Dermoscopy-dominant correction fusion for paired clinical and dermoscopic skin lesion image classification

Paired clinical and dermoscopic images provide complementary information for skin lesion classification. We developed Dermoscopy-Dominant Correction Fusion (DDCF), an asymmetric fusion framework that preserves dermoscopic features as the base representation and incorporates clinical information as a gated corrective signal. We performed a retrospective evaluation using nine diagnostic categories from MILK10k. Three stratified lesion-level partitions with a 7:1:2 training/validation/test ratio and fixed class-wise counts were evaluated for each of three backbones: Swin-Tiny, DenseNet121, and EfficientNet-B4. The correction coefficient was fixed separately at each evaluated value, and the checkpoint with the lowest validation loss was selected within each configuration before test evaluation. Macro-averaged one-vs-rest AUROC was the primary metric; balanced accuracy and macro-F1 were complementary metrics. DDCF and naive-sum fusion were compared using paired stratified bootstrap confidence intervals within each backbone and partition. At \(λ=0.7\) , mean test AUROCs for DDCF were \(0.9308 ± 0.0066\) , \(0.9099 ± 0.0017\) , and \(0.9084 ± 0.0037\) for Swin-Tiny, DenseNet121, and EfficientNet-B4, respectively. This coefficient had the highest observed mean test AUROC among the evaluated values for all three backbones. In the nine backbone-by-partition comparisons with naive-sum fusion, AUROC differences were positive; nominal 95% bootstrap intervals excluded zero in two comparisons. Macro-F1 differences were also positive in all nine comparisons. Swin-Tiny balanced accuracy was lower with DDCF in all three partitions (mean 0.5736 versus 0.6229 for naive-sum fusion). Class-wise improvements were particularly evident for BKL and BCC, but increased BCC sensitivity was accompanied by reduced specificity. No uniform benefit was demonstrated for melanoma or keratinocytic tumors as a whole. DDCF showed consistently positive numerical macro-AUROC differences relative to naive-sum fusion at the reported configuration, while nominal statistical support and classification effects varied across partitions and diagnostic categories. The results support further evaluation of modality-asymmetric fusion without establishing a universally optimal coefficient or general clinical superiority. Not applicable.