Segmentation of Retinal Low-Cost Optical Coherence Tomography Images using Deep Learning

A multi-institutional research team has published a paper demonstrating a deep learning system designed to segment retinal images from a novel, low-cost optical coherence tomography (OCT) device intended for at-home patient monitoring. The system targets age-related macular degeneration (AMD), a leading cause of vision loss whose treatment requires frequent monitoring of disease-specific biomarkers visible in OCT scans. The core of their method is a convolutional neural network (CNN) trained to identify and segment the total retina and pathological structures like pigment epithelial detachments (PEDs).

While the CNN achieved high accuracy for overall retinal segmentation, the team found segmenting the more complex PED biomarkers to be challenging due to image artifacts. To address this, they implemented a second-stage refinement using a convolutional denoising autoencoder (CDAE). This model was trained on clean retinal shape data and could correct errors in the initial CNN predictions, effectively cleaning up noisy or artifact-laden images from the portable SELF-OCT device. The work, accepted for SPIE Medical Imaging 2020, represents a critical step toward enabling effective computer-aided diagnosis in home-monitoring scenarios.

The technical approach highlights a practical solution for deploying AI in resource-constrained environments. By combining a primary segmentation network with a shape-prior refinement model, the system becomes more robust to the lower image quality expected from cost-effective, patient-operated devices. This pipeline is essential for translating frequent clinical monitoring into a viable home-care protocol, which could significantly improve treatment outcomes by allowing for earlier detection of disease progression.