Detection of brain cancer from MRI images using Deep Learning

Abstract

The diagnosis of brain tumors from Magnetic Resonance Imaging (MRI) images is a challenging clinical task, due to the visual similarity of the tumor subtypes and the variability of the medical imaging data. Deep learning has demonstrated great potential in automating this process, but most of the existing frameworks treat detection and classification as a one-stage task, which may result in inter-class confusion and hinder the diagnostic reliability. This work proposes an automatic two-stage hybrid deep learning framework for brain tumor analysis from MRI images to overcome these limitations. In the first stage, a hybrid CNN-VGG19 model performs binary classification to differentiate normal brain scans from tumor-positive cases. In the second stage, the detected tumors are classified into three clinically relevant subtypes, glioma, meningioma, and pituitary tumor, by a fine-tuned ResNet50 model incorporating dual pooling and progressive fine-tuning strategy. All images are pre-processed by CLAHE-based contrast enhancement and data augmentation before training to improve the input quality and generalization of the model. Experimental evaluation on two publicly available MRI datasets shows that the proposed framework achieves 99% accuracy in Stage 1 and 99% accuracy in Stage 2 with consistently high precision, recall and F1-score across all classes. The two-stage architecture reduces the inter-class confusion and provides a clinically interpretable decision making process, making it a promising computer-aided diagnosis tool to assist radiologists in brain tumor detection and classification