White patchy skin lesion classification using feature enhancement and interaction transformer module

Abstract

White patchy skin lesions have always been difficult to distinguish, yet precise identification of specific types can enable targeted treatment and alleviate patient anxiety. Deep convolutional neural networks (DCNNs) show great potential in this regard. However, DCNNs still exhibit limitations such as incapacity to capture global correlations, inability to discern invisible data distributions, and difficulty in handling imbalanced datasets. To address these challenges, we propose a feature enhancement and interaction transformer module for accurately identifying common white patches. Our approach begins with the development of a dual-position encoding attention and convolution hybrid submodule, which aims to model the global information of the feature domain and enhance feature representation. Subsequently, we construct a feature interaction submodule on the batch dimension to enable the DCNN to explore sample relationships, learn about invisible distribution from the dataset, and reduce the imbalance problem. Based on the dataset comprising four types of white patchy skin lesions, the proposed approach achieved an accuracy, precision, recall, F1 score, and AUC of 92.65 %, 92.83 %, 92.65 %, 92.74 %, and 0.98, respectively. These results demonstrate the superior performance of our approach compared to other state-of-the-art models, underscoring its potential to enhance the classification of white patchy skin lesions and expand their clinical applications without compromising the integrity of the DCNN structure.