Beyond Transfer Learning: A Lightweight Convolutional Architecture for Dermatoscopic Skin Cancer Detection

Abstract

Skin cancer, specifically melanoma, represents a growing public health challenge in Chile, with a 116% increase in the mortality rate over the last two decades. Early detection is critical, but clinical diagnostic accuracy varies significantly, and access to specialists is limited. This work presents the development of an artificial intelligence model for clinical decision support based on deep learning for the automatic classification of skin lesions (benign vs. malignant). The performance of a custom-built Convolutional Neural Network (CNN) architecture, trained from scratch, was compared to a transfer learning model based on InceptionV1. The experimental results indicated that, while the transfer learning model achieved greater overall sensitivity, the proposed architecture attained superior accuracy (69.75% vs. 67.33%), demonstrating a greater capacity to reduce false positives. This validates the effectiveness of designing lightweight and specialized architectures, which achieve competitive and efficient performance without relying on massive pre-training, opening new avenues for the implementation of computer-assisted diagnostic tools in environments with limited computational resources.