A Nested U-Net Approach for Brain Tumour Segmentation

Abstract

With the emergence of deep learning methods for image segmentation, the potential of approaches for automatic brain tumour delineation has increased substantially. This paper presents a model which is inspired by U-Net++ for this task which makes training more efficient whilst also returning better accuracy. Our approach obtained Dice Scores of 0.90, 0.85, and 0.68 on the whole tumour, tumour core, and enhanced tumour core classes. These results were obtained on a holdout set of 68 scans from the BraTS 2019 training dataset. Our model also uses half the parameters of a popular U-Net adaptation which makes use of residual blocks, resulting in faster training. On average, our model performed 8.44% better than the latter for Dice scores for all three classes within our setup.