Deep Learning based Intraretinal Layer Segmentation using Cascaded Compressed U-Net

AO_SCPLOWBSTRACTC_SCPLOWIntraretinal layer segmentation on macular optical coherence tomography (OCT) images generates non invasive biomarkers querying neuronal structures with near cellular resolution. While first deep learning methods have delivered promising results with high computing power demands, a reliable, power efficient and reproducible intraretinal layer segmentation is still an unmet need. We propose a cascaded two-stage network for intraretinal layer segmentation, with both networks being compressed versions of U-Net (CCU-INSEG). The first network is responsible for retinal tissue segmentation from OCT B-scans. The second network segments 8 intraretinal layers with high fidelity. By compressing U-Net, we achieve 392- and 26-time reductions in model size and parameters in the first and second network, respectively. Still, our method delivers almost similar accuracy compared to U-Net without additional constraints of computation and memory resources. At the post-processing stage, we introduce Laplacian-based outlier detection with layer surface hole filling by adaptive non-linear interpolation. We trained our method using 17,458 B-scans from patients with autoimmune optic neuropathies, i.e. multiple sclerosis, and healthy controls. Voxel-wise comparison against manual segmentation produces a mean absolute error of 2.3m, which is 2.5x better than the devices own segmentation. Voxel-wise comparison against external multicenter data leads to a mean absolute error of 2.6m for glaucoma data using the same gold standard segmentation approach, and 3.7m mean absolute error compared against an externally segmented reference data set. In 20 macular volume scans from patients with severe disease, 3.5% of B-scan segmentation results were rejected by an experienced grader, whereas this was the case in 41.4% of B-scans segmented with a graph-based reference method.