An Improved Capsule Network for Glioma Segmentation on Mri Images: A Curriculum Learning Approach Publisher Pubmed



Amiri Tehrani Zade A^{1, 2} ; Aziz MJ^{1, 2} ; Masoudnia S^{3, 5} ; Mirbagheri A^{2, 4} ; Ahmadian A^{1, 2} Show Affiliations
Source: Computers in Biology and Medicine Published:2022

Abstract

Glioma segmentation is an essential step in tumor identification and treatment planning. Glioma segmentation is a challenging task because it appears with blurred and irregular boundaries in a variety of shapes. In this paper, we propose an efficient and novel model for automatic glioma segmentation based on capsule neural networks. We improved the architecture and training of the SegCaps model, the first capsule-based segmentation network. The proposed architecture is improved by introducing dilation blocks in the primary capsule block to get deeper features while avoiding resolution reduction. The prediction layer of the network is also modified using one-dimensional convolution filters, enabling the network to not only maximize tumor existence likelihood but also regularize the capsule orientations within the tumor. Our main contribution, however, is to introduce an enhanced curriculum-based training algorithm into the proposed SegCaps model. We adapt the curriculum learning for the model by suggesting a new pacing mechanism based on a roulette-wheel selection algorithm that enriches randomness in the network and prevents bias. A hybrid dice loss function is also employed, which is better adapted to the introduced curriculum-based training procedure. We evaluated the performance of improved SegCaps on the BraTS2020, a multimodal benchmark dataset for brain tumor segmentation. The experimental results confirmed that the improvements yield a top-performing yet memory-efficient deep capsule architecture. The proposed model outperformed the best-reported accuracies on the BraTS2020, achieving improved dice scores of 85.16% and 81.88% for tumor core and enhancing tumor segmentation, respectively. Using 90%, fewer parameters than the popular U-Net also confirmed the high memory efficiency of our proposed model. © 2022