A denoising autoencoder-enhanced ResNet 50 framework for explainable brain tumor classification using MRI.
Authors
Affiliations (6)
Affiliations (6)
- Department of Radiological Sciences, College of Applied Medical Science, King Khalid University, Abha, Saudi Arabia.
- Department of Radiological Sciences, College of Applied Medical Sciences, Taif University, Taif, Saudi Arabia.
- Radiological Sciences department, College of Applied Medical Sciences, Najran University, Najran, Saudi Arabia.
- Radiological Sciences Department, College of Applied Medical Sciences, King Saud University, Riyadh, Saudi Arabia.
- Radiologic Sciences Department, Faculty of Applied Medical Sciences, King Abdulaziz University, Jeddah, Saudi Arabia.
- College of Medicine, Alfaisal University, Riyadh, Saudi Arabia.
Abstract
Brain tumors are a major issue in neurological health where the timely and accurate diagnosis is critical in the effectiveness of the clinical intervention and the outcome of the patients. It is believed that magnetic resonance imaging (MRI) is the imaging modality that is predominantly used to evaluate brain tumors due to the fact that it has excellent soft-tissue contrast, however, the analysis of automated MRI is hindered by noise, inhomogeneity in intensities, inter-scanner variation, and even the visual similarity that is exhibited by different tumor subtypes. In order to overcome these challenges, the current research presents a powerful, understandable deep learning architecture to be used in the context of multi-classification of brain tumors using MRI scans. The proposed methodology is a denoising autoencoder (DAE) based image improvement step with a transfer-learning based ResNet-50 classifier. The DAE is trained without any supervision to suppress noise and emphasize salient anatomical structures before classification. The improved images then undergo refinement of a pretrained ResNet-50 network to label MRI scans with four clinically relevant classes. The experimental assessment of a publicly available benchmark dataset shows that the framework has a test accuracy of 98.93% and the precision, recall, and F1-scores are high in all classes. Furthermore, the gradient mapping technique (Grad-CAM) is used to provide visual explainability of the model predictions by highlighting image regions that influence the classification outcomes. These results show the potential of the proposed framework as an explainable AI-assisted decision-support approach for brain tumor MRI classification. Further validation using external and multi-center clinical datasets is required before clinical deployment.