Supplementary Figures 1-5 and Table I (Expert-level identification of pemphigus and bullous pemphigoid based on a multimodal artificial intelligence framework)

Published: 16-06-2020| Version 2 | DOI: 10.17632/437v5dp32m.2
Xiaoyu He


Figure 1: The co-attention structure in the proposed multimodal network. Figure 2: Test results of the classification of the five categories of skin diseases and normal skin. NORMAL, AD, BD, RID, ED, DE denote normal skin, acne dermatoses, bullous dermatoses, rheumatic immune dermatoses, erythematous dermatoses, and drug eruptions, respectively. A, Confusion matrix of the clinical image classification model. It shows the detailed classification accuracy of each class. B, The receiver operating characteristic (ROC) curves and the area under the ROC curves (AUROC) of each class. Figure 3: Test results of the classification of pemphigus and bullous pemphigoid. A, The sensitivity, specificity, and AUROC of the metadata-only model, clinical-image-only model, and our multimodal model. B, Feature importance ranking of the metadata-only model. The horizontal axis includes 19 features in the metadata feature vector, and the vertical axis includes the feature importance values calculated by the random forest algorithm. Figure 4: Comparison between other multimodal models and our multimodal model. Figure 5: Test results of the model interpretability experiment. A, Some image samples of pemphigus and bullous pemphigoid and their corresponding saliency maps. B, The distributions of points before (left side) and after (right side) the decision-making strategy. Table I: Investigation of the acceptance of dermatologists at the primary areas to artificial intelligence tools for bullous dermatoses.