Early detection of multiple sclerosis through transfer learning: a weighted snapshot fusion approach with ResNet101

Multiple sclerosis (MS) is a chronic neurological disorder that affects the central nervous system. Early detection of MS is crucial for timely intervention and better patient health outcomes. However, the irregular classification mapping of distinct features of MS poses challenges to accurate predictions. This paper introduces a novel Weighted Snapshot Fusion Ensemble (WSFE) technique that leverages transfer learning with a pre-trained Residual Network (ResNet101) to enhance the classification accuracy of MS. The proposed WSFE applies five different learning rates to the pre-trained ResNet101 model and computes the accuracy of probability prediction by continuously changing the learning rate from maximum to minimum and vice versa. Higher priority is given to the instantly acquired maximum accuracy; accordingly, the weights of the corresponding learning rate are updated. This information is then utilized in an additional layer placed in the pre-trained ResNet101 model to extract unique features through transfer learning. The effectiveness of the WSFE approach is visualized using the GradCAM technique, and additional layers are included in the pre-trained ResNet101 model, resulting in a Modified Residual Network (M-ResNet101) model. Experimental results demonstrate the effectiveness of the proposed WSFE technique with M-ResNet101, achieving a high precision of 99.34%, accuracy of 99.23%, recall of 98.63%, and F1 score of 99.46%, surpassing the performance reported by previous frameworks, including VGG-16, GoogleNet, MobileNetv2, ExceptionNet, SqueezeNet, ResNet50, and AlexNet. The proposed approach shows promise for improving the early detection possibilities of MS compared to existing methods.