AUTHOR=Wang Hui , Qi Qianqian , Sun Weijia , Li Xue , Yao Chunli 

TITLE=Classification of clinical skin lesions with double-branch networks

JOURNAL=Frontiers in Medicine

VOLUME=Volume 10 - 2023

YEAR=2023

URL=https://www.frontiersin.org/journals/medicine/articles/10.3389/fmed.2023.1114362

DOI=10.3389/fmed.2023.1114362

ISSN=2296-858X

ABSTRACT=Malignant skin lesions are a great threat to patients' health. Due to the limitations of existing diagnostic techniques such as poor accuracy and invasive operations, and the high similarity between malignant skin lesions and other skin lesions, there is a low diagnostic efficiency and a high rate of misdiagnosis. For this problem, we propose a DBN (Double-Branch Networks) based on a double-branch network model, which uses backbone with the same structure as the original network branch and the fusion network branch. The feature maps of each layer of the original network branch are extracted by our proposed CFEBlock (Common Features Extraction Block) to extract the common features of the feature maps between adjacent layers, after which these features are combined with the feature maps of the corresponding layers of the fusion network branch by FusionBlock, and finally the total prediction results are obtained by weighting the prediction results of the two branches. In addition, we constructed a new dataset CSLI (Clinical Skin Lesion Images), combining the publicly available dataset PAD-UFES-20 with our collected dataset, the CSLI dataset contains 3361 images of clinical skin diseases in 6 disease categories: Actinic Keratosis (730), Basal Cell Carcinoma of skin (1136), Malignant Melanoma (170), Melanocytic Nevus of Skin (391), Squamous Cell Carcinoma (298) and Seborrheic Keratosis (636). We divide the CSLI dataset into a training set and a test set, and we performed accuracy, precision, sensitivity, specificity, f1score, AUC summary, different model training visualization, ROC curve and confusion matrix for various diseases, and finally showed that the network performed better overall on test data.