Back

Deep Radiomics in 18F-FDG-PET/CT for Survival Prediction in Multiple Myeloma: Novel Embeddings Using A Foundational Models Memory

Guinea-Perez, J.; Uribe, S.; Peluso, S.; Castellani, G.; Nanni, C.; Alvarez, F.

2025-11-06 radiology and imaging
10.1101/2025.11.04.25339482 medRxiv
Show abstract

PurposeTo test whether internal memory states from a medical founda-tional segmentation model can serve as compact, mask-aware embeddings for predicting progression-free survival (PFS) in multiple myeloma (MM) from whole-body [18F]FDG PET/CT, and how late fusion of PET, CT, and clinical data enhances prognostic performance. MethodsWe 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: channelxmemory 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 Harrells c-index (mean {+/-} SE across folds). ResultsImage-only models using the averaging downsampler achieved up to 0.659 {+/-} 0.015 c-index (PET, spine-dilated), comparable to baseline ra-diomics 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. ConclusionMask-aware memory embeddings extracted from a founda-tional segmentation model provide effective, data-efficient imaging biomark-ers for MM PFS and, when fused with routine clinical covariates, significantly improve risk stratification over clinical-only or radiomics baselines. This of-fers a practical path to prognostic modeling on small medical cohorts without feature design.

Published in Computerized Medical Imaging and Graphics · not in our set (fewer than 10 published preprints to learn from) · training set

Matching journals

The top 5 journals account for 50% of the predicted probability mass.

50% of probability mass above

"Similar papers" are the closest papers from that journal in the model's embedding space. They show what the match is built on, but the ranking comes mostly from a classifier over the whole training set, not from these examples alone.