Multimodal Skin Lesion Diagnosis via Dynamic Consensus and Self-Correcting Representation Learning.

Abstract

Multimodal skin-lesion classification requires integrating complementary clinical and dermoscopic images while addressing inter-branch disagreement and the semantic gap. We propose MDCL-SCRL, a modality-specific dual-stream framework using paired images and patient metadata. Multimodal Dynamic Consensus Learning (MDCL) derives detached sample-level weights from inter-branch prediction disagreement to modulate the consensus classification loss, whereas Self-Correcting Representation Learning (SCRL) uses concept scores as training-only auxiliary targets. Inference requires only the paired images and metadata and combines branch probabilities through a fixed normalized geometric mean. Under matched fixed-split five-fold cross-validation on MILK10k, MDCL-SCRL improved mean Macro-F1 by 1.59-7.14 percentage points and Macro-AUROC by 3.02-4.86 points over matched GlobalConcat baselines across CNN and Transformer backbones. Component and sensitivity analyses characterized individual contributions. On Derm7pt, dataset-specific retraining improved both metrics across two backbones, supporting additional-dataset reproducibility. Out-of-fold analyses associated lower MDCL weights with errors and showed correlations between SCRL predictions and semantic targets. These findings support disagreement-aware optimization and training-only semantic supervision, while prospective independent clinical validation remains necessary.