Anatomy-Slot: Unsupervised Deep Learning for Retinal Diagnosis via Bilateral Eye Comparison
A new deep learning approach called Anatomy-Slot has been introduced by researchers to enhance retinal diagnostics by directly comparing similar structures in both eyes, similar to how clinicians operate. This model features an unsupervised anatomical bottleneck that breaks down patch tokens into slots, aligning them through bidirectional cross-attention. When tested on the ODIR-5K dataset, Anatomy-Slot demonstrated a 4.2% increase in AUC compared to a corresponding ViT-L baseline (95% CIs; Wilcoxon signed-rank test, W=0, p=0.002). Additional experiments included pairing disruption, stress testing with Gaussian noise, quantitative optic disc grounding on REFUGE, and analyzing cross-attention localization. This research has been published on arXiv in the field of computer vision and pattern recognition.
Key facts
- Anatomy-Slot improves AUC by 4.2% over ViT-L baseline on ODIR-5K
- Uses unsupervised anatomical bottleneck with slot decomposition and cross-attention
- Tested under Gaussian noise for robustness
- Optic disc grounding evaluated on REFUGE dataset
- Published on arXiv in computer vision category
Entities
Institutions
- arXiv