Enhancement of the U-net Architecture for MRI Brain Tumor Segmentation

Abstract

The magnetic resonance images (MRI) brain tumor segmentation is one of the most difficult medical images segmentation, and it has many challenges because the tumor has no specific shape or size, not found in a specific place of the brain and contains three sub-regions (full tumor FT, tumor core TC, and enhanced tumor ET). The manual segmentation is extremely difficult and prone to mistakes. In this research, the semantic segmentation is used by implementing the U-net model, which is a fully convolutional network (FCN) algorithm on BraTS 2018 that contains four modalities (T1, T2, T1c, Flair). The U-net model is used two time: The first is using 9-layers U-net with some enhancements on the original model for segment the full tumor, and the second is using 7-layers U-net model for segment the tumor core and enhanced tumor. The results were promised by segmenting the brain tumor within three sub-regions (full tumor, tumor core, and enhanced tumor), and the results were evaluated using standard brain tumor segmentation metrics. The proposed system achieves mean dice similarity coefficient metric; it is 0.87, 0.76, and 0.71 for full tumor, tumor core, and enhanced tumor, respectively. Additionally, the median dice similarity coefficient metric is 0.90, 0.84, and 0.80 for full tumor, tumor core, and enhanced tumor, respectively.KeywordsBrain tumorBraTS 2018U-netFCNDeep learningMRI