Uncertainty Prediction in Brain Tumour Segmentation

Abstract

Gliomas are one of the most common brain tumour at different levels of the province, with Magnetic Resonance Imaging (MRI) used for diagnosis. In this project, It was asked to try to find uncertainty in the Brain Tumour Segmentation on MRI images using the BraTs19 Dataset and to look at how machine learning algorithms can work with these MRI images. Since these tissues are so large in shape and appearance, their separation becomes a challenge. The concept of distribution of predictive predictions, methods of extending the test time, and consolidation methods are used to reduce uncertainty and increase confidence in model predictions. This work proposes to be based on U-Net trained in integrated techniques to measure back memory usage and reduce the impact of unequal information. Accurate classification of different sub-regions of gliomas, as well as oedema, necrotic core, to improve and not improve the tumour total from imaging of modal scans, is necessary for clinical association in diagnosing, predicting, and treating brain tissue. Estimated measurements and planned uncertainties during this work were used in the Brain Tumour Segmentation Challenge (2019) related to lump separation and uncertainty estimates and used technique for uncertainty to obtain high accuracy and real-time performance.