arXiv · Jul 17, 2026
Spatial Normalization for Cross-Domain Retinal Layer Segmentation in Optical Coherence Tomography
A study investigates spatial normalization as a preprocessing strategy to mitigate geometric domain shifts in retinal layer segmentation from Optical Coherence Tomography (OCT). The proposed fovea-centered normalization framework aligns OCT volumes into a common anatomical reference. The research evaluates state-of-the-art deep learning architectures and combines conventional overlap-based metrics for comprehensive assessment.
What happened
Retinal layer segmentation in OCT is crucial for extracting quantitative biomarkers, especially in neurodegenerative disease research. Deep learning methods achieve high performance but struggle with robustness and generalization across heterogeneous datasets due to domain shifts. This work introduces a fovea-centered spatial normalization framework inspired by neuroimaging practices to align OCT volumes, aiming to improve segmentation consistency. A comprehensive evaluation of deep learning architectures is performed using both overlap-based and other metrics.
Technical significance
The approach adapts spatial normalization techniques from neuroimaging to OCT data, centering volumes on the fovea to reduce anatomical variability. This preprocessing step is designed to be model-agnostic and could enhance the generalization of existing segmentation architectures without modifying their core design.
Industry impact
Improving cross-domain segmentation robustness can accelerate the clinical translation of OCT-based biomarkers for neurodegenerative diseases, where diverse acquisition protocols and patient populations currently limit reliability.
What to watch
If validated across multiple public and clinical datasets, this normalization framework could become a standard preprocessing step in OCT analysis pipelines, potentially enabling more reliable multi-site studies and regulatory approval of AI-based diagnostic tools.
Decision value
Enhanced segmentation consistency may reduce the need for costly manual corrections and retraining of models for each new device or population, lowering barriers for AI-assisted OCT analysis in both research and clinical settings.