Classifying Brain Tumours: A Deep Learning Approach with Explainable AI

Abstract

Brain tumours stand out as one of the deadliest cancers, urging the need for early detection and intervention. However, the diverse characteristics of these tumours, ranging from variations in shape, location, type, to size, alongside the inherent variations in intensity across magnetic resonance imaging (MRI) images, contribute to the complexity of MRI analysis. To tackle these difficulties, this study proposes the development of a deep learning model for autonomous brain tumour classification, leveraging Convolutional Neural Networks (CNNs) specifically, VGG16, ResNet50, and MobileNetV2. This study fine-tuned hyperparameters, including learning rate, optimizer, number of epochs, and dropout rate, to enhance model performances. All three CNN models exhibited good performance for brain tumour classification, with the VGG16 model, configured with the Adam optimizer, 150 epochs training time, a 0.3 dropout rate, and a learning rate of 0.0001, demonstrating superior accuracy (98%) in predicting brain tumours. The study recognizes the challenge of interpretability in deep neural models, addressing it by employing Local Interpretable Model-agnostic Explanations (LIME). LIME's explanation, employing superpixels and coloured indicators, reveals crucial regions considered by the CNN for brain tumour class prediction, offering insights into the decision-making process.

Topics & Concepts