Prediction of melanoma skin cancer using convolutional neural networks implemented in a mobile application

Cinthya Y.-S.; Alexis G.-S.; Jorge M.-M.; Angel M.-M.; Jose S.-T. · Proceedings of the 2025 IEEE 5th International Conference on Advanced Learning Technologies on Education and Research, ICALTER 2025 (2025) · DOI: 10.1109/icalter69698.2025.11355137

Resumen

Abstract:Melanoma, a skin cancer with an increasing incidence, can be fatal if not detected early; convolutional neural networks (CNNs) offer high accuracy in detection, but still face challenges. This applied study uses MobileNetV2 architecture to develop a model for implementation in a mobile application, designed as a support tool for dermatologists in the early detection of melanoma using dermatological images. The study population comprised patients aged 1 8 or older at Goyeneche Hospital in Arequipa, with a sample of 129 participants. The CRISP-DM methodology was employed to develop the convolutional neural network model, utilizing TensorFlow and Google Colab. Given challenges such as limited datasets and class imbalance, the images were processed using Keras' ImageDataGenerator, which was configured solely for pixel rescaling (1 / 255) to standardize the input data. No additional image transformations were applied. This preprocessing ensured consistent input quality while maintaining the integrity of the original dermatological images. The SCRUM methodology was used for application development. As a result, an accessible mobile application was developed, achieving an overall accuracy of 89.9%. The model achieved 84.2% sensitivity for melanoma detection and 95.4% specificity for correctly identifying benign lesions. These performance metrics reflect a strong ability to distinguish between malignant and benign tumors, with a melanoma-specific accuracy of 93%, indicating reliable detection of malignant cases.

Instituciones: Universidad Católica de Santa María