Disagreement-Aware Adaptive Boundary Fusion for skin lesion segmentation

Accurate segmentation of skin lesions in dermoscopic images is important for reproducible morphological assessment, but low contrast, artifacts, and ambiguous margins remain challenging. We present Disagreement-Aware Adaptive Boundary Fusion (DAABF), in which a lightweight FusionNet combines the probability maps of a CNN-based U-Net and a Transformer-U-Net with the original RGB image and their pixelwise absolute difference. To prevent lesion leakage, HAM10000 was partitioned by lesion identifier into 7,013 training, 1,501 validation, and 1,501 test images; FusionNet was trained using five-fold lesion-grouped out-of-fold expert predictions. On the held-out HAM10000 test set, DAABF achieved a Dice score of 0.9296, an IoU of 0.8826, and an HD95 of 15.54 pixels in the resized 256 × 256 image space. Relative to simple probability averaging (Dice 0.9284, IoU 0.8806, HD95 16.33), the differences were statistically detectable after Holm correction but modest in magnitude. Stronger single-model baselines, particularly SegFormer, U-Net++, and DeepLabV3+, obtained higher absolute overlap scores. In cross-dataset evaluation without dataset-specific retraining, DAABF achieved Dice scores of 0.9130 on PH2 and 0.8223 on ISIC 2018. Absolute disagreement was an interpretable spatial cue, but the experiments did not establish that it was superior to entropy or solely responsible for the fusion gain.