A Novel Deep Learning Model Architecture Design that Improves Skin Lesion Diagnosis

Mainstream deep learning models do not perform well in skin lesion image classification for skin disease diagnosis due to multiple factors, such as noise and illumination intensity variations of skin lesion images, similarity of lesion features, morphological diversity, small volume of datasets and unbalanced image samples. In order to improve the diagnosis performance, in this paper, we first preprocess skin lesion images through a new color constant algorithm to mitigate their illumination intensity variations. Then, we apply a novel deep learning model, which combines multi-scale convolution and attention mechanisms. In the model, stacked multi-scale convolution are used to obtain local and global features of lesion images. Both local and global features are fused as input of the flatten transformer shift window. This approach expands local receptive field of the shift window, focuses more on local and global lesion feature areas of a skin lesion and enhances relationship among local and global features. Extensive experiments demonstrate the efficacy of our method. Compared with other models, our proposed model pays more attention to features in the skin lesion areas. The overall accuracy rate, specificity, and sensitivity have increased by 2.91%, 0.58%, and 4.02% respectively. The overall F1-Score has reached the state-of-the-art value of 0.8841.