DPAFNet: A Residual Dual-Path Attention-Fusion Convolutional Neural Network for Multimodal Brain Tumor Segmentation

Abstract

Brain tumors are highly hazardous, and precise automated segmentation of brain tumor subregions has great importance and research significance on the diagnosis and treatment of diseases. Rapid advances in deep learning make accurate and efficient automatic segmentation more possible, but there are challenges. In this paper, an efficient 3D segmentation model (DPAFNet) based on dual-path (DP) module and multi-scale attention fusion (MAF) module is proposed. In DPAFNet, the dual path convolution is applied to broaden the network scale and residual connection is introduced to avoid network degradation. An attention fusion module is proposed to aggregate channel level global and local information, in which feature maps of different scales are fused to obtain features that are enriched in semantic information. This makes the object information of small tumors get full attention. Furthermore, the 3D iterative dilated convolution merging (IDCM) module expands the receptive field and improves the ability of context awareness. Ablation experiments verify the optimal combination of dilation rate for the dilated convolution merging module and demonstrate the enhancement of segmentation accuracy due to the post-processing method. Comparative experiments of this study on BraTS2018, BraTS2019 and BraTS2020 are promising and provide a promising precision and Dice score compared to related work. The proposed DPAFNet achieves Dice score of 79.5%, 90.0% and 83.9% in the enhancing tumor, whole tumor and tumor core on BraTS2018, respectively. On BraTS2019, it achieves Dice score of 78.2%, 89.0% and 81.2% in the enhancing tumor, whole tumor and tumor core, respectively.