinical metrics and handcrafted sonographic markers guided by clinical protocols. We developed a multi-modal Two-Stream fusion framework ( TSNet ) utilizing a ResNet-50 backbone with Spatial Attention Modules (SAM) for visual extraction, fused with tabular data. Class imbalance was addressed using Focal Loss. Performance was evaluated via 5-fold stratified cross-validation. Results The optimal image-only model ( ResNet-SAM ) achieved an accuracy of 0.8612, whereas the best machine learning baseline ( RF-Sel ) attained 0.7570. The proposed TSNet achieved a robust patient-level accuracy of 0.8529. Notably, for the difficult-to-diagnose AEC subtype, TSNet achieved an AUC of 0.8031 and an F1-score of 0.4800. This represents a 17.1% improvement over the best machine learning baseline, confirming its superior capability in capturing acute-on-chronic pathology. Conclusion TSNet effectively differentiates appendicitis subtypes by leveraging the complementary strengths of deep visual representations and expert clinical knowledge. By significantly improving the identification of the elusive AEC subtype, the framework offers a robust tool for optimizing surgical decision-making and reducing the risk of mismanagement in complex clinical cases.