Cross-Domain Retinopathy Classification Based on Optical Coherence Tomography Sensors via Domain Adversarial Graph Convolutional Network

Optical coherence tomography (OCT) sensor system, known for their high precision and noninvasive imaging capabilities, can capture detailed structural information of retinal tissue. Analyzing OCT sensing images allows for the identification of various retinopathies, providing significant references for clinical diagnosis and treatment. In a typical OCT sensing system, a sample beam illuminates the retina to capture reflected interference signals, which are converted to electronic form. The interference pattern is processed using the Michelson interferometry and diffraction gratings, and the resulting spectral data are transformed into a B-scan image through the Fourier transform, revealing detailed cross-sectional and 3-D organization. The application of deep learning techniques has become a prevalent methodology for the analysis of OCT sensing images. However, given the varying light source wavelengths, detector sensitivities, and imaging modalities of OCT sensing devices, images generated for the same patient at different medical centers may show significant discrepancies. These differences can affect image quality and resolution, as well as the representation of specific pathological features. This can result in distributions that are incongruent between the training and test datasets, thereby impeding the efficacy of deep learning models. Existing studies have focused mainly on domain and class labels, overlooking the importance of data structure and limiting their effectiveness. To address this issue, we propose a novel approach, namely, domain adversarial graph convolutional network (DAGCN), introducing two innovative modules to enhance the adaptability of OCT sensing image classification models. The first module, data structure-aware alignment, leverages small batch OCT image data to reduce domain shift. The second, class centroid alignment (CA), minimizes the distance between class centroids across domains, improving generalization and adaptability. We validated DAGCN’s performance through extensive testing on three public datasets. The results show that DAGCN enhances the adaptability of OCT retinopathy image classification techniques by effectively learning domain invariants and semantic representations that mitigates the impact of feature fluctuations resulting from discrepancies in sensor sensitivity and noise levels across OCT sensing devices.