Trust-Aware Semi-Supervised Pseudo-Labeling and Multi-Modal Fusion for Skin Lesion Classification using GNNs

Abstract Accurate and trustworthy classification of skin lesions is critical for early diagnosis of melanoma. While existing deep learning models yield strong performance, they often lack transparency, underutilize metadata, and struggle with limited labeled data. This paper proposes a novel Trust-Aware Semi-Supervised Multi-Modal Graph Neural Network (GNN) that unifies three innovations: (1) a metadata-aware attention fusion mechanism to integrate deep visual features, handcrafted descriptors, and clinical metadata; (2) an entropy-based trust calibration strategy that adaptively segments predictions into high, medium, and low confidence levels; and (3) a trust-guided pseudo-labeling framework that leverages only high-trust samples for retraining. Extensive experiments conducted on the ISIC 2020 dataset demonstrate that the proposed method consistently outperforms conventional machine learning baselines and recent state-of-the-art approaches. After incorporating high-trust pseudo-labels, the model achieves a Test Accuracy of 98.27% and a Macro F1 score of 93.23%, confirming the effectiveness of the trust-aware learning strategy. The proposed framework offers a promising solution for enhancing skin lesion classification performance under limited labeled data conditions.