Methodological Considerations for Improving Deep Learning‐Based Classification of Skin Lesions in High‐Frequency Ultrasound Imaging

To the Editor:

We read with interest the article by Silva et al on employing convolutional neural networks (CNNs) for the binary classification of skin lesions using high‐frequency ultrasound (HFUS) in B‐mode and Doppler modalities. ¹ The comparative study of single‐input and fused architectures provides a valuable empirical foundation. To further advance this line of research toward more robust and clinically integrated systems, we wish to propose 3 methodological considerations.

First, the study explores model fusion (Unity, Cascade) but does not explicitly address how to identify and weight the most influential feature pathways within the multimodal network. In complex systems such as fused CNNs, not all integrated features contribute equally to a reliable diagnosis. Techniques that evaluate both local feature saliency and global network influence, concepts also explored in complex network analysis, ² could help prioritize the most informative feature maps from B‐mode and Doppler branches. We suggest implementing attention mechanisms or learnable weighting layers during fusion to dynamically emphasize diagnostically relevant features. Such an approach may enhance the Cascade model's interpretability and potentially its performance by reducing noise from less contributory modalities.

Second, the training process for the deep learning models, while employing class weighting, may not optimally handle challenging samples near the decision boundary, the identified “zone of uncertainty.” General‐purpose optimizers like Adam may not sufficiently focus on these ambiguous cases, which are critical for clinical safety. Recent advances in optimization methods designed for deep networks, including those that adaptively adjust gradient contributions based on learning difficulty, ³ offer promising avenues. We recommend exploring specialized training paradigms that increase the model's focus on borderline cases. This could help sharpen the decision boundary and improve both recall and specificity, thereby increasing clinical reliability.

Third, the CNN architectures utilized standard pooling layers (eg, Global Average Pooling) for downsampling. While effective, these operations may discard subtle but diagnostically relevant textural and spectral information inherent to ultrasound imaging. Domain‐specific pooling strategies that better retain spectral characteristics have shown value in medical ultrasound analysis. ⁴ We propose evaluating whether replacing conventional pooling with ultrasound‐tailored alternatives, such as Singular Pooling, could improve sensitivity to morphological details in B‐mode or vascular patterns in Doppler. Such a modification may enhance feature preservation and boost classification accuracy, especially for structurally complex lesions.

In summary, the work by Silva et al effectively demonstrates the potential of deep learning for HFUS‐based skin lesion classification. Building upon this foundation, integrating feature‐aware fusion mechanisms, targeted optimization for uncertain cases, and domain‐adapted network components represent clear pathways toward more robust and clinically applicable AI tools. Addressing these methodological aspects could further strengthen the reliability and deployment readiness of such diagnostic systems.

Data Availability Statement

Data sharing not applicable to this article as no datasets were generated or analyzed during the current study.