Enhanced 3D U-Net Segmentation Architecture for the Detection and Localization of Brain Tumor

Brain Tumor is rapidly growing problem these days for every age group of people, and thus to accurately detect a tumor, it is essential that we render to automatic brain tumor segmentation. In this study, semantic 3D U-Net segmentation architecture is proposed using Magnetic Resonance Imaging (MRI). Basically this method not only detects the tumor but also convey the location of a tumor. In this, we have applied multiple preprocessing techniques for better results and the resultant of these multiple techniques is fed to the proposed architecture of U-Net. This paper has also tried to minimize the mean-variance problem, handle the different sizes of tumor and also to deal with address imbalance problem. In this study, different parameters are evaluated in terms of brain tumor segmentation i.e. Dice similarity coefficient, Sensitivity, Specificity, Accuracy and Mean IoU. In the study, three classification models have also been used to evaluate the survival rate. The experiments have been evaluated using BraTS 2020 dataset with 98.91% accuracy for segmentation task.