A Comparative Study of DenseNet-201 and Swin Transformer for Malignant and Benign Skin Lesion Classification

Abstract

Skin cancer has a high global prevalence, underscoring the need for accurate and efficient early detection systems to support screening. This study presents a comparative analysis of DenseNet-201 and Swin Transformer for binary classification of malignant and benign skin lesions using the BCN20000 dataset, which contains 12,413 dermoscopic images. The proposed workflow includes image preprocessing and augmentation, transfer learning-based model training, and evaluation under a 5-fold stratified cross-validation protocol. Performance is assessed using Accuracy, Precision, Sensitivity (Recall), F1-score, and the area under the receiver operating characteristic curve (AUC-ROC). In addition, computational efficiency is examined in terms of parameter count, model size, and training time. Across five folds, DenseNet-201 achieved 88.05% Accuracy, 88.90% Precision, 89.48% Sensitivity, 89.17% F1-score, and 94.73% AUC, whereas Swin Transformer achieved 87.42% Accuracy, 89.77% Precision, 87.06% Sensitivity, 88.39% F1-score, and 94.33% AUC. A paired t-test at α = 0.05 indicated no statistically significant performance difference between the two models. Model interpretability was investigated using Grad-CAM for DenseNet-201 and EigenCAM for Swin Transformer to verify that predictions were driven by lesion-relevant regions. Overall, the results suggest that both architectures are suitable candidates for dermoscopic image-based skin lesion screening support systems, including teledermatology applications.