SEMC-Net: A Shared-Encoder Multi-Class Learner

Abstract

Brain tumour segmentation is a crucial task in medical imaging that involves identifying and delineating the boundaries of tumour tissues in the brain from MRI scans. Accurate segmentation plays an indispensable role in the diagnosis, treatment planning, and monitoring of patients with brain tumours. This study presents a novel approach to address the class imbalance prevalent in brain tumour segmentation using a shared-encoder multi-class segmentation framework. The proposed method involves training a single encoder class learner and multiple decoder class learners, which are designed to learn feature representation of a certain class subset, in addition to a shared encoder between them that extracts common features across all classes. The outputs of the complement-class learners are combined and propagated to a meta-learner to obtain the final segmentation map. The authors evaluate their method on a publicly available brain tumour segmentation dataset (BraTS20) and assess performance against the 2D U-Net model trained on all classes using standard evaluation metrics for multi-class semantic segmentation. The IoU and DSC scores for the proposed architecture stands at 0.644 and 0.731, respectively, as compared to 0.604 and 0.690 obtained by the base models. Furthermore, our model exhibits significant performance boosts in individual classes, as evidenced by the DSC scores of 0.588, 0.734, and 0.684 for the necrotic tumour core, peritumoral edema, and the GD-enhancing tumour classes, respectively. In contrast, the 2D-Unet model yields DSC scores of 0.554, 0.699, and 0.641 for the same classes, respectively. The approach exhibits notable performance gains in segmenting the T1-Gd class, which not only poses a formidable challenge in terms of segmentation but also holds paramount clinical significance for radiation therapy.