RSCO-Net: A refined single candidate optimizer for multimodal colorectal cancer postoperative recurrence prediction

Postoperative recurrence remains a critical challenge in the clinical management of colorectal cancer (CRC), affecting approximately 30%–50% of patients who undergo curative-intent surgery. Precise risk stratification is essential for personalized treatment and improved prognosis, yet traditional markers often fall short in predictive accuracy. This study introduces RSCO-Net , an integrated multimodal framework that leverages a Refined Single Candidate Optimizer (RSCO) for enhanced CRC recurrence prediction. Our approach uniquely combines 512-dimensional visual features from whole-slide images (WSIs) via attention-based multiple-instance learning (MIL) with 512-dimensional semantic embeddings from pathology reports via transformer-based language models. To navigate the high-dimensional, non-convex optimization landscape of the fused 1024-dimensional feature space, we propose the RSCO metaheuristic. RSCO employs a two-phase search strategy that balances global exploration with local exploitation, coupled with an escape mechanism to avoid stagnation. Experimental evaluations on the TCGA-COAD/READ dataset demonstrate that RSCO-Net achieves a state-of-the-art AUROC of 0.839 and an AUPRC of 0.731, significantly outperforming benchmark population-based optimizers and unimodal baselines. Ablation studies confirm that RSCO provides a 2.1% gain over simple candidate optimization, while multimodal fusion contributes a 9.1% improvement over WSI-only models. Furthermore, we provide clinical interpretability through SHAP-based feature importance and attention-driven WSI visualization. Our findings establish RSCO-Net as a robust and computationally efficient tool for precision oncology, advancing the frontier of multimodal medical AI for CRC prognosis.