Do Medical Vision Models Reason About Anatomy? Probing the Spatial Inductive Biases of Learned Visual Representations
Ranking
Overall
81
Content
100
Popularity
37
Observed public metrics from 1 member.
Merged summary
TL;DR - SPAR-Bench tests whether medical vision models can reason spatially about abdominal CT anatomy, finding that they mostly encode typical organ locations rather than compare structures within individual patients. This exposes a key limitation hidden by diagnostic accuracy and standard probing methods.
- Eight probes separate coordinate localization, relational reasoning, and spatial queries across multi-organ abdominal CT scans.
- Within-slice comparison tasks remained near chance across five architectures and three medical foundation models, despite scaling and finetuning.
- Apparent in-domain success vanished under zero-shot transfer, suggesting memorization of canonical anatomy rather than image-based spatial computation.
- Using full token features instead of pooled representations raised relational recovery from 0.7% to 67.8%, showing pooled probes can substantially underestimate encoded information.
Sources (1)
Do Medical Vision Models Reason About Anatomy? Probing the Spatial Inductive Biases of Learned Visual Representations
Public signals
Semantic Scholar citations 0 · Semantic Scholar influential citations 0
TL;DR - SPAR-Bench tests whether medical vision models can reason spatially about abdominal CT anatomy, finding that they mostly encode typical organ locations rather than compare structures within individual patients. This exposes a key limitation hidden by diagnostic accuracy and standard probing methods.
- Eight probes separate coordinate localization, relational reasoning, and spatial queries across multi-organ abdominal CT scans.
- Within-slice comparison tasks remained near chance across five architectures and three medical foundation models, despite scaling and finetuning.
- Apparent in-domain success vanished under zero-shot transfer, suggesting memorization of canonical anatomy rather than image-based spatial computation.
- Using full token features instead of pooled representations raised relational recovery from 0.7% to 67.8%, showing pooled probes can substantially underestimate encoded information.