Difficulty-Aware Brain Lesion Segmentation from MRI Scans

摘要

The automatic segmentation of lesions from brain MR images is critical in diagnosing and treating diseases of the brain. Compared with laborious and time-consuming manual segmentation, computer-aided segmentation provides more efficient and reliable predictions. Recently, Deep Convolutional Neural Networks were proposed to show state-of-the-art performance. Training a Deep Convolutional Neural Network demands a large amount of labeled data from experts. However, limited by poor spatial resolution, low contrast, etc., identifying lesion boundaries is difficult with ambiguity, such difficult-to-segment regions pose difficulty and variation in training a segmentation model. In this paper, we present a novel teacher–student framework. A teacher model based on Bayesian Neural Network is used to identify these regions and quantify the degree of difficulty, and a dynamically weighted training loss is applied on the student model according to such difficulty. The experimental results on the BRATS 15 dataset and SPES 2015 dataset demonstrate the state-of-the-art performance of our method.