Known-Operator-Learning Networks For Retinal Vessel Segmentation
Deep learning has advanced rapidly in recent decades, but its lack of interpretability remains a real obstacle in high-stakes fields where mistakes can be costly or dangerous. Among the ways researchers try to make deep models more interpretable, this thesis focuses on known operator learning: building network architectures that explicitly mirror the logic of well-understood conventional algorithms, so the network stays interpretable by design while still being trainable on data to adapt to the task at hand.
In ophthalmology, glaucoma, macular degeneration, and diabetic retinopathy are leading causes of vision loss, and segmenting retinal vessels from eye images is a key step in diagnosing and analyzing these conditions. Fundus imaging is the most studied modality, with plenty of public annotated datasets and existing deep learning methods to build on. OCTA, by contrast, is a newer, higher-resolution imaging modality that remains comparatively under-explored, and it’s a harder segmentation problem: its en-face projections have a low signal-to-noise ratio and show thin, densely overlapping vessel structures from multiple retinal layers at once.
This dissertation pursues two lines of work. The first applies the well-known U-Net segmentation architecture to vessel segmentation on fundus images. After common ways of boosting U-Net’s performance failed to help, the thesis instead identifies which components and configurations actually matter for U-Net to work well as a black-box model. The second, following the known-operator-learning approach, introduces Frangi-Net, a convolutional neural network version of the classic Frangi filter used for enhancing vessel structures. Building on this, the thesis constructs a three-step hybrid pipeline, basic preparation, a preprocessing network, and segmentation, that stays fully interpretable while still performing well. Notably, the preprocessing network is trained within the full pipeline as a general-purpose, edge-preserving denoiser, learning from paired images and vessel maps rather than from clean/noisy image pairs, and shows strong robustness and generalization as a result. Combining test-data-specific preparation, this learned denoiser, and the classic Frangi filter, the thesis shows that interpretable, high-performing segmentation pipelines can be flexibly built across different ophthalmic imaging databases, and even across imaging modalities.