Automated Brain Disease Classification using Transfer Learning based Deep Learning Models

Abstract

Brain MRI (Magnetic Resonance Imaging) classification is one of the most significant areas of medical imaging. Among different types of procedures, MRI is the most trusted one to detect brain diseases. Manual and semi-automated segmentations need highly experienced radiologists and much time to detect the problem. Recently, deep learning methods have taken attention due to their automation and self-learning techniques. To get a faster result, we have used different algorithms of Convolutional Neural Network (CNN) with the help of transfer learning for classification to detect diseases. This procedure is fully automated, needs less involvement of highly experienced radiologists, and does not take much time to provide the result. We have implemented six deep learning algorithms, which are InceptionV3, ResNet152V2, MobileNetV2, Resnet50, EfficientNetB0, and DenseNet201 on two brain tumor datasets (both individually and manually combined) and one Alzheimer’s dataset. Our first brain tumor dataset (total of 7,023 images-training 5,712, testing 1,311) has 99-100 percent training accuracy and 98-99 percent testing accuracy. Our second tumor dataset (total of 3,264 images-training 2,870, testing 394) has 100 percent training accuracy and 69-81 percent testing accuracy. The combined dataset (total of 10,000 images-training 8,000, testing 2,000) has 99-100 percent training accuracy and 98-99 percent testing accuracy. Alzheimer’s dataset (total of 6,400 images-training 5,121, testing 1,279, 4 classes of images) has 99-100 percent training accuracy and 71-78 percent testing accuracy. CNN models are renowned for showing the best accuracy in a limited dataset, which we have observed in our models.