Deep Learning‐Based Skin Lesion Classification

High-frequency ultrasound (HFUS) is valuable for assessing skin lesions, supporting diagnosis, treatment monitoring, and surgical planning. This study evaluates deep learning models for binary classification of HFUS images acquired in B-mode and Doppler mode. Two single-input CNNs were trained with each modality, while Unity and Cascade architectures combined both. The HFUS-Doppler model achieved the best performance (95.0% accuracy, AUC 0.98), followed by Unity (90.5% accuracy, AUC 0.97). Cascade showed lower accuracy but greater confidence in malignant predictions. Probability distribution analysis revealed differences in model certainty near the decision threshold. Results indicate that combining B-mode and Doppler can enhance diagnostic performance, depending on network design and data quality, supporting the potential of customized deep learning for non-invasive HFUS-based skin lesion classification.