Genetic Algorithm Guided Image Channel Selection for Skin Lesion Segmentation

Abstract

Skin cancer stands as the most widespread form of cancer globally, and early detection significantly enhances its treatability. Even though deep learning techniques have greatly improved the precision of segmentation, there is still potential for enhancement by tackling substantial challenges like the variability in lesion sizes, colors, shapes, and differences in contrast levels. This paper introduces an innovative approach to Genetic Algorithm-guided feature selection in skin lesion segmentation. The Modified TMU-Net (Transformer Meets U-Net) architecture overcomes limitations by replacing the ResNet block with a custom block for improved adaptability to variable input sizes. The dual-pipeline design integrates transformers for global representations and spatial dependency with convolutional neural networks for local contextual representations. A Genetic Algorithm (GA) is employed alongside the architecture to optimize image channel selection for enhanced segmentation. The GA iteratively refines solutions encoded as binary vectors, representing combinations of image types. This framework combines the original RGB images with data derived from the principles of skin illumination and imaging. Our approach involves incorporating data from distinct color bands, grayscale images immune to variations in illumination, and images with reduced shading effects. The combination of $ R_{n} $ , GRAY, SA, and RGB features produces qualitatively superior results compared to the RGB images. The proposed model was trained on the ISIC 2017 publicly available dataset and achieved successful outcomes.

Keywords

• Transformer
• semantic segmentation
• genetic algorithm
• dermoscopic images
• attention