Abstract
Segmenting melanoma lesions accurately is harder than it might appear. Even in high-quality dermoscopic images, lesion edges can be gradual rather than sharp, color differences between lesion and healthy skin are sometimes minimal, and no two lesions look quite alike. These are not just technical inconveniences; they directly affect how reliably a computer-aided system can support a clinician's decision. In this work, we explored whether combining deep learning with swarm intelligence could address some of these difficulties. Specifically, we paired three CNN architectures, U-Net, SegNet, andDeepLabv3+ with ResNet-50, with the Grasshopper Optimization Algorithm (GOA), using it not as a preprocessing step but as a refinement stage applied after the network produces its initial prediction. The idea was that GOA's color-based clustering could tighten boundaries in regions where the CNN's confidence was low. We tested this hybrid model on the ISBI 2016 and ISIC 2018 dermoscopic datasets. Not all architectures responded equally U-Net struggled noticeably, while SegNet performed reasonably well. In previous approaches, deep learning and swarm intelligence were treated as separate stages, whereas our proposed framework, the GOA, is integrated directly into the post-prediction step of DeepLabv3+-ResNet-50, which enables boundary refinement thatcomplements rather than replaces CNN predictions. The results from DeepLabv3+ withResNet-50 were the best, achieving Dice coefficients of 91.7% and 88.2% and Jaccard indices of 84.9% and 79.7% on the ISBI 2016 and ISIC 2018 datasets, respectively, demonstrating the effectiveness of the proposed hybrid approach for lesion segmentation.