Abstract
Medical foundation models pretrained via self-supervised learning on large-scale imaging data capture transferable anatomical and pathological representations, enabling few-shot frameworks to inherit these features and shift focus toward constructing geometrically robust decision boundaries from limited labels. We introduce \textbf{PROFIT} (\textbf{PRO}totype-guided \textbf{F}ew-shot \textbf{I}nference and \textbf{T}ransfer), a parameter-efficient framework tailored for multimodal stroke diagnosis that decouples representation inheritance from decision boundary optimization. Its core component, the \textbf{ProtoSphere Head}, is a lightweight hyperspherical classifier that repurposes support-set prototypes as geometry-aware priors on the unit hypersphere and employs a learnable temperature to calibrate decision sharpness. By freezing the foundation encoder and restricting adaptation to only approximately 1K trainable parameters, PROFIT preserves pretrained knowledge while stabilizing optimization under extreme data scarcity. We evaluate PROFIT along the stroke diagnostic continuum, from infarct detection to severity grading that reflects distributed neurological impairment. On the MICCAI FOMO25 Challenge, PROFIT achieves top AUROC on infarct detection (Open Track) using only 21 labeled cases for fine-tuning. On the public Stroke Outcome Optimization Project (SOOP) dataset, it consistently outperforms ProtoNet, Linear Probe, and Baseline++ across all $K$-shot settings ($K \in {1,5,10,15,20,25}$), achieving an AUC of 0.661 at $K{=}25$—surpassing the fully supervised UniFormer (AUC of 0.651)—with low cross-seed variance, demonstrating a data-efficient pathway for deploying foundation models for annotation-scarce stroke diagnosis.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/SWITCH_004.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: https://papers.miccai.org/miccai-2026-sat/supp/SWITCH_004_supp.pdf
Link to Open Review
Open Review Page: https://openreview.net/forum?id=eOSyJ8aHcR
BibTex
@InProceedings{WeiSiq_PROFIT_MICCAISAT2026,
author = { Wei, Siqi AND Jia, Fucang},
title = { { PROFIT: Prototype-Guided Few-shot Inference and Transfer for Multimodal Stroke Diagnosis } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17255},
month = {pending},
page = {pending}
}