AN ATTENTION-ENHANCED NESTED 3D U-NET WITH COMPOUND LOSS FOR TUMOR CORE SEGMENTATION IN MULTISEQUENCE VOLUMETRIC BRAIN MRI

Abstract

Maximal removal of the tumor core tissues while preserving healthy brain tissues is essential to decrease tumor recurrence and improve patient outcomes. Therefore, a comprehensive pixel-wise understanding of brain MR images becomes imperative for the precise and automated identification of tumor core regions, such as enhancing tumor, non-enhancing tumor, and necrosis. Attention mechanisms allow deep learning models to focus on relevant regions, thereby improving the accuracy of tumor delineation, especially in the presence of class imbalance. In this work, we extend our previously proposed attention-enhanced nested U-Net and compound loss framework to a fully volumetric formulation for multisequence MRI segmentation. We trained and tested the proposed 3D model using 190 multisequence volumetric brain MR images in the BraTS 2019 benchmark dataset. The proposed model achieved Dice scores of 0.88, 0.78, and 0.84 on the BraTS 2019 dataset, for the whole tumor, enhancing tumor, and tumor core, respectively. Experimental results demonstrate competitive and consistent improvements in tumor core segmentation compared with recent 3D deep learning models