RD2A: densely connected residual networks using ASPP for brain tumor segmentation

Abstract

The variations among shapes, sizes, and locations of tumors are obstacles for accurate automatic segmentation. U-Net is a simplified approach for automatic segmentation. Generally, the convolutional or the dilated convolutional layers are used for brain tumor segmentation. However, existing segmentation methods of the significant dilation rates degrade the final accuracy. Moreover, tuning parameters and imbalance ratio between the different tumor classes are the issues for segmentation. The proposed model, known as Residual-Dilated Dense Atrous-Spatial Pyramid Pooling (RD2A) 3D U-Net, is found adequate to solve these issues. The RD2A is the combination of the residual connections, dilation, and dense ASPP to preserve more contextual information of small sizes of tumors at each level encoder path. The multi-scale contextual information minimizes the ambiguities among the tissues of the white matter (WM) and gray matter (GM) of the infant’s brain MRI. The BRATS 2018, BRATS 2019, and iSeg-2019 datasets are used on different evaluation metrics to validate the RD2A. In the BRATS 2018 validation dataset, the proposed model achieves the average dice scores of 90.88, 84.46, and 78.18 for the whole tumor, the tumor core, and the enhancing tumor, respectively. We also evaluated on iSeg-2019 testing set, where the proposed approach achieves the average dice scores of 79.804, 77.925, and 80.569 for the cerebrospinal fluid (CSF), the gray matter (GM), and the white matter (WM), respectively. Furthermore, the presented work also obtains the mean dice scores of 90.35, 82.34, and 71.93 for the whole tumor, the tumor core, and the enhancing tumor, respectively on the BRATS 2019 validation dataset. Experimentally, it is found that the proposed approach is ideal for exploiting the full contextual information of the 3D brain MRI datasets.