Multimodal classification of skin lesions based on lightweight gated low-rank hyper-fusion network

Skin cancer imposes a substantial global public health burden, and early, non-invasive, and accurate lesion classification is a critical prerequisite for improving patients' long-term survival outcomes. In routine dermatological practice, multimodal integration of dermoscopic images and clinical metadata (including patient age, sex, and lesion anatomical location) has been widely adopted. Nevertheless, current state-of-the-art multimodal fusion frameworks still encounter a set of unresolved critical bottlenecks in real-world clinical deployment. These unaddressed limitations primarily include cross-modal negative transfer, overfitting under limited annotated sample regimes, insufficient inter-modal feature interaction, and the “black-box” nature of model decision-making, all of which collectively degrade the operational robustness and real-world clinical deployment feasibility of existing multimodal diagnostic frameworks. To systematically address these unmet challenges, this work presents a Gated Low-Rank Hyper-Fusion Network (GLHF-Net) for automated skin lesion classification. The proposed architecture leverages a lightweight Swin Transformer variant as its visual feature extraction backbone, built around a two-stage dynamic multimodal fusion pipeline. First, a Metadata-Driven Channel-wise Gating Mechanism (Meta-Gate) adaptively suppresses task-irrelevant channel-level noise in high-dimensional visual feature maps, mitigating cross-modality distribution discrepancy between tabular clinical metadata and latent image embeddings. Second, an embedded Low-Rank Hyper-Dynamic Fusion (LR-HDF) module executes constrained deep feature reconstruction under explicit low-rank regularization, effectively reducing empirical overfitting risk on small annotated dermatological datasets. Extensive validation experiments carried out on the public benchmark dataset HAM10000 demonstrate that the proposed GLHF-Net model achieves robust diagnostic accuracy with an AUC of 0.9200, a pAUC of 0.8405, a Sensitivity of 0.8935 and an NPV of 0.9682, and demonstrates highly competitive performance compared to both conventional feature concatenation methods and methods based on complex attention mechanisms. Grad-CAM visualization results confirm that the proposed model can accurately localize core lesion regions while effectively suppressing interference artifacts such as occluding hair, which provides crucial interpretable evidence for the clinical reliability of the model. With strong anti-overfitting ability and high inference efficiency (121.1 FPS on GPU and 20.35 FPS on CPU), this framework provides a reliable lightweight solution for the deployment of intelligent skin cancer screening in clinical scenarios.