Intelligent brain tumor detection and classification to assist physician in clinical diagnostic system

Abstract

This project is an artificial intelligence model that classify brain tumor MR images into three different classes, namely Glioma, Meningioma and Pituitary Tumors. The existing method of analysing MRIs is manual classification, which suffer from difficulties such as the long time it takes to classify and the accuracy that can vary based on the experience of the physicians. The researchers are working on classification of MRIs since years, and each of them are competing to get a higher accuracy and performance results. However, the competition in this field is widely focusing on getting higher accuracy and better performance and trying different datasets to get variety of all possible combinations. After doing a successful experiment on Alexnet network and reaching an accuracy better than the state of the art. After noticing that the research field is full of researches, but no real application is applied in the hospitals, it is the time to start thinking practically about moving the research one step toward practical side, which is the medical application of this problem. In this project, an application is developed for giving multiple opinions about MR image of a brain tumor of the three types, helping the physicians with not only 2nd opinion, but with 4 different opinions from four different AI entities, increasing the accuracy that can be obtained in deciding which tumor is in the image, in an easy to use environment with few clicks, making the numbers and technical aspect of the AI technology to us as engineers and the solution is simplified as possible in the hands of physicians. The pre-trained networks used in the project are Googlenet, Alexnet, Mobilenetv2, Resnet101, and the training accuracy obtained using the Figshare dataset on all of them are 100%, 97.66%, 100%, 100% respectively, and a validation accuracy of 92.27%, 86.87%, 94,34%, and 94.23% respectively.