Graph-of-Differences: Anatomy-Structured Difference Alignment for Medical Image Re-Identification

Researchers propose Graph-of-Differences (GoD), a novel approach to medical image re-identification that grounds patient matching in explicit anatomical structures rather than arbitrary spatial features. The method demonstrates significant accuracy improvements on fundus and chest X-ray images while providing clinically auditable explanations tied to named anatomical regions.

Graph-of-Differences addresses a critical vulnerability in medical image re-identification systems: their reliance on shortcut learning that produces non-interpretable decisions clinicians cannot validate against known anatomy. Traditional approaches extract global features from images, which can correlate with patient identity through spurious patterns rather than clinically meaningful structures. GoD restructures this problem by representing each medical image as an anatomy graph where nodes correspond to named anatomical regions, then establishing soft correspondence between matched anatomies across image pairs.

This work emerges from growing recognition that AI interpretability is non-negotiable in clinical settings. Medical institutions increasingly require explainable AI systems that clinicians can audit and understand. The paper's approach converts pixel-level heatmaps—which are notoriously unstable and difficult to validate—into structure-level evidence grounded in anatomical nomenclature. This shift from pixel spaces to semantic anatomical spaces represents a maturation in how medical AI handles the explainability-accuracy trade-off.

The empirical results demonstrate tangible improvements: +7.1 percentage points on fundus images and +3.1 points on chest X-rays over strong baselines. Critically, the method shows better generalization on zero-shot external transfers, suggesting that anatomy grounding captures more robust, transferable features than spatial tokens. This has direct implications for clinical deployment, where models trained on one institution's data must perform reliably on diverse imaging equipment and protocols.

The open-source release signals potential adoption by medical imaging researchers. Future development likely involves extending anatomical graph representations to other imaging modalities and integrating these interpretable embeddings into clinical decision support systems. The work demonstrates how domain-specific architectural constraints can simultaneously improve both model performance and auditability.

• →Graph-of-Differences anchors medical image re-identification in explicit anatomical structures rather than arbitrary spatial features, improving both accuracy and interpretability.
• →The method achieves +7.1pp improvement on fundus images and +3.1pp on chest X-rays compared to frozen-backbone baselines.
• →Anatomy-grounded explanations replace unstable pixel heatmaps with verifiable structure-level evidence that clinicians can audit directly.
• →Superior zero-shot transfer performance indicates that anatomical representations capture more generalizable, robust features across institutions.
• →Open-source code release enables broader adoption in medical imaging research and potential clinical deployment pipelines.