Approach for Skin Cancer Detection Based on Dermatoscopic Images Applying YOLOv5, YOLOv8, MobileNetV2 & ResNet50

Sanchez J.-P.; Vivanco S.; Wong L. · 2026 9th International Conference on Artificial Intelligence and Big Data, ICAIBD 2026 (2026) · DOI: 10.1109/icaibd69640.2026.11637275

Resumen

Melanoma, though just 1.9% of Peru's skin cancer cases, poses a significant public health challenge due to its high mortality and barriers to early diagnosis. Given this context, this study proposes an approach for skin cancer classification using YOLOv8, YOLOv5, MobilNetV2 and ResNet50 models applied to dermatoscopic images. The approach was divided into four phases: (i) dataset acquisition, (ii) data preprocessing, (iii) models training and (iv) performance evaluation. The models focused on the use of a combination of three public ISIC datasets, which were subsequently augmented, resulting in a total of 7,216 dermatoscopic images classified into four classes: 'benign' (0), 'squamous cell carcinoma' (1), 'basal cell carcinoma' (2) and 'melanoma' (3). The training results showed that YOLOv8 and MobileNetV2 models had similar performance, achieving an average accuracy of 92.07% and 92.34%, respectively, with MobileNetV2 being the model with the most balanced results in the detection of 'benign','squamous cell carcinoma', 'basal cell carcinoma' and 'melanoma'. The main objective of this approach is to find a model that combines accuracy with the lowest possible computational resources so that it can be used for the early detection of skin cancer in resource-limited scenery.

Instituciones: Universidad Peruana de Ciencias Aplicadas