EfficientNet for retinal blood vessel segmentation

Abstract

Automated techniques for retinal vessel segmentation is an active research area for the past three decades. Features associated with retinal blood vessels like morphology, area, diameter, tortuosity are important to assess the onset and progression of many eye-related and cardiovascular diseases. For retinal vessel segmentation, we propose two deep neural networks: U-net with EfficientNet as the backbone and EfficientNet encoder with LinkNet decoder. Gamma adjustment and contrast limited histogram equalization is the pre-processing stages adopted. EfficientNetB3 with U-net provide significant improvement. Results are evaluated on benchmark fundus image datasets like DRIVE [1], STARE [2], HRF [3], and CHASE_DB1 [4]. The proposed architecture obtained 96.35% accuracy, 86.35% sensitivity, 97.67% specificity, and an F1 score of 0.8465 on the DRIVE dataset.