Brain Tumor Classification: XAI with Lime

Seow, Zhi Hung (2023) Brain Tumor Classification: XAI with Lime. Final Year Project (Bachelor), Tunku Abdul Rahman University of Management and Technology.

Text Full Text -Seow Zhi Hung.pdf Restricted to Registered users only Download (6MB)

Abstract

Brain tumors are among the most common cancers in humans. As a result, early intervention is required to diagnose and treat brain tumors. Diagnosing brain tumors based on visual judgments has been the norm practiced by medical professionals. Nonetheless, manual interpretation of brain images suffers the risk of misdiagnosis owing to human mistakes caused by severe workload, weariness, and interpretation bias. Professionals with a high level of medical competence are also necessary to determine the type of tumor. Such challenges call for the development of a brain tumor classification system that uses artificial intelligence (AI) to conduct autonomous brain tumor categorization. As a result of evaluating several Convolutional Neural Network (CNN) architectures and documented literature on their applications in biological imaging, three CNN models for brain tumor classification are suggested to be developed in this study: VGG-16, ResNet-50, and MobileNetV2. The hyperparameters would be tweaked to improve the CNN models’ performance. The various architecture performances with the given hyperparameters will be tested and compared using the well-established deep learning (DL) assessment measures like precision, recall, F1 score, and accuracy. The system will be built in Python and run on Kaggle utilizing images from the Brain Tumor Magnetic Resonance Image (MRI) dataset, which has the MRI split into four groups: glioma, meningioma, no tumor, and pituitary. The study report with the successful use of VGG16 CNN model, with a specific configuration of several epochs of 150, a dropout rate of 0.3, and a learning rate of 0.0001, and the optimizer is Adam, for predicting brain tumors with high accuracy. The model outperforms other CNN models, such as ResNet50 and MobileNetV2. The study demonstrates that VGG16 can predict all types of brain tumors with an accuracy rate of 99.7%, making it a reliable model for diagnosing brain tumors. However, the absence of the explanation ability of deep neural models is thought to be the main barrier to using these approaches in clinical practice. Hence, local interpretable model-agnostic explanation (LIME), a form of explainable AI (XAI) is used for the explanation in the classification of the brain tumor. Throughout this study, applied LIME explanation can identify which regions are crucial for accurate brain tumor prediction. The LIME explanation uses colored indicators to differentiate between important and unimportant regions, demonstrating which areas are significant in identifying brain tumors.