Ensemble learning based-features extraction for brain mr images classification with machine learning classifiers

In general, different neuroimaging methods including Computed Tomography (CT), Magnetic Resonance Imaging (MRI), and Positron Emission Tomography (PET) are employed to assess the tumor in the brain. For this study, MR images are utilized to diagnose brain tumors. To assist doctors and radiologists in automatic brain tumor diagnosis and to overcome the need for manual diagnosis, a brain MR image automated classification system is being developed. The latest advancement of deep learning has helped medicine and neuroimaging in the diagnosis of many diseases, and Convolutional Neural Networks (CNN) is an extremely frequently employed deep learning approach in the automatic image classification area. This paper presents a novel approach for classifying brain MR images utilizing a dataset consisting of 5712 MR images. The dataset was partitioned into training and validation sets using an 80-20 ratio to ensure unbiased evaluation. To extract informative features from brain MR images, we employ Transfer Learning, utilizing seven pre-trained CNNs, including ResNet-50, DenseNet-121, VGG-19, EfficientNetB1, EfficientNetV2B1, Inception V3, and MobileNet. Multiple machine learning classifiers are utilized to evaluate the extracted features, and the three best-performing features are chosen to form a robust Features Ensemble. This ensemble is then combined with the best-performing machine learning classifier, the Multilayer Perceptron (MLP), to classify brain MR images into four categories: Glioma tumor, Meningioma tumor, Pituitary tumor, and Normal brain. Our experimental findings indicate that the fusion of extracted features from ResNet-50, VGG-19, and EfficientNetV2B1, using the Features Ensemble approach with the MLP classifier, significantly enhances performance. We achieve an impressive accuracy of 96,67% (+/- 1,04) with a confidence level of 95%. Furthermore, our proposed technique surpasses existing state-of-the-art methods. These results underscore the effectiveness of our approach in accurately classifying brain MR images and its potential to improve diagnostic capabilities.