New Benchmark Reveals Anatomy Over Model Complexity for Cardiac AI

Himanshu Singh's paper, accepted at the ICML 2026 Workshop GlobalSouthML (Seoul, South Korea; PMLR 306, 2026), tackles a fundamental question in medical imaging: when labeled data and compute are limited, what drives performance—more expressive models or better representation of clinically meaningful anatomy? The benchmark focuses on 5-class cardiac pathology prediction using the public ACDC MRI dataset. It extracts segmentation-derived patient descriptors from three cardiac structures: right ventricle, myocardium, and left ventricle. These descriptors are then compared across anatomy-specific and multi-structure representations using three classifier families: linear, kernel, and tree-based. The study explicitly controls for model complexity to isolate the effect of anatomical representation.

The results are clear: under limited label settings, representation dominates complexity. Models using the most informative anatomical regions consistently outperform more complex models that use generic or poorly chosen features. This finding has direct implications for AI deployment in resource-constrained healthcare settings, where large labeled datasets and compute are often unavailable. Rather than chasing ever-larger models, developers should focus on identifying and representing the most diagnostically relevant anatomy. The full paper is available on arXiv (2606.06509) and includes code and data links.