Proportion Estimation and Multi-Class Classification of Abnormal Brain Cells

Abstract

Vagueness in the determination of the tumor size creates significant hindrances in planning and quantitatively assessing brain tumor treatments. Non-invasive magnetic resonance imaging has become a primary non-ionizing radiation diagnostic tool for brain cancers. It takes a long time to manually segment the extent of a brain tumor from 3D M.R.I. volumes, and the performance heavily depends on the operator's skill. A precise and automated brain tumor segmentation tool is needed desperately. In this case, an accurate assessment of the tumor's extent requires a reliable automated segmentation method for the brain tumor. The Multimodal Brain Tumor Image Segmentation (BRATS 2020) dataset is used in this paper to demonstrate an automated deep convolutional network, or U-Net, method for brain tumor segmentation. Deep Learning and Transfer Learning are utilized to improve the accuracy and effectiveness in detecting and recognizing different types of brain cancers. The unobserved images' F1 scores were 98% and 99%, respectively.