A Study on Effective Segmentation Network in MRI Images for Diagnosis of Brain Tumor

Brain glioma type's tumors are among the most toxic cancers worldwide. Now-a-days, due to the advance of convolutional neural networks, many brain tumor segmentation systems aid clinicians to diagnose and operate. To support clinicians in the accurate automatic semantic segmentation of brain tumors, we investigate two convolutional encoding-decoding neural network i.e. UNet and PSPNet for brain glioma tumour segmentation from multimodal MRI images. The study is trained from the 2020 Brain Tumour Segmentation (BraTS) challenge. A rigorous assessment and comparative evaluation on the BraTS'20 validation set and test set, respectively, evolved with the PSPN et architecture achieving dice scores of 0.876 and 0.905 for enhancing tumour region and necrotic tumour/core region, respectively and with the UN et architecture achieving dice scores of 0.842 and 0.908 for enhancing tumour region and necrotic tumour/core region, respectively.