Which Anatomy Matters Under Limited Labels? A Data-Efficient Anatomy-Aware Benchmark for Cardiac Pathology Prediction

Technical Summary: Which Anatomy Matters Under Limited Labels?

Problem Statement
In the domain of medical imaging, particularly under conditions of limited labels and constrained computational resources, a critical question remains unresolved: are performance gains driven primarily by the complexity of the machine learning models or by the quality of the representation of clinically meaningful anatomy? While researchers often resort to increasingly complex models to improve accuracy in low-data regimes, it is unclear whether the bottleneck lies in model expressiveness or in how clinical structures are represented. This issue is particularly pertinent in cardiac imaging, where pathology is expressed through anatomically meaningful structures rather than arbitrary input variations. Furthermore, practical bottlenecks in healthcare often extend beyond model design to data preparation, annotation, and deployment infrastructure, especially in resource-constrained settings.

Methodology
To address this, the authors constructed a reproducible, low-data, anatomy-aware benchmark using the public Automated Cardiac Diagnosis Challenge (ACDC) MRI dataset. The study focuses on a balanced 5-class pathology prediction task involving dilated cardiomyopathy (DCM), hypertrophic cardiomyopathy (HCM), myocardial infarction (MINF), normal subjects (NOR), and abnormal right ventricle (RV), comprising 100 subjects (20 per class).

The core methodology involves:

1. Anatomy-Aware Feature Extraction: Instead of using raw images, the authors derive patient-level descriptors from segmentation masks of three principal anatomical structures: the Right Ventricle (RV), Myocardium (MYO), and Left Ventricle (LV). For each structure, simple shape descriptors (e.g., area, aspect ratio, circularity, elongation, compactness, radial distance) are extracted from binary masks and aggregated across slices.
2. Ablation Study: The benchmark evaluates four feature configurations: RV-only, MYO-only, LV-only, and ALL-structures (concatenation of all three).
3. Model Comparison: The study compares three lightweight model families across these representations:
   • Multinomial logistic regression (Linear baseline).
   • RBF-SVM (Nonlinear kernel baseline).
   • Random Forest (Tree-based nonlinear baseline).
4. Evaluation Protocol: Models are assessed using 5-fold stratified cross-validation with metrics including accuracy, balanced accuracy, and macro-F1. The study also investigates label efficiency by varying the fraction of available labels, tests the utility of explicit inter-phase dynamic features (delta/ratio descriptors), and performs sanity checks via label shuffling.

Key Results
The experiments yield several distinct findings:

• Representation Dominates Complexity: Under limited label settings, the choice of anatomical representation has a significantly larger impact on performance than the choice of classifier family. The variation in performance across different anatomical representations (e.g., RV-only vs. MYO-only) far exceeds the variation observed between linear, kernel, and tree-based models once the representation is fixed.
• Myocardial Dominance: Among single-structure representations, the Myocardium (MYO) provides the strongest predictive signal, substantially outperforming RV-only and LV-only features. This suggests that the pathology signal in this benchmark is not uniformly distributed but is concentrated in myocardial morphology.
• Optimal Performance: While MYO is the best single-structure source, the combined "ALL-structures" representation yields the best overall performance.
• Limited Value of Simple Dynamics: Augmenting static anatomical features with handcrafted inter-phase delta and ratio descriptors did not materially improve performance over static multi-structure representations. This suggests that for this specific benchmark, static morphology captures the dominant signal, or that the simple dynamic summaries were too compressed to capture richer spatial deformation patterns.
• Robustness and Sanity: The benchmark is neither trivial nor noise-dominated; performance scales with label fraction and remains well above chance. Sanity checks (random label permutation) resulted in performance dropping to chance levels (~0.23 balanced accuracy), confirming that gains arise from genuine anatomical signals. The pipeline also demonstrated robustness to mild segmentation perturbations (erosion/dilation).
• Comparison to End-to-End Learning: A lightweight end-to-end ResNet-18 trained on raw MRI slices performed substantially worse (Balanced Acc. ~0.41) than the anatomy-aware baselines (Balanced Acc. ~0.85–0.87), reinforcing the value of explicit anatomical representation in low-label regimes.

Significance and Claims
The paper claims that in resource-constrained healthcare settings, identifying and representing the most informative anatomy may be more critical than increasing model complexity. The authors argue that for low-data cardiac pathology prediction, the "representation choice matters more than classifier complexity."

Specifically, the work suggests:

1. Scientific Insight: The benchmark is driven less by generic whole-heart geometry and more by structure-specific morphology, with the myocardium serving as the most informative individual component for the pathologies studied.
2. Practical Implication: For low-resource medical machine learning, prioritizing the myocardium as a single anatomical target may be the most effective strategy for building simplified, compute-efficient pipelines when labels are scarce.
3. Broader Impact: The findings support a shift in perspective for medical AI design, particularly in the Global South or settings with limited annotation resources. Instead of assuming better performance requires heavier end-to-end models, practitioners may achieve more practical gains by explicitly identifying and representing the anatomical structures that carry the clinically meaningful signal.

The authors remain modest regarding limitations, noting that the study relies on a single public dataset, uses handcrafted descriptors rather than raw-image end-to-end learning, and evaluates dynamics only through simple inter-phase summaries. Future work is suggested to extend this framework to additional datasets, uncertainty-aware analysis, and richer temporal descriptors.