Introduction. Malignant skin neoplasms are one of the most pressing problems in modern healthcare, characterized by a steady increase in morbidity. Early diagnosis of initial forms of melanoma by primary care physicians who lack dermoscopy skills presents a particular challenge.
The aim of the study was to develop and validate a screening methodology using mobile dermoscopy and machine learning algorithms for the early differential diagnosis of skin neoplasms.
Materials and Methods. To train neural networks, a combined dataset was formed, including 24,765 dermoscopic images from the ISIC-2019 repository and 657 clinically verified images collected by the authors, taking into account the skin phototypes of the Russian population. A developed optical smartphone attachment module was used to collect proprietary data. The software part of the system is implemented in a cloud architecture using the Vision Transformer (ViT) deep learning model. The efficiency of two analysis modes was investigated: multiclass classification (8 classes) and cascade binary classification (sequential separation into melanocytic/non-melanocytic lesions and differentiation of melanoma/nevus).
Results. Experimental evaluation showed the advantage of the cascade strategy. The accuracy of the model at the critically important stage of differentiating melanoma and nevus was 0.964 (F-measure 0.951), which exceeds the indicators of the multiclass approach (Accuracy 0.932). During clinical approbation on a sample of more than 200 patients, 9 cases of melanoma and 6 cases of basal cell carcinoma were detected. Comparison of the system's results with the conclusions of expert oncologists demonstrated a diagnosis agreement in 89% of cases.
Conclusion. The proposed intelligent decision support system (CDSS) based on mobile dermoscopy ensures high diagnostic accuracy comparable to expert levels. The implementation of the methodology in primary care practice will increase cancer alertness, screening accessibility, and the efficiency of patient routing to specialized oncological institutions.
