Response to Letter: “Skin Cancer Detection Using Deep Learning Approaches” by Haque et al.

Dear Editor:

We were delighted to read the insightful review by Haque et al., recently published in Cancer Biotherapy & Radiopharmaceuticals, on the application of deep learning (DL) algorithms to skin cancer detection, which admirably surveys artificial neural networks (ANNs), convolutional neural networks (CNNs), k-nearest neighbors (KNNs), and generative adversarial networks (GANs) for lesion classification and highlights their promise in improving noninvasive diagnostic accuracy. ¹ We are living in an AI-fueled revolution, in which emergent, AI-driven disciplines such as the so-called “radiomics” promise to transform every facet of skin cancer management and, most critically, to upend the course of melanoma—the most prevalent and feared cutaneous malignancy. ²

In particular, the authors correctly emphasize the power of GANs for synthetic image generation to mitigate data scarcity; however, the “black box” nature of GANs, combined with documented challenges such as mode collapse, instability during training, and the risk of introducing subtle artifacts, raises important questions about the integrity and reproducibility of augmented datasets. Future studies would benefit from incorporating quantitative metrics to evaluate GAN outputs and ensure that generated images faithfully represent the diverse lesion morphologies encountered in clinical practice.

Moreover, while the review appropriately notes the utility of public repositories such as ISIC and PH2, these datasets are known to suffer from demographic and pathological biases. Predominantly populated by images of fair-skinned individuals and common lesion subtypes, they leave DL models vulnerable to diminished performance when applied to underrepresented groups (e.g., individuals with darker phototypes) or rare melanoma variants such as acral, mucosal, and amelanotic subtypes. ³ Curating inclusive, multicenter image collections that span a broader age range, skin phototypes, and histopathological categories—including nonmelanoma carcinomas common in elderly populations—will be key to addressing these gaps.

In parallel with these digital strategies, 3D total body photography has emerged as a robust tool for the early detection and ongoing monitoring of skin cancer. A 2025 prospective study by Peter et al. demonstrated the effectiveness of AI-assisted 3D dermoscopic imaging in tracking nearly 2800 melanocytic nevi during pregnancy, showcasing its ability to identify dynamic morphological changes indicative of malignancy. ⁴ Similarly, studies have shown that 3D imaging, when combined with CNNs, ensures highly reproducible nevus counts and consistent surveillance over time. Additionally, the clinical value of this approach has been underscored in population-based cohorts, with significant improvements in detection rates and excision outcomes. ⁵ These findings highlight the clinical importance of 3D imaging as a noninvasive, reproducible, and accurate method for early skin cancer diagnosis and long-term monitoring.

Finally, the path to clinical translation hinges on addressing challenges such as high computational demands, interpretability of DL models, and ethical considerations regarding data use. Efficient model architectures, prospective validation, explainable AI techniques, and transparent governance frameworks will be essential to build clinician trust and support regulatory approval. ⁶

Together, these technological advancements bring us closer to a future where AI-driven dermatology tools enable equitable, scalable, and life-saving skin cancer care—one pixel at a time.