Multi-view brain tumor segmentation (MVBTS): An ensemble of planar and triplanar attention UNets

Authors: SNEHAL RAJPUT, RUPAL KAPDI, MEHUL RAVAL, MOHENDRA ROY

Abstract: 3D UNet has achieved high brain tumor segmentation performance but requires high computation, large memory, abundant training data, and has limited interpretability. As an alternative, the paper explores using 2D triplanar (2.5D) processing, which allows images to be examined individually along axial, sagittal, and coronal planes or together. The individual plane captures spatial relationships, and combined planes capture contextual (depth) information. The paper proposes and analyzes an ensemble of uniplanar and triplanar UNets combined with channel and spatial attention for brain tumor segmentation. It investigates the significance of each plane and analyzes the impact of uniplanar and triplanar ensembles with attention to segmentation. We tested the performance of these variants on the BraTS2020 training and validation datasets. The best dice similarity coefficients for enhancing tumor, whole tumor, and tumor core over the training set are 0.712, 0.897, and 0.837, while they are 0.699, 0.875, and 0.782, over the validation set, respectively (obtained through BraTS model evaluation platform). The scores are at par with the leading 2D and 3D BraTS models. Therefore, the proposed approach with fewer parameters (almost 3× less) demonstrates comparable performance to that of a 3D model, making it suitable for brain tumor segmentation in resource-limited settings.

Keywords: Attention network, gliomas, triplanar ensemble, brain tumor segmentation, UNet

Full Text: PDF