HEDS-Net: Minimax-Based Segmentation-Guided Melanoma Detection from Dermoscopic Images

Abstract

Hair, ruler marks and marker ink corrupt lesion borders in dermoscopic images and mislead both segmentation and classification, while malignant lesions are scarce in public archives. This paper presents HEDS-Net (Hilbert Edge, Difference-convolution and Structure-aware Network), a three-stage pipeline for melanoma detection. First, a Hilbert–Sobel artefact removal (HSAR) module detects thin artefacts from the difference between the directional analytic-signal amplitudes of the image and of an oriented-closing reference, inpaints them, restores colour and contrast, and outputs a Hilbert edge prior. Second, a difference-convolution re-parameterised segmentation decoder (DRS-Decoder) uses tri-branch difference convolutions and an edge-guided dispersion gate to delineate low-contrast borders, and folds every multi-branch kernel into a single 3×3 convolution at inference. Third, a knowledge-guided classifier fuses ABCD shape, colour and texture descriptors computed from the predicted mask with deep features of a two-phase fine-tuned ResNet-50. On ISIC-2018, HEDS-Net reaches a lesion Dice of 91.84%. On ISIC-2020 it attains 99.21% accuracy under the protocol of the SegFusion base framework (99.01%) and 92.31% under a leakage-free lesion-level protocol, where the re-run base framework obtains 89.74%.