Cardiac MRI segmentation with focal loss constrained deep residual networks

Abstract

Delineating anatomical structures for cardiac magnetic resonance imaging (CMRI) is crucial for various medical applications such as medical diagnoses, treatment, and pathological studies. CMRI segmentation, which aims to automatically and accurately segment the heart structures, is highly beneficial for cardiologists. However, it is non-trivial to perfectly segment the ventricles, especially for the heart apex slices, considering their small sizes compared to the input images. For example, the endocardium in the Sunnybrook dataset only occupies 4% of the entire image by average. During the training process, these target pixels, or other hard samples, are buried by the massive backgrounds that make the model mostly receive optimization signals from easy samples. In this paper, we propose a focal loss constrained residual network (FR-Net) to tackle the problem. In order to mitigate the fact that the gradients of the hard samples can be easily overwhelmed by the easy samples, we use a pixel-wise re-weighting strategy to balance the gradients. Furthermore, considering focal loss constraints for each pixel independently, we propose an alternative training fashion that trains the model with focal loss and dice loss alternatively. The segmentation model can not only benefit from the pixel-wise focal loss but also from the region-wise dice loss to comprehensively optimize the model. We conducted thorough experiments on the Sunnybrook dataset, CMRI dataset, right ventricle dataset, and ACDC dataset to verify the effectiveness of the proposed method.

Keywords: cardiac magnetic resonance imaging; deep learning; heart segmentation.