Attention Mechanism, Linked Networks, and Pyramid Pooling Enabled 3D Biomedical Image Segmentation

Abstract

We present an approach to detect and segment tumorous regions of the brain by establishing three varied segmentation architectures for multiclass semantic segmentation along with data-specific customizations like residual blocks, soft attention mechanism, pyramid pooling, linked architecture, and 3D compatibility to work with 3D brain MRI images. The proposed segmentation architectures namely, Attention Residual U-Net 3D (ARU-Net 3D), LinkNet 3D and PSPNet 3D, segment the MRI images and successfully isolate three classes of tumors. By assigning pixel probabilities, each of these models differentiates between pixels belonging to tumorous and non-tumorous regions of the brain. By experimenting and observing the performance of each of the three architectures using metrics like Dice Loss and Dice Score, on the BraTS2020 dataset, we successfully establish the following validation scores: 0.6488, 0.6485, and 0.6501 for the ARU-Net, LinkNet, and PSPNet 3D architectures respectively. Code has been made available at: https://github.com/indiradutta/BrainTumorSeg-III-D.