Abstract: In view of the weak ability of multi-scale lesion processing ability of the traditional deep convolutional neural networks for fully automatic brain tumor segmentation, the improved recurrent residual convolutional units are used to build the backbone for feature learning to improve the spatial relevance of feature learning and alleviate the network degradation and gradient dispersion caused by too complex network model. The hierarchical feature pyramid is constructed by 3D atrous-convolution with different expansion rates and cross model attention mechanism, combined with context features to improve the recognition ability of the overall model for tumors of different sizes. Combined with multi-layer feature map, the tumor is predicted to obtain the final segmentation result. Abundant ablation experiments carried on BraTS 2019 datasets demonstrate that the average DSC values of WT, TC and ET were 0.897, 0.852 and 0.823 respectively. Compared with the existing efficient brain tumor segmentation methods, RAPNet has better effect in learning the multi-scale features of lesions.

Key words: brain tumor segmentation, feature pyramid, 3D recurrent residual convolution unit, attention mechanism, 3D atrous-convolution