Benchmarking Multimodal Medical Data Analysis for Cancer-Related Disease Diagnosis

Abstract

Multimodal medical data provide complementary information for cancer diagnosis and prognosis; however, effectively integrating heterogeneous modalities remains challenging because real-world clinical datasets frequently contain incomplete modality observations. We present a unified benchmarking framework for multimodal cancer-related diagnostic tasks that (i) learns modality-specific embeddings (WSI, CT, MRI, Clinical), (ii) trains a learnable reconstruction module to impute missing embeddings, and (iii) evaluates multiple fusion strategies (early and late fusion with learnable weighting). We validate the approach on two independent cohorts (MMIST-ccRCC — clear cell renal cell carcinoma dataset; Lung-Immunotherapy — lung cancer dataset) using patient-level stratified five-fold cross-validation and extensive ablation across 44+ configuration variants. The framework consistently improves diagnostic performance over unimodal baselines and is robust to missing modalities (on MMIST-ccRCC achieving a BACC of 0.9144 ± 0.0476 and Lung-Immunotherapy obtaining a BACC of 0.8462 ± 0.0717). Hence, this study provides a reproducible evaluation protocol and practical design guidelines for multimodal cancer-related diagnostic tasks. Code is available here https://github.com/CherryNhu/Benchmarking-Multimodal-Medical-Data-Analysis-for-Cancer-Related-Disease-Diagnosis