Challenges in the diagnosis of malignant skin neoplasms using artificial intelligence software and approaches to their resolution on the example of Derma Onko Melanoma Check software

Background. Artificial intelligence (AI)-based computer vision software is becoming an increasingly important tool in dermato-oncological practice. Deep learning models have demonstrated the ability to classify skin neoplasms with accuracy comparable to that of experienced dermatologists. However, their implementation in real clinical practice has revealed several significant challenges. Objective: To analyze and systematize the major challenges arising in the AI-assisted diagnosis of malignant skin neoplasms and to propose solutions to each identified challenge using the Derma Onko Melanoma Check software as an example. Material and methods. The clinical experience of applying AI software for the diagnosis of malignant skin neoplasms was analyzed, and six major categories of challenges were identified. Specific solutions to each of these challenges are proposed on the example of the Derma Onko Melanoma Check clinical decision support system (a multimodel AI-based computer vision system) for the preliminary diagnosis and triage of patients with skin neoplasms. Results. Six major challenges were identified and systematized: (1) imbalanced training datasets, with an overrepresentation of dermoscopic images and an insufficient number of clinical photographs; (2) poor quality of images acquired by physicians (e.g., overexposure, blur, and underexposure), limiting diagnostic reliability; (3) submission of images in which the lesion occupies an insufficient area of the frame for reliable analysis; (4) the single-model architecture of most AI systems and the resulting lack of adequate oncological alertness; (5) the absence of explainable artificial intelligence (XAI) in AI-generated reports; and (6) the inability of physicians to modify or refine the AI-generated conclusion using information obtained from the patient’s medical history and clinical examination. For each challenge, a specific solution implemented in Derma Onko Melanoma Check is proposed, including automated image quality assessment, lesion size verification, a multimodel ensemble of neural networks, XAI-based visualization of high-risk regions, interactive entry of clinical history data, and support for multimodal analysis. Conclusion. The implementation of AI systems incorporating the proposed solutions into routine clinical practice, particularly in primary care settings involving general practitioners, may significantly improve the early detection of malignant skin neoplasms by raising oncological vigilance for suspicious lesions and increasing the clinical interpretability and reliability of AI-generated diagnostic conclusions.