Abstract
Patient‑level classification of clinically significant prostate cancer (csPCa) using biparametric MRI (bpMRI) is still challenging owing to heterogeneous MRI visual characteristics and limited lesion annotations. This study presents a lesion‑annotation‑free framework that adapts a frozen MedSigLIP vision encoder for T2‑weighted, DWI, and ADC inputs via shared/private LoRA; dynamic slice masking handles examinations with variable depth, whereas shared and sequence‑specific adapters capture general and complementary representations. The resulting embeddings are fused with clinical variables, and an orthogonality loss is applied to mitigate feature redundancy. Trained on a retrospective cohort (N = 6458), the model yields AUC/MCC of 0.849/0.505 on the held‑out retrospective test set (n = 969) and 0.830/0.449 on an independent prospective cohort (N = 2540), surpassing conventional LoRA and shared‑encoder baselines. These findings validate that shared/private LoRA enables scalable patient‑level csPCa classification without lesion masks.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/MWM_120.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
Open Review Page: Not Available
BibTex
@InProceedings{DimAvt_SequenceSpecific_MICCAISAT2026,
author = { Dimitriadis, Avtantil AND Kalliatakis, Grigorios AND Osuala, Richard AND Kessler, Dimitri AND Diaz, Oliver AND Lekadir, Karim AND Fotiadis, Dimitrios I. AND Tsiknakis, Manolis AND Papanikolaou, Nikolaos AND Marias, Kostas},
title = { { Sequence-Specific LoRA Adaptation of a Vision Foundation Model for Patient-Level csPCa Classification from Biparametric MRI } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17262},
month = {pending},
page = {pending}
}