Multi-Class Brain Tumor Detection Using Deep Learning with Explainable AI Visualization

Abstract

The detection of brain tumors is a fundamental problem in medical image analysis that has a significant part in patient survival since identification at early stages increases the chances of surviving the disease. MRI scan interpretation takes a a great deal of time and can be prone due to human error. In this paper, an automated multi-class brain tumor detection system is introduced based on deep learning that visualizes the system with Explainable Artificial Intelligence (XAI). VGG16, DenseNet121 and ResNet50 pre-trained CNN models are tested through transfer learning on brain MRI dataset with 7200 images, which are divided into 4 classes: Glioma, Meningioma, No Tumor and Pituitary. Training images: 5,600, Testing images: 1,600. ResNet50 also obtains the highest classification accuracy of 88.69%, better than VGG16 (78.69%) and DenseNet121 (84.38%). The Grad-CAM visualization of the tumor region where the model uses it for prediction helps with interpretability in the clinical context. A Streamlit web app allows for real-time tumor detection. The proposed system shows promising results in computer-aided diagnosis in clinical settings.