U-Net multi-modality glioma MRIs segmentation combined with attention

Abstract

Glioma, the most common primary intracranial tumor, is known as the “brain killer,” accounting for 27% of all central nervous system tumors and 80% of malignant tumors, and is one of the most difficult and refractory tumors to treat in neurosurgery. The development of medical imaging technology has simplified the diagnosis of the disease, and in order to avoid or reduce the errors of manual segmentation, deep learning based segmentation of glioma has become the hope of radiologists and clinicians. Accurate segmentation of gliomas is an important prerequisite for making glioma diagnosis, providing treatment plans and evaluating treatment outcomes. To effectively target the characteristics of multimodal glioma MRI and the shortcomings of CNNs-based, U-Net-based glioma segmentation methods, a method of 2D-CNNs segmentation results based on attention mechanism is proposed. In this study, the datasets of BraTS2018 and BraTS2019 were included and the segmentation results were evaluated using three metrics: Dice coefficient, positive predictive value, and sensitivity. The experimental results show that the proposed segmentation method can accurately segment gliomas.