Minimizing Hybrid Dice Loss for Highly Imbalanced 3D Neuroimage Segmentation

Recent advances in medical image segmentation have largely been driven by the success of deep learning algorithms. However, one main challenge for the training of one- stage segmentation networks is the serious imbalance between the number of examples that are easy and hard to classify or in positive and negative classes. In this paper, we first investigate and compare the strategies that were proposed parallelly to handle one or two of these imbalance problems. And we propose a hybrid loss that addresses these two imbalance problems together by combining the merits of Exponential logarithmic Dice and weighted Cross entropy Loss (EDCL). Without any whistles and bells, the proposed EDC loss with 3D Unet achieves mean dice of 57.38%, which surpasses the other state-of-the- art methods with 5-fold cross-validation on a public dataset for 3D brain lesion segmentation, Anatomical Tracings of Lesions After Stroke (ATLAS) v1.2.