Two-Stage Deep Learning Pipeline for Brain Tumor MRI Classification with Explainable AI

Abstract

Brain tumor recognition from magnetic resonance imaging (MRI) is one of the crucial applications of medical image analysis. MRI scans are difficult to analyze manually and require considerable time, while variations in tumor appearance can make classification challenging. Deep learning models can facilitate this process by learning relevant visual patterns directly from MRI images. However, classification systems that provide only a predicted class may offer limited information about the regions influencing the decision. In this work, a two-stage deep learning pipeline for brain tumor classification from MRI images is proposed. In Stage 1, a MobileNetV2 model performs normal and abnormal screening. Images identified as abnormal are further processed in Stage 2 using a VGG16 network enhanced with a Squeeze-and-Excitation (SE) attention block. The Stage 2 model performs four-class classification into glioma, meningioma, pituitary tumor, and non-tumor. Grad-CAM is incorporated to visualize the image regions contributing to the classification result. The trained models are deployed in a Flask-based web application for image uploading and result visualization. Experiments were conducted using the Kaggle Brain Tumor MRI Dataset containing 7,023 MRI images. The MobileNetV2 screening model achieved an accuracy of 96.38%. Among the models evaluated in Stage 2, the VGG16 model with SE attention achieved the highest accuracy of 93.94%. The system integrates normal/abnormal screening, detailed classification, visual explanation, and web-based access within a single two-stage pipeline.