Multi-class Gradient Harmonized Dice Loss with Application to Knee MR Image Segmentation

Abstract

The Dice loss function is widely used in volumetric medical image segmentation for its robustness against the imbalance between the numbers of foreground and background voxels. However, it is not able to differentiate hard examples from easy ones, which usually comprise the majority of training examples and therefore dominate the loss function. In this work, we propose a novel loss function, termed as Gradient Harmonized Dice Loss, to both address the quantity imbalance between classes and focus on hard examples in training, with further generalization to multi-class segmentation. The proposed loss function is employed in a 3D fully convolutional neural network for multiple object segmentation of MRI knee joint images and validated on both public SKI10 dataset and 637 MRI knee scans collected from local hospitals. The experimental results show that the Gradient Harmonized Dice Loss outperforms the popular loss functions, such as Dice loss and Focal loss, and achieves the state-of-the-art results on the validation data of SKI10.