Anatomy-Slot: Unsupervised Anatomical Factorization for Homologous Bilateral Reasoning in Retinal Diagnosis

Technical Summary: Anatomy-Slot for Bilateral Retinal Diagnosis

Problem Statement
Retinal diagnosis is inherently a bilateral clinical task, where clinicians compare homologous structures (e.g., optic disc asymmetry) across left and right eyes to detect disease. However, most existing deep learning models operate on monocular representations, treating each eye independently. The authors hypothesize that texture-based monocular classification is insufficient for bilateral disease detection and that explicit structural comparison is necessary for clinically aligned performance. While bilateral modeling and object-centric slot decomposition exist independently in the literature, their integration as explicit correspondence units for homologous anatomical reasoning has not been empirically tested under statistically controlled conditions.

Methodology: Anatomy-Slot
The proposed Anatomy-Slot framework operationalizes the hypothesis of explicit cross-eye correspondence through an unsupervised anatomical bottleneck. The pipeline consists of the following components:

1. Backbone and Tokenization: Given a pair of bilateral fundus images $(x_L, x_R)$, the model utilizes a shared RETFound ViT-L backbone to extract patch tokens.
2. Unsupervised Anatomical Factorization (Slot Attention): For each eye, patch tokens are decomposed into $K$ object-centric slots using Slot Attention (with $I=3$ iterations). This process initializes slots as learnable parameters and updates them via a competitive attention mechanism (softmax over slots, renormalized over tokens) and a GRU, yielding a set of slot representations $S \in \mathbb{R}^{K \times D}$ per eye.
3. Bidirectional Cross-Eye Correspondence: To align homologous structures, the model employs bidirectional cross-attention. The left eye's slots query the right eye's slots (and vice versa) using multi-head cross-attention ($H=8$). This allows the model to explicitly align slots across eyes, creating a correspondence-aware representation.
4. Dual Objective Learning: The model is trained with a combined objective:
   • Classification Loss ($L_{cls}$): Standard supervised loss for disease diagnosis.
   • Reconstruction Loss ($L_{recon}$): An unsupervised auxiliary task where each slot predicts per-patch RGB values and mask logits. A lightweight decoder reconstructs a low-resolution RGB image. This loss stabilizes slot learning and prevents slot collapse.
   • Total Objective: $L = L_{cls} + 0.5 L_{recon}$.
5. Training Protocol: The slot module undergoes unsupervised pretraining on a pool of ~20k unlabeled retinal images. During supervised fine-tuning on the target dataset, only the last ViT block and layer normalization parameters are unfrozen, while the slot iterations are reduced to 3.

Key Contributions

• Architecture: A novel bilateral slot-based architecture where slots serve as correspondence units, and cross-attention facilitates explicit cross-eye information exchange.
• Statistical Rigor: A robust evaluation protocol involving $n=10$ independent seeds, 95% confidence intervals, and non-parametric paired testing (Wilcoxon signed-rank test) to ensure statistical significance.
• Mechanistic Analysis: Comprehensive tests probing the model's reliance on correspondence, including pairing disruption experiments, stress testing under Gaussian noise, and quantitative analysis of optic disc grounding.

Results
Evaluated on the ODIR-5K dataset (bilateral multi-label diagnosis), the method demonstrates significant improvements over matched baselines:

• Performance: Anatomy-Slot achieves a macro AUC of 0.865, a 4.2% improvement over the matched ViT-L baseline (0.823). The improvement is statistically significant ($W=0, p=0.002$).
• Ablation Studies:
  • Bilateral processing without slots yields only a modest gain (0.828 AUC).
  • Slot decomposition without cross-eye interaction underperforms (0.754 AUC).
  • Removing the reconstruction loss causes slot collapse, degrading performance to 0.729 AUC.
  • The full model's performance peaks at $K=8$ slots; fewer slots under-represent anatomy, while more slots dilute correspondence.
• Correspondence Dependence: When left-right pairing is disrupted (shuffled during evaluation), Anatomy-Slot's performance drops by 18.3 points (0.865 to 0.682), whereas the baseline drops by only 3.2 points. This supports the hypothesis that the model relies on homologous structural matching.
• Robustness: Under Gaussian noise ($\sigma=0.2$), Anatomy-Slot maintains an AUC of 0.685, outperforming both the baseline (0.578) and the "No Slots" variant (0.595).
• Grounding: On the REFUGE dataset (zero-shot transfer), the model achieves a mean Intersection over Union (mIoU) of 0.635 for optic disc localization using Slot 1, significantly outperforming a K-Means feature clustering baseline (0.452).
• Separability: The Fisher Ratio in the penultimate representation space increases from 0.85 (baseline) to 1.62 (Anatomy-Slot), indicating improved class separability.

Significance and Claims
The paper claims that explicit modeling of cross-eye correspondence via unsupervised anatomical factorization significantly improves diagnostic performance in retinal disease detection. The results suggest that the performance gain arises specifically from the combination of structured object-centric slots and cross-eye alignment, rather than from bilateral processing or slot decomposition alone.

The authors position this work as a step toward "clinically aligned retinal AI," demonstrating that models can learn to rely on homologous anatomical matching in an unsupervised manner. The mechanistic analyses (pairing disruption, attention localization) provide evidence that the model's behavior is consistent with the clinical reasoning process of comparing homologous structures. The authors acknowledge limitations, including evaluation primarily on a single dataset and the need for further testing on clinical misalignment, but conclude that the approach offers a path toward interpretable diagnostic support systems.