Introduction. Diagnostic errors between nevic lesions and skin tumors are frequent for non-specialist physicians. Therefore, there is a need to create a simple and practical tool based on artificial intelligence to assists them in distinguishing potentially malignant lesions from benign ones, improving early detection of skin cancer.
Objective. To evaluate the performance of SkinGuard, an algorithmic framework in prospectively collected images of skin lesions from patients at the Otamendi and Miroli Sanatorium and the MEDICUS Center (OTAMED test dataset).
Materials and methods. The images used for SkinGuard training were downloaded from the ISIC-workshop2 and Asan-Hallym Dataset websites. The ISIC-Asan dataset was composed of 60,837 images (benign: 42,077; malignant: 18,760) with histological confirmation of the following lesions: melanocytic nevi, benign keratoses, vascular lesions, dermatofibroma, intraepithelial carcinoma, basal cell carcinoma, squamous cell carcinoma, and melanoma. The ISIC-Asan dataset was divided into a training set (90.1%; n=54,830) and a test set (9.9%; n=6,007). OTAMED contained 233 images (161 benign and 72 malignant) obtained with cameras from different cell phones. SkinGuard is an assembly of two deep neural networks (EfficientNetV2S and MobilenetV2) preconfigured with ImageNet and trained on 2 GeForce RTX3090 GPUs. Once trained, the model was evaluated on both test sets.
Results. In the ISIC-Asan test set, the system obtained 91.2%, 91.4%, 90.8% and 87.0% for global accuracy, global sensitivity, global specificity and malignant class F1-score. More specifically, metrics for melanoma class were 96.0%, 93.0%, 97.0%, and 95.0% for accuracy, sensitivity, specificity and F1-score. In OTAMED, the system obtained 89.2%, 87.0%, 94.4% and 84.0% for global accuracy, global sensitivity, global specificity and malignant class F1-score. More specifically, metrics for melanoma class were 94.0%, 77.0%, 96.0%, and 78.0% for accuracy, sensitivity, specificity and F1-score.
Conclusion: SkinGuard has the ability to classify with high precision the different neoplastic skin lesions from images taken with a cell phone in a real clinical setting. This method could allow rapid and accurate assessments of these neoplastic lesions within a primary care visit, improving classification, timely referral to a dermatologist, and early treatment of the different forms of skin cancer.
