Answers are grounded in this paper and cite their sources.
Teledermatology enables remote skin lesion assessment from dermoscopic images, but associated clinical metadata may be missing, inconsistently encoded, or clinically implausible. Conventional multimodal fusion often assumes that available metadata are reliable, which can cause negative transfer when structured side information is degraded. This paper proposes a reliability-gated multimodal fusion (RG-MF) framework that estimates sample-wise metadata reliability from availability and plausibility cues and attenuates metadata influence before mutual-attention fusion with image tokens. RG-MF preserves the Vision Transformer image encoder, soft-label metadata encoder, and mutual-attention fusion backbone for controlled comparability, while adding a lightweight reliability module. Experiments on HAM10000 used three independent seeds, lesion-level grouped splits, deterministic nested metadata missingness and corruption at rates of 0.25, 0.50, and 0.75, bootstrap confidence intervals, reliability-gate verification, gate-mode ablation, metadata-contribution diagnostics, targeted non-MCAR stress tests, and inference-cost measurement. RG-MF achieved mean macro-F1 scores of 0.6353, 0.6356, and 0.6352 in the clean, severe missingness, and severe corruption settings, compared with 0.6113, 0.6111, and 0.6110 for the prior fusion baseline. The image-only ViT baseline remained competitive, and diagnostics showed limited metadata contribution, indicating image-dominated behavior under the present protocol. Therefore, RG-MF should be interpreted primarily as a fail-safe, robustness-oriented metadata-control and diagnostic evaluation mechanism that reduces the influence of unreliable structured inputs rather than as conclusive dominance over image-only learning or a pure accuracy-enhancement module. The reliability score decreased as metadata quality deteriorated, and the module added only 145 trainable parameters.
No takes yet. Share an insight, caveat, or question.
Furqan et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: