Multi-features refinement and aggregation for medical brain segmentation

Abstract

Due to the complexity of the anatomical structure for human organs, medical image segmentation is always a challenging computer vision task. The Convolutional Neural Network (CNN) requires a rich feature representation, which not only needs the convolutional layers from shallow to deep,but also requires the resolution from small to large. Although CNN can be used to fuse mid-level features that are employed short-cutting, this just is a simple 'shallow' connection. Thus, how to obtain useful features and how to utilize these features to improve the segmentation processes are still the key issues. In this paper, Multi-features Refinement and Aggregation (MRA) makes full use of hierarchical features by using the features fusion on several levels, and reveal the importance of refinement and aggregation of features in the medical image segmentation process. The network get low-level, high-level and even mid-level features by sampling. After aggregation and re-extraction, these features are more effectively combined. Experiment results show that our method can significantly improve segmentation accuracy compared to existing feature fusion schemes. And our approach is generalized to different backbone networks with consistent accuracy gain in brain segmentation, and it sets a new state-of-the-art in the Brat-2015 benchmarks.