Back

Deep Learning-Based Classification of Melanoma and Cutaneous Lesions Using NFNet Architecture: Development and Clinical Validation

Cavadia, D.; Castro, M.; Llull, M. O.

2025-11-17 dermatology
10.1101/2025.11.15.25340317 medRxiv
Show abstract

Melanoma remains the most lethal form of skin cancer, necessitating early detection for optimal patient outcomes. This study presents an advanced automated diagnostic system utilizing state-of-the-art convolutional neural networks--NFNet-L0, ResNeSt-101e, and MogaNet-XT--to classify nine types of cutaneous lesions from dermoscopic images. Leveraging a diverse dataset of 22,618 images from the International Skin Imaging Collaboration (ISIC) and Venezuelan clinical centers, the system demonstrates 96.2% accuracy in multi-class classification, outperforming conventional methods. Robust data augmentation and class-balancing strategies ensured balanced performance across all lesion categories. Clinical validation by five board-certified Venezuelan dermatologists confirmed the systems diagnostic accuracy, clinical relevance, and usability. The applications deployment as an accessible web interface highlights its potential for supporting dermatological diagnosis, particularly in resource-limited settings. Our findings underscore the promise of deep learning-driven tools for enhancing skin cancer detection and improving healthcare equity in Latin America and beyond.

Matching journals

The top 4 journals account for 50% of the predicted probability mass.

50% of probability mass above

"Similar papers" are the closest papers from that journal in the model's embedding space. They show what the match is built on, but the ranking comes mostly from a classifier over the whole training set, not from these examples alone.