An Explainable Deep Learning Framework for Brain Tumor Detection Using MRI Images

Brain tumor detection using Magnetic Resonance Imaging (MRI) is a critical diagnostic procedure that demands high interpretability, accuracy, and efficiency. This paper presents a system of brain tumor classification that is based on interpretable deep learning model. The Convolutional Neural Network (CNN) used is a customized one trained on two publicly obtained Br35H dataset along with a four-class brain tumor MRI dataset. There are four-class brain tumor, glioma, meningioma, pituitary and no tumor. Image denoising, data augmentation, and normalization are applied using the methodology to enhance robustness and generalizability in models. Gradient-weighted Class Activation Mapping (Grad-CAM) and Local Interpretable Model-agnostic Explanations (LIME) are combined to help tackle the problem of deep learning model opacities. These devices visualise class-discriminating areas and important local superpixels enhancing clinical transparency. The proposed CNN has been experimentally demonstrated to achieve about 94% and 98% accuracy on the Br35H dataset and the multiclass brain tumor MRI dataset respectively. The accuracy, recognition, and the F1-scores are comparable across classes. The results indicate that the framework is capable of capturing the features, type of tumor, and generates interpretable visual data to be relevant to the clinical world. The paper presents a full and interpretable deep learning model in the diagnosis of brain tumors acquired through MRI. It helps close the high diagnostic accuracy/reliable model explainability gap.