Hierarchical Attention for Sparse Volumetric Anomaly Detection in Subclinical Keratoconus

The detection of weak, spatially distributed anomalies in volumetric medical imaging remains challenging due to the difficulty of integrating subtle signals across non-adjacent regions. This study presents a controlled comparison of sixteen architectures spanning convolutional, hybrid, and transformer families for subclinical keratoconus detection from three-dimensional anterior segment optical coherence tomography (AS-OCT). The results demonstrate that hierarchical architectures achieve 21-23% higher sensitivity and specificity, particularly in the difficult subclinical regime, outperforming both convolutional neural networks (CNNs) and global-attention Vision Transformer (ViT) baselines. Mechanistic analyses indicate that this advantage arises from spatial scale alignment: hierarchical windowing produces effective receptive fields matched to the intermediate extent of subclinical abnormalities, avoiding the excessive locality observed in convolutional models and the diffuse integration characteristic of pure global attention. Attention-distance measurements show that subclinical cases require longer spatial integration than healthy or overtly pathological volumes, with hierarchical models exhibiting lower variance and more anatomically coherent focus. Representational similarity further indicates that hierarchical attention learns a distinct feature space that balances local structure sensitivity with flexible long-range interactions. Auxiliary age and sex prediction tasks demonstrate moderately high cross-task consistency, supporting the generalizability of these inductive principles. The findings provide design guidance for volumetric anomaly detection and highlight hierarchical attention as a principled approach for early pathological change analysis in medical imaging.

💡 Research Summary

This paper presents a comprehensive and controlled investigation into the optimal deep learning architectures for detecting sparse, subtle anomalies in 3D medical imaging, using subclinical keratoconus (SKC) detection from anterior segment optical coherence tomography (AS-OCT) as a rigorous test case. The core research question addresses the unresolved debate on inductive biases: whether strong locality (CNNs), unconstrained globality (Vision Transformers), or an intermediate hierarchical approach is best suited for integrating weak, spatially distributed signals across volumetric data with limited samples.

The study constructs a unified benchmark comparing sixteen distinct models across convolutional (e.g., 3D ResNet, 3D ConvNeXt), pure transformer (ViT), and hierarchical transformer (e.g., 3D Swin Transformer, PVT) families, implemented in both 2D (single-slice) and 3D (full-volume) configurations. Training and evaluation are performed on a dataset of 12,579 AS-OCT volumes, using continuous risk scores derived from Gaussian Mixture Modeling as labels, thereby avoiding the noise associated with categorical labels for a progressive disease.

The key empirical finding is that 3D hierarchical attention models, specifically the Swin Transformer, consistently outperform all other architectures. They achieve a 21-23% relative improvement in sensitivity and specificity over both 3D CNNs and 3D ViT baselines, with the advantage being most pronounced in the challenging subclinical regime. Furthermore, 3D models universally surpass their 2D counterparts across all architectural families, confirming the importance of cross-slice context.

Beyond mere performance metrics, the paper’s significant contribution lies in its mechanistic analysis explaining why hierarchical attention excels. Through effective receptive field analysis, the authors demonstrate that hierarchical windowing creates receptive fields matched to the “intermediate spatial extent” of subclinical abnormalities. This avoids the excessive locality of CNNs (which cannot integrate non-adjacent cues) and the diffuse, unfocused attention of global ViTs (which wastes capacity on irrelevant regions).

Additional analyses solidify this conclusion: 1) Attention-distance measurements show that subclinical cases require longer-range spatial integration than healthy or manifest disease volumes, and hierarchical models meet this need with lower variance and more anatomically coherent attention maps. 2) Representational similarity analysis (using Centered Kernel Alignment) reveals that hierarchical models learn a distinct feature space that hybridizes local structural sensitivity (a CNN strength) with flexible long-range interactions (a transformer strength). 3) Auxiliary task consistency (age and sex prediction) is moderately high for hierarchical models, supporting the generalizability of their learned representations.

In summary, the paper provides strong evidence that for sparse volumetric anomaly detection—a task characterized by weak signals of intermediate spatial scale—hierarchical attention offers a principled architectural advantage. It successfully aligns the model’s effective receptive field with the physical scale of the target pathology. The work translates empirical results into generalizable design principles, advocating for hierarchical attention as a superior inductive bias for early-disease analysis in medical imaging and other domains involving 3D anomaly detection.