Leveraging Color Space Transformations for Enhanced Skin Lesion Segmentation in Computational Dermatology Using Deep Learning
Main Article Content
Abstract
Accurate segmentation of skin lesions from dermoscopic imagery represents a foundational challenge in computational dermatology, with direct implications for the early detection and clinical management of melanoma and other cutaneous malignancies. Conventional deep learning segmentation frameworks predominantly operate within the standard RGB color space, which frequently conflates luminance and chrominance information, thereby limiting the discriminative capacity of learned feature representations in regions exhibiting subtle contrast variations or complex morphological boundaries.
This work presents a systematic investigation into the application of color space transformations—including HSV, LAB, and YCbCr encodings—as preprocessing and augmentation strategies to enhance the performance of deep convolutional segmentation architectures applied to dermoscopic skin lesion datasets. By decoupling chromatic and luminance channels, the proposed framework enables neural networks to extract more robust and semantically meaningful representations of lesion boundaries, texture gradients, and pigmentation distributions. A multi-branch encoder design is introduced that jointly processes complementary color representations, fusing information across spaces through learned attention mechanisms.
Extensive experiments conducted on the ISIC 2018 benchmark dataset demonstrate that integrating color space diversity into the segmentation pipeline yields statistically significant improvements over baseline models. The proposed approach achieves a mean Intersection over Union (mIoU) of $0.891$ and a Dice Similarity Coefficient (DSC) of $0.923$, surpassing contemporary single-space methods by margins of approximately $3.2\%$ and $2.7\%$, respectively.
These findings substantiate the hypothesis that color space heterogeneity constitutes a principled and computationally efficient inductive bias for dermoscopic segmentation, offering a complementary axis of improvement orthogonal to architectural scaling. The proposed methodology holds promise for integration into clinical decision-support systems requiring high-fidelity lesion delineation.