3D dense connectivity network with atrous convolutional feature pyramid for brain tumor segmentation in magnetic resonance imaging of human heads. (June 2020)

• Record Type: Journal Article
• Authors: Zhou, Zexun; He, Zhongshi; Shi, Meifeng; Du, Jinglong; Chen, Dingding
• Is Part Of: Computers in biology and medicine. Volume 121(2020)
• Publication Date: 2020-06
• Subjects: Brain tumor segmentation -- MRI -- 3D dense connectivity network -- Feature pyramid -- Deep supervision; Medicine -- Data processing -- Periodicals; Biology -- Data processing -- Periodicals; 610.285
• DOI: 10.1016/j.compbiomed.2020.103766
• Languages: English
• ISSNs: 0010-4825

Abstract

Abstract: The existing deep convolutional neural networks (DCNNs) based methods have achieved significant progress regarding automatic glioma segmentation in magnetic resonance imaging (MRI) data. However, there are two main problems affecting the performance of traditional DCNNs constructed by simply stacking convolutional layers, namely, exploding/vanishing gradients and limitations to the feature computations. To address these challenges, we propose a novel framework to automatically segment brain tumors. First, a three-dimensional (3D) dense connectivity architecture is used to build the backbone for feature reuse. Second, we design a new feature pyramid module using 3D atrous convolutional layers and add this module to the end of the backbone to fuse multiscale contexts. Finally, a 3D deep supervision mechanism is equipped with the network to promote training. On the multimodal brain tumor image segmentation benchmark (BRATS) datasets, our method achieves Dice similarity coefficient values of 0.87, 0.72, and 0.70 on the BRATS 2013 Challenge, 0.84, 0.70, and 0.61 on the BRATS 2013 LeaderBoard, 0.83, 0.70, and 0.62 on the BRATS 2015 Testing, 0.8642, 0.7738, and 0.7525 on the BRATS 2018 Validation in terms of whole tumors, tumor cores, and enhancing cores, respectively. Compared to the published state-of-the-art methods, the proposed method achieves promising accuracy and fast processing, demonstrating good potential for clinical medicine. Highlights: A densely connected convolutional neural network is developed as the backbone to learn dense features of brain tumors via MR images. A 3D atrous convolution feature pyramid is connected with the end of the backbone to learn multiscale features of brain tumors. A 3D deep supervision mechanism is further proposed to guide the optimization of the gradient flow during training.