DIFFERENTIABLE NEURAL ARCHITECTURE SEARCH FOR OPTIMAL CHANNEL SELECTION IN DERMOSCOPIC SKIN LESION SEGMENTATION

Accurate skin lesion segmentation from dermoscopic images is a fundamental prerequisite for early-stage melanoma detection. The diagnostic richness of intrinsic, shading-attenuated, and retinex-decomposed representations is overlooked by current deep learning techniques, which mostly rely on fixed RGB input channels. In this work, we offer a Differentiable Neural Architecture Search (DARTS) framework for a modified Transformer–CNN hybrid segmentation network (TMU-Net) that automatically selects the best input channel combinations based on gradients. The suggested method relaxes the channel-selection space to a continuous domain, allowing end-to-end joint optimization of both the segmentation network parameters and the channel weights, in contrast to discrete combinatorial search techniques like Genetic Algorithms. Tests on the ISIC 2016, ISIC 2017, and ISIC 2018 benchmark datasets show that DARTS-guided channel selection consistently performs better than GA-guided selection and fixed-channel baselines, with a Dice coefficient of up to 0.912 on ISIC 2018. These findings prove that differentiable architectural search is a scalable and efficient method for choosing input modalities in medical image segmentation.