A novel multi-agent spatiotemporal fusion framework for intelligent skin cancer diagnosis

Skin cancer is one of the most common malignancies worldwide, and early-stage diagnosis remains challenging due to its morphological similarity to benign lesions. Most existing computer-aided diagnostic systems rely on single static images, overlooking temporal information that is critical for distinguishing progressive malignancy.

We propose a novel multi-agent spatiotemporal fusion framework to enhance diagnostic accuracy. The framework consists of three key components: (1) a spatial agent based on a convolutional neural network for high-fidelity static feature extraction; (2) a temporal agent employing gated recurrent units to model longitudinal lesion evolution; and (3) a collaboration agent that dynamically fuses spatial and temporal representations via an attention-based weighting strategy.

Experiments on large-scale public dermoscopic datasets showed that our method achieved an accuracy of 94.5%, an F1-score of 93.8%, and an AUC of 0.97-outperforming traditional machine learning models, CNN classifiers, and 3D-CNN baselines. Ablation studies further confirmed the critical contribution of temporal modeling and adaptive fusion, particularly in differentiating early melanoma from atypical nevi.

This work highlights the potential of spatiotemporal modeling to improve early skin cancer detection and provides a promising direction for AI-assisted diagnosis of other chronic diseases requiring longitudinal monitoring.