Integrating multimodal clinical data with a large model for prostate cancer diagnosis

Abstract

Accurate prostate cancer (PCa) diagnosis remains difficult because of tumor heterogeneity and the challenge of integrating multimodal clinical information. We developed Prost-LM, a multimodal large language model that jointly embeds MRI-derived features, numerical PSA values, and free-text clinical reports into a unified semantic space to enable deep cross-modal reasoning. Trained and validated on a large multi-center cohort of 3940 patients, Prost-LM achieved strong diagnostic performance, with an internal validation AUC of 0.954 for distinguishing PCa from benign conditions, outperforming MRI-only models (AUC = 0.868, P < 0.001). For detecting clinically significant PCa (Gleason score ≥ 7), Prost-LM reached an AUC of 0.955. Additionally, the model provides interpretable diagnostic decisions to support clinical verification. These results suggest Prost-LM can improve automated PCa diagnosis and support precision oncology through multimodal AI.

Ähnliche Arbeiten