Transformer-Based Multimodal Fusion Integrating Computed Tomography, Pathology, and Clinical Features for Oral Squamous Cell Carcinoma Recurrence Prediction

Background Postoperative recurrence is the leading cause of treatment failure in oral squamous cell carcinoma (OSCC), yet current risk-stratification schemes offer only modest predictive performance. We developed a Transformer-based multimodal fusion framework integrating computed tomography (CT), histopathology, and clinical variables for individualized recurrence prediction. Methods A total of 134 patients with OSCC were retrospectively enrolled. Using a dual-stream strategy, deep semantic representations were derived from preoperative contrast-enhanced CT and haematoxylin and eosin-stained histopathological images via MnasNet-0.5, while quantitative radiomic descriptors were extracted using PyRadiomics. Together with structured clinical variables, these constituted five parallel feature streams, which – after dimensional harmonization – were passed through a three-layer Transformer encoder employing self-attention and learnable modality weights for nonlinear cross-modal integration. Performance was assessed by fivefold stratified cross-validation. Results The full-modality TransformerFusion model achieved an AUC of 0.909 ± 0.069, with a sensitivity of 0.791, specificity of 0.834, and precision of 0.854. Decision curve analysis demonstrated superior net clinical benefit over logistic regression and the treat-none strategy across threshold probabilities of 0.20 to 0.70. Ablation experiments revealed that the pathology dual-stream, CT dual-stream, and clinical-only submodels attained AUCs of 0.840, 0.751, and 0.773, respectively – each inferior to the integrated model. Self-attention-based fusion (AUC = 0.909) also outperformed feature concatenation (AUC = 0.854). Conclusion The proposed framework captures complementary prognostic signals across CT, histopathology, and clinical variables, outperforming single-modality counterparts in recurrence prediction. As it relies solely on routine preoperative CT and standard haematoxylin and eosin slides, it offers a clinically deployable tool for postoperative risk stratification in OSCC.