Supplementary material for: An Explainable Deep Learning Model Classifies Eight Categories of Pigmented Skin Lesions on Clinical Photographs: A Multicenter Retrospective Internal Validation Study in Japan

Description

This dataset contains the Supplementary Materials for the manuscript "An Explainable Deep Learning Model Classifies Eight Categories of Pigmented Skin Lesions on Clinical Photographs: A Multicenter Retrospective Internal Validation Study in Japan," accepted for publication in JAAD International. The Supplementary Materials provide extended methodological details and additional results supporting the findings reported in the main manuscript, including: - Supplementary Methods: ensemble architecture and prediction aggregation (20 models comprising EfficientNet-B4, EfficientNet-B7, Vision Transformer, and Swin Transformer across five cross-validation folds), dataset and reference standard, patient-level data splitting, and model interpretability analysis using Gradient-weighted Class Activation Mapping (Grad-CAM). - Supplementary Table 1: Accuracy of individual architectures and the full ensemble on the independent test set. - Supplementary Figure 1: (a) Receiver operating characteristic curves and (b) precision-recall curves for each pigmented lesion category. - Supplementary Table 2: Detailed classification performance metrics for each pigmented lesion category on the independent test set. - Supplementary Table 3: Quantitative Grad-CAM-based attention metrics across lesion categories. - Supplementary Figure 2: Representative Grad-CAM visualizations demonstrating lesion-centered, disease-specific activation patterns. - Supplementary Figure 3: Representative misclassified cases between seborrheic keratosis and basal cell carcinoma. Note: This dataset contains supplementary documentation only. The underlying clinical image dataset is not publicly available due to institutional and ethical restrictions and is available from the corresponding author upon reasonable request.

Steps to reproduce

The supplementary materials were generated during preparation of the final manuscript. They include additional methodological details, performance tables, and Grad-CAM visualizations supporting the analyses reported in the associated manuscript. The underlying deep learning models were developed using five-fold cross-validation with EfficientNet-B4, EfficientNet-B7, Vision Transformer, and Swin Transformer architectures. This dataset contains documentation only and does not include the underlying clinical image dataset.

Institutions

• Kindai University Hospital

  Osaka, Osaka

Categories

Licence