Fig. 2: A vascular structure segmentation problem.
From: Rethinking deep learning in bioimaging through a data centric lens
All grayscale images or patches are displayed after Auto-Contrast in ImageJ. a Various image samples showing high diversity in the dataset. The size of each image is 7 × 1024 × 1024 (ZYX). b 25 representative patches (i.e., a core set) are selected from 90,860 patches to build model M0, 6 of which are shown here. Manual annotations are marked in green. c The top row contains image patches with high uncertainty. The bottom row shows the corresponding predicted segmentation. Three patches (within the red boxes) are manually selected from these image patches as the critical set to fine-tune model M1. d Segmentation results of model M0 (blue boxes) and model M1 (red boxes). Fine-tuning with the critical set significantly reduced model hallucinations. e Segmentation results evaluated using reverse classification accuracy (RCA). Two random examples are shown, one with relatively high RCA (right) and the other with relatively low RCA (left). Visual inspection can confirm that a higher RCA corresponds to a better segmentation result.
