Efficient Brain Tumor Segmentation with Dilated Multi-fiber Network and Weighted Bi-directional Feature Pyramid Network

Brain tumor segmentation is critical for precise diagnosis and personalised treatment of brain cancer. Due to the recent success of deep learning, many deep learning based segmentation methods have been developed. However, most of them are computationally expensive due to complicated network architectures. Recently, multi-fiber networks were proposed to reduce the number of network parameters in U-Net based brain tumor segmentation through efficient graph convolution. However, the efficient use of multi-scale features has not been well explored between contracting and expanding paths except simple concatenation. In this paper, we propose a light-weight network where contracting and expanding paths are connected with fused multi-scale features through bi-directional feature pyramid network (BiFPN). The backbone of our proposed network has a dilated multi-fiber (DMF) structure based U-net architecture. First, conventional convolutional layers along the contracting and expanding paths are replaced with a DMF network and an MF network, respectively, to reduce the overall network size. In addition, a learnable weighted DMF network is utilized to take into account different receptive sizes effectively. Next, a weighted BiFPN is utilized to connect contracting and expanding paths, which enables more effective and efficient information flow between the two paths with multi-scale features. Note that the BiFPN block can be repeated as necessary. As a result, our proposed network is able to further reduce the network size without clearly compromising segmentation accuracy. Experimental results on the popular BraTS 2018 dataset demonstrate that our proposed light-weight architecture is able to achieve at least comparable results with the state-of-the-art methods with significantly reduced network complexity and computation time. The source code of this paper will be available at Github.