Abstract
Few-shot medical anomaly detection aims to identify/localize pathological regions using only a few annotated examples, reflecting real clinical scenarios where expert annotations are scarce and expensive. When data is limited, fine-tuning large vision models can lead to over fitting, whereas purely visual adaptation strategies may overlook sub tle pathological deviations from normal anatomical patterns. To address this challenge, we propose a vision–language framework that explicitly models abnormality as a semantic deviation from normality. Built upon a frozen CLIP backbone, proposed framework learns medical textual prompts describing normal and abnormal states and constructs a dif ferential semantic representation that isolates pathology-relevant infor mation rather than treating disease as an absolute concept. To align vi sual features with this semantic deviation, we introduce text-gated visual adapters inserted at multiple transformer layers, allowing textual devia tion cues to dynamically modulate intermediate visual features and sup press dominant normal anatomical patterns. We evaluate the proposed framework on six medical anomaly detection benchmarks spanning five imaging modalities under multiple shot settings. Our method achieves the best average performance in most shot settings and remains compet itive across datasets, with particularly strong gains in low-shot regimes.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/AMAI_028.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
Open Review Page: https://openreview.net/profile?id=%7EArooba_Maqsood1
BibTex
@InProceedings{ShaSho_Modeling_MICCAISAT2026,
author = { Shafi, Shoiab AND Maqsood, Arooba AND Tariq, Abdullah AND Gilani, Syed Zulqarnain},
title = { { Modeling Abnormality as Semantic Deviation for Few-Shot Medical Anomaly Detection } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17273},
month = {pending},
page = {pending}
}
