ABSTRACT
The need for accurate and intuitive diagnostic tools in the field of biomedical science has incited many researchers to focus on integrating the various models offered by deep learning to suit therapeutic utility. Neural networks, in particular, have been exhaustively worked upon to aid in image segmentation and classification operations. This project aims to address the need for automatic segmentation of MRI images of gliomas in cancer patients. The DenseNet architectural variant of convolutional neural networks has been utilized to build a highly accurate 3D segmentation tool. Our program also functions to perform preliminary classification of the tumors into high-grade tumors and low-grade tumors to afford the medical community an initial insight into the severity of the gliomas. We have utilized the open-source Python library function to build our architecture. The training and testing of this architecture were performed on the benchmark MICCAI BRaTS dataset. Our model showed high segmentation precision where we obtained accuracy close to 100% (~99.94% with enhancements). The same values were obtained for sensitivity and positive predictive values (PPVs). This level of segmentation accuracy can be considered ideal for practical medical use.
