Lighting Enhancement and Skin Lesion Analysis in Macroscopic Images using Genetic Algorithms and Deep Neural Networks

Vanesa Gómez-Martínez, Cristina Soguero-Ruiz

Keywords: Illumination enhancement, skin lesion segmentation, macroscopic images, genetic algorithms, deep neural network

TL;DR: We propose a hybrid framework that combines unsupervised GA-based enhancement, U-Net segmentation, and ResNet-50 classification for macroscopic skin lesion analysis.

Abstract: Macroscopic skin lesion images, often captured with mobile cameras, are more accessible than dermoscopic images but suffer from poor lighting and reduced sharpness, which can affect lesion segmentation and diagnostic performance. This work presents a hybrid framework combining unsupervised image enhancement using Genetic Algorithms (GAs) with deep learning models for skin lesion segmentation (U-Net-based models) and melanoma classification (convolutional neural networks-based models). Experiments conducted on the Waterloo dataset demonstrate that GA-enhanced images improve segmentation performance across all U-Net variants, with the best dice score (0.871), outperforming original images (0.868) and IECET-enhanced images (0.851). For classification, the highest AUCROC (0.80) was achieved when using GA-enhanced and segmented images. Additionally, Grad-CAM analysis confirms that our pipeline enhances interpretability by guiding the model’s attention toward lesion areas, rather than irrelevant regions such as borders or dataset logos. These results show that simple yet effective preprocessing with GAs can enhance both performance and interpretability in skin lesion analysis using accessible macroscopic images.

Serve As Reviewer: ~Robert_Jenssen1

Submission Number: 14