Enhanced Monkeypox Skin Lesion Classification Using Pooling-Based Vision Transformer Architecture

Abstract

Monkeypox is viral disease transmitted from animals to man and presents symptoms of smallpox especially rashes and lesions on the skin. The recent mutation that has led to human-to-human transmission has caused international concern and therefore enhanced method of diagnosing is required and proved. In this part of the work, we bring forward a powerful approach for the monkeypox classification with a pooled-based vision transformer mode called as Pooling-based Vision Transformer (PiT) architecture merged with MobileNetV3 that is trained with Adam optimizer. By merging the strengths of both architectures can enhance representation power by integrating local and global feature extraction. This hybrid approach significantly reduces computational load through techniques like token pooling, leading to higher accuracy without a proportional increase in computational costs. The Lion optimizer is employed to enhance the accuracy of the model and enhance convergence response and performance in contrast to other optimizers. For the classification task, the performance of the proposed model was 94.23, 91, 93.5 and 90.75 % occupancy for accuracy, precision, recall, and F1 score.