Self-Guided Adversarial Network for Domain Adaptive Retinal Layer Segmentation

The precise segmentation and measurements of the retina and choroid layers from optical coherence tomography (OCT) imaging are crucial tasks for the downstream analysis of eye-related diseases. However, low contrast across different tissue layers and domain gaps between various OCT imaging instruments pose challenges to reliable segmentation. To address this issue, we propose a novel and robust domain adaptation learning method for OCT layer segmentation that bridges the domain gaps between images acquired from the most commonly used OCT instruments in a clinic. Specifically, we introduce a selective guided adversarial adaptation strategy that generalizes the learned parameters from different image domains. Our approach utilizes dual encoders to ensure that the encoding paths of target and source images remain separate. This allows the parameters to be transmitted based solely on the weight and the gradient during information propagation, which further reduces domain discrepancy within an adversarial framework. We validate our proposed method using various domain OCT data acquired from multiple instruments and demonstrate that it outperforms other advanced methods in both retinal and choroidal layer segmentations. Furthermore, we provide comprehensive pioneer correlation measurements and analysis based on the obtained segmentations, effectively illustrating the associations between layer thickness measurements and neurodegenerative diseases, such as white matter hyperintensities (WMHs) and Parkinson’s disease (PD).