Known-Operator-Learning Networks For Retinal Vessel Segmentation

Deep learning methods have been developing successfully in recent decades. Despite their excellent performance in various fields, deep models have been criticized for their weak interpretability. Their lack of explainability hinders their application in crucial areas where errors could be costly or even fatal. Among the different approaches to improving the interpretability of deep models, we choose to research the direction of known operator learning. Known operator learning is the most explicit and direct way to incorporate well-understood analytical prior knowledge into a deep network. The constructed network architectures follow the functionality of conventional algorithms and accommodate inherent interpretability. Building on the foundation of their classic algorithm counterparts, these networks are data-driven, allowing them to adapt more effectively to the specific requirements of various applications. In ophthalmology, glaucoma, macular degeneration, and diabetic retinopathy are the leading causes of vision loss and blindness. Segmenting retinal vessels from ophthalmic photographs is important for the analysis and diagnosis of these retina-related diseases. Among different imaging modalities, fundus imaging is the most extensively studied, as there are many public fundus image databases with annotations, and numerous methods, including deep networks, have been applied to fundus images. In contrast, OCTA is a novel imaging modality that remains under-studied despite its high resolution and diagnostic potential. Comparatively, the OCTA en face projection has a low signal-to-noise ratio and presents thin, dense vascular structures from multiple overlapping plexuses. As a result, the task of retinal vessel segmentation is more complicated and challenging for OCTA than for fundus data. This dissertation explores two directions. Firstly, the state-of-the-art medical image segmentation model U-Net is applied to the task of retinal vessel segmentation from fundus images. After unsuccessful attempts to boost the model's performance using common additive network variants, we focus on identifying the necessary components and extreme configurations for the U-Net to function effectively as a black box. Secondly, in line with known operator learning, we design the Frangi-Net, a convolutional neural network translation of the classic vessel-enhancing Frangi filter. Based on this, we construct a hybrid pipeline for retinal vessel segmentation including three steps, namely basic preparation, network preprocessing and segmentation. The hybrid pipeline is not only fully interpretable but also achieves good performance. Note that the preprocessing network is trained within the segmentation pipeline to behave as an universal edge-preserving denoiser, using image and vessel pairs rather than clean/noise-corrupted image pairs. The preprocessing network demonstrates strong robustness and good generalization capability. By combining a preparation procedure tailored to test data, the preprocessing denoiser and the classic Frangi filter, interpretable and high-performing segmentation pipelines can be flexibly constructed across various ophthalmic imaging databases, even across modalities.