Missing-Modality-Robust AI Framework Predicts NSCLC Survival Using CT, Pathology, and Clinical Data

Multimodal deep learning for unresectable stage II-III NSCLC survival reaches C-index 74.42 using CT, whole-slide pathology, and clinical data — without dropping patients who lack complete modality sets. Retrospective; external validation pending.

• Retrospective study of unresectable stage II-III NSCLC; exact n not reported in source. Framework uses foundation models for CT, whole-slide histopathology images, and structured clinical variables with a missing-aware encoding strategy.
• Trimodal intermediate fusion achieved C-index 74.42, outperforming unimodal, early-fusion, and late-fusion baselines; risk scores produced statistically significant log-rank stratification of progression and metastatic risk across all modality combinations.
• Key limitation: small cohort size explicitly noted by authors; no external or prospective validation reported, limiting generalizability.

RadPigeon summaries are original and for information only. They are not clinical advice.