From Whole to Parts: Medical Imaging Semantic Segmentation with Very Imbalanced Data

Label imbalance is an unavoidable issue for deep learning in the field of medical imaging. In terms of semantic segmentation, diseased tissues occur infrequently with very limited volumes, which exerts a strong bias to the learning of convolutional neural networks (CNNs). Dice loss functions are often applied for training to alleviate the problem, however they become very unstable when the dataset is extremely imbalanced. In this work, we propose a novel method that first locates the abnormal tissues, and performs segmentation in local regions. We also propose the Ising conditional random fields (CRFs) for post- processing, and apply the Digital Annealer (DA) for optimization. Our experiments demonstrate that the proposed approach is effective in learning with very imbalanced data.