LHU-VT: A Lightweight Hypercomplex U-Net with Vessel Thickness-Guided Dice Loss for retinal vessel segmentation

Waqar Ahmed, Panos Liatsis

Abstract: Highlights•We propose a lightweight Hypercomplex U-Net designed for retinal vessel segmentation.•Hypercomplex number system to extract complex inter-channel relationships within data.•Pixel-wise vessel thickness-guided weighted dice loss to tackle the class imbalance.•Achieved reductions of 2.4×<math><mo is="true">×</mo></math> in FLOPs, 4.4×<math><mo is="true">×</mo></math> in parameters, and 2.6×<math><mo is="true">×</mo></math> in model size.•Achieved AUCs of 0.9938, 0.9879, 0.9988, and 0.9808 on CHASE, DRIVE, STARE, and HRF.