Mask-aware foundational-model embeddings for 18F-FDG-PET/CT prognosis in multiple myeloma

PURPOSE: F]FDG PET/CT, and how late fusion of PET, CT, and clinical data enhances prognostic performance. METHODS: We analyzed 227 newly diagnosed MM patients with PET/CT and clinical data. For two regions of interest (spine-dilated and full skeleton), we prompted MedSAM2 slice-wise using mask-derived bounding boxes and cached the final spatio-temporal memory tensor per modality. We compared two downsampling strategy to obtain per-study embeddings: channel×memory averaging with a small CNN head, and depth-attention pooling. PET and CT embeddings were combined by late fusion and passed to a DeepSurv head. We evaluated image-only and multimodal (image+clinical) models with stratified 5-fold cross-validation. The primary endpoint was Harrell's c-index (mean ± SE across folds). RESULTS: Image-only models using the averaging downsampler achieved up to 0.659±0.015 c-index (PET, spine-dilated), comparable to baseline radiomics results. Multimodal models improved discrimination to 0.710±0.032 (CT, spine-dilated), with similar performance for other PET/CT+clinical variants (0.703-0.710), improving clinical-only baselines ∼6.5%. Averaging consistently outperformed depth-attention; concatenation and gated fusion performed comparably. PET outperformed CT within the same mask in image-only settings. CONCLUSION: Mask-aware memory embeddings extracted from a foundational segmentation model provide effective, data-efficient imaging biomarkers for MM PFS and, when fused with routine clinical covariates, significantly improve risk stratification over clinical-only or radiomics baselines. This offers a practical path to prognostic modeling on small medical cohorts without feature design.