Learning Melanocytic Proliferation Segmentation in Histopathology Images From Imperfect Annotations

Kechun Liu, Mojgan Mokhtari, Beibin Li, Shima Nofallah, Caitlin May, Oliver Chang, Stevan Knezevich, Joann Elmore, Linda Shapiro; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2021, pp. 3766-3775

Melanoma is the third most common type of skin cancer and is responsible for the most skin cancer deaths. A diagnosis of melanoma is made by the visual interpretation of tissue sections by a pathologist, a challenging task given the complexity and breadth of melanocytic lesions and the subjective nature of biopsy interpretation. We leverage advances in computer vision to aid melanoma diagnosis by segmenting potential regions of lesions on digital images of whole slide skin biopsies. In this study, we demonstrate a Mask-R-CNN-based segmentation framework for such a purpose. To alleviate the cost of data annotation, we leverage a sparse annotation pipeline. Our model can be trained on sparse and noisy labels and achieves state-of-the-art performance in identifying melanocytic proliferations, producing a segmentation with Dice score 0.719, mIOU 0.740 and overall pixel accuracy 0.927.