Adaptive Weighted Loss Makes Brain Tumors Segmentation More Accurate in 3D MRI Volume

Abstract

Accurately segmenting brain tumors in Magnetic Resonance Imaging (MRI) volume can benefit the diagnosis, monitoring, and surgery planning of the disease. Manual delineation practices require anatomical knowledge, are expensive, time consuming and can be inaccurate due to human error. In the data era, machines with learning capabilities can achieve automatic brain tumor segmentation in MRI volume with promising performance on large region. However, it is not very effective for tumor segmentation on small region. Although more and more methods use pixel-level loss to guide algorithms to pay more attention to accurate segmentation of small regions, this problem still exists. In this paper, we propose an adaptive weighted loss, which can automatically adjust the proportion of loss generated by different region segmentation, thereby making small region segmentation more accurate. We added the adaptive weighted loss to a 3D MRI brain tumor segmentation network using auto-encoder regularization (3D-AE), and performed extensive validation on the MICCAI Brain Tumor Segmentation Challenge 2018 dataset (BRATS 2018). The achieved dice score is 0.769 for core tumor, 0.904 for the whole tumor and 0.887 for enhanced tumor. The overall results show better performance than the state-of-the-art in both dice score and precision on BRATS 2018.