Resource Constrained U-Net for Extraction of Retinal Vascular Trees

TL;DR

The paper tackles automated retinal vascular tree extraction from fundus images under tight compute and data constraints. It adapts the U-Net by adding Batch Normalization and Dropout, uses grayscale conversion and CLAHE, and expands the dataset via augmentation, training with a combined Dice and Binary Cross-Entropy loss. On DRIVE, it reports a Jaccard of about 0.677 and ROC-AUC of 0.893, showing competitive performance given the resource limits, though minor vessels remain challenging. The work provides a practical, fast-converging baseline for automated retinal vessel segmentation and highlights directions for improving sensitivity to fine vessels and broader generalization.

Abstract

This paper demonstrates the efficacy of a modified U-Net structure for the extraction of vascular tree masks for human fundus photographs. On limited compute resources and training data, the proposed model only slightly underperforms when compared to state of the art methods.

Figures (13)

• Figure 1: Original image-mask pairs.
• Figure 2: Mean image-mask pair
• Figure 3: Image pixel value distribution
• Figure 4: Correlation between training image pixel values
• Figure 5: Original and grayscale image.