MOAT: MobileNet-Optimized Attention Transfer for Robust and Scalable Dermatology Image Classification

In recent times, dermatological disease is a common health issue around the world. Timely and accurate dermatological disease detection is mandatory for proper treatment planning and improving patient outcomes. Prior to this, various detection and classification methodologies were developed for the early prediction of skin disease using dermatology images. However, they have struggled with poor detection accuracy and require high computational time. This research proposes a novel MobileNet-Optimized Attention Transfer framework for the accurate classification of skin disease. In this study, the MobileNet model is deployed for feature extraction, which integrates the self-attention and cross-attention mechanisms. The attention mechanisms prioritize essential features within the images and allow the model to identify the subtle patterns associated with various skin conditions. For hyperparameter tuning, an Optical Microscope Algorithm with an initial search strategy is applied. The algorithm iteratively fine-tunes the parameters to balance global and local search and prevent the model from converging on suboptimal configurations. The performance of the proposed method is validated using the Skin Cancer ISIC dataset and the Skin Cancer MNIST: HAM10000 dataset and compared to existing skin disease classification methodologies in terms of some common assessing metrics. The experimental results demonstrate that the MobileNet-Optimized Attention Transfer framework effectively classified the dermatological disease and achieves a high accuracy of 98.89%, a lower mean squared error of 0.186, and a low computational time of 1.53 s compared to existing methodologies. This result indicates that the proposed framework is a scalable, adaptable solution for clinical applications, aiding dermatologists with reliable diagnostic support and enabling early intervention.