Adaptive Color Space Selection for Deep Learning-Based Skin Lesion Segmentation in Dermoscopic Imaging

Abstract

Accurate segmentation of skin lesions in dermoscopic images remains a fundamental challenge in computer-aided diagnosis of melanoma and other cutaneous malignancies. Existing deep learning segmentation frameworks predominantly operate within the standard RGB color space, neglecting the potential discriminative advantages offered by alternative color representations such as HSV, LAB, and YCbCr. This oversight may limit model sensitivity to subtle chromatic and luminance variations that are diagnostically significant in dermoscopic analysis.

In this work, we propose an adaptive color space selection framework that systematically evaluates and integrates multiple color space representations to enhance the performance of deep learning-based skin lesion segmentation models. Our approach introduces a lightweight attention-guided fusion module that dynamically weights feature maps extracted from heterogeneous color spaces, enabling the network to exploit complementary spectral information without incurring prohibitive computational overhead. The framework is compatible with established encoder-decoder architectures, including U-Net and its variants, and can be integrated with minimal architectural modification.

Extensive experiments conducted on three publicly available benchmark dermoscopy datasets—ISIC 2017, ISIC 2018, and PH\textsuperscript{2}—demonstrate that adaptive multi-color-space fusion achieves statistically significant improvements over single-space RGB baselines across key segmentation metrics, including Dice similarity coefficient, Jaccard index, and sensitivity. Ablation studies confirm the individual contribution of each color representation and validate the efficacy of the proposed adaptive weighting mechanism.

These findings establish that thoughtful color space design constitutes a meaningful yet underexplored axis of optimization in dermoscopic segmentation pipelines. The proposed framework offers a computationally efficient, modality-agnostic strategy for improving lesion boundary delineation, with direct implications for clinical decision-support system development.