Fig. 1.

Schematic diagram of the training pipeline. Lower panel: features were extracted from raw images with deep-learning and filter-based methods. Upper panel: under-sampling methods were performed to balance the number of positive (vessel) and negative (background) training classes