Physics-Informed Machine Learning in Vessel Segmentation

Abstract

Retinal vessel segmentation is essential for identifying eye diseases, but existing methods struggle with image variations, noise, and the demand for large amounts of labeled data. This study presents a new multi-step, physics-informed deep learning framework to overcome these challenges. The approach begins with two specialized Physics-Informed Neural Networks (PINNs) for feature extraction: an EdgeNet module, guided by a diffusion-reaction partial differential equation (PDE) with adjustable parameters and flexible operators, to clearly outline vessel edges and a C8-equivariant texture PINN, shaped by Perona-Malik anisotropic diffusion and statistical goals, to capture texture features that remain consistent despite rotations. The latent representations of these two models have rich representations of retinal features, which are carefully combined and used as an enhanced input for an advanced UTNet model, which use convolutional feature extraction with Transformer-based global context understanding for the final segmentation. Thorough testing on the DRIVE dataset, especially when using a Vessel Thickness Guided Dice Loss (VTDL), proved the framework’s effectiveness, achieving an AUC-ROC of 0.9768 and a strong recall of 0.8801 when using vessel thickness loss and 0.8547 recall when using combined BCE and Dice loss, demonstrating excellent ability in detecingt small and delicate blood vessels. This work shows that embedding physical knowledge through PINNs for targeted feature extraction and its combination greatly improves the accuracy, reliability, and detailed capture of deep learning models for retinal vessel segmentation, providing a structured step forward in automated medical image analysis.

Date of Award	2025
Original language	American English
Supervisor	Panagiotis Liatsis (Supervisor)