Attention-Guided Artificial Intelligence Multi-Agent architecture for Explainable and resilient Skin Disease Classification

Main Article Content

Madhavan Babu, Prashant Kumar Yadav

Abstract

Artificial intelligence (AI) is gained considerable attention in dermatological image analysis because deep learning models identify visual patterns associated with the different skin diseases. However, variations in lesion appearance, skin tone, image quality, and disease similarity affect classification reliability. Conventional single-model architectures often struggle in combining the matching visual features, distinguish clinically similar conditions, and are providing consistent predictions beyond the heterogeneous dermatological images. This study is introducing an Attention-Guided Cooperative Multi-Agent Network (AGC-MAN) for skin disease classification. The proposed model is combining the four specialized agents responsible for global feature extraction, local lesion characterization, attention-based representation learning, and cooperative classification. A hierarchical attention mechanism is emphasizing discriminative lesion regions, while an adaptive agent coordination module is combining individual predictions through the confidence-weighted consensus learning. A feature consistency mechanism is reducing the disagreements among the agents. The architecture is employing the HAM10000 and ISIC dermatological image datasets for multiclass evaluation. The proposed framework is providing a combined classification strategy that is combining localized visual attention, collaborative inference, and interpretable decision-making for automated skin disease recognition. The comparative evaluation reports 96.10% accuracy, 95.60% precision, 95.20% recall, and 95.40% F1-score, with the macro-AUC of 0.974 at 25 training epochs. when it is compared with the EFAM-Net, accuracy is improving by 1.90 percentage points. The architecture is establishing a cooperative classification strategy for seven-category skin lesion recognition using HAM10000, with the ISIC-2019 providing an additional evaluation benchmark.

Article Details

Section
Articles