Segmenting instances of microvascular structures, including capillaries, arterioles, and venules. https://www.kaggle.com/competitions/hubmap-hacking-the-human-vasculature/data
Model is trained on 2D PAS-stained histology images from healthy human kidney tissue slides. U-net architecture is used for semantic segmentation. Since labelled data is limited, Pretraining with random occlusions was utilized to achieve better performance in the downstream task.
Our trained model gave IoU of 0.7 on validation set