ABGF-Net: An Adaptive Bio-Contextual Graph Fusion Network for Interpretable Skin Lesion Classification

Manoj, R. Karthick (2026) ABGF-Net: An Adaptive Bio-Contextual Graph Fusion Network for Interpretable Skin Lesion Classification. Journal of Innovation and Technology, 2026 (12). pp. 107-114. ISSN 2805-5179

	Text joit2026_12.pdf - Published Version Available under License Creative Commons Attribution. Download (144kB)
	Text 885 - Published Version Available under License Creative Commons Attribution. Download (30kB)

Abstract

Skin disease detection using image-based deep learning has achieved significant progress; however, most existing approaches rely on convolutional neural networks (CNNs) or vision transformers (ViTs) that primarily capture pixel-level features while neglecting the underlying biological and structural characteristics of skin lesions. This limitation reduces interpretability and may affect generalization across diverse skin types and disease categories. To address these challenges, this paper proposes an Adaptive Bio-Contextual Graph Fusion Network (ABGF-Net), a novel framework that integrates structural lesion representation, biological priors, and visual contextual information for enhanced skin disease classification. The proposed method decomposes dermoscopic images into multiple lesion components, including pigment regions, texture patterns, and boundary structures, which are represented as nodes within a dynamically constructed graph. A bio-prior encoding module incorporates dermatological knowledge such as melanin distribution, vascular patterns, and lesion asymmetry, while a graph neural network performs relational reasoning over lesion structures. In parallel, a vision transformer extracts global contextual features, and both representations are integrated through a cross-attention fusion mechanism. The proposed model was evaluated on the HAM10000 and ISIC dermoscopic image datasets containing multiple skin lesion categories. Experimental results demonstrate that ABGF-Net achieved an accuracy of 94.8%, precision of 94.1%, recall of 93.7%, F1-score of 93.9%, and an AUC of 0.96, outperforming conventional CNN- and ViT-based approaches. Furthermore, attention heatmaps and Grad-CAM visualizations confirm that the model focuses on clinically relevant lesion regions, improving interpretability and diagnostic reliability. These results demonstrate that ABGF-Net provides an effective, robust, and biologically informed solution for automated skin disease detection and clinical decision support

Item Type:	Article
Uncontrolled Keywords:	Skin Disease Detection; Dermoscopic Image Analysis; Graph Neural Networks (GNN); Vision Transformer (ViT); Bio-Contextual Learning
Subjects:	Q Science > QA Mathematics > QA75 Electronic computers. Computer science R Medicine > RA Public aspects of medicine R Medicine > RL Dermatology T Technology > T Technology (General)
Depositing User:	Unnamed user with email masilah.mansor@newinti.edu.my
Date Deposited:	10 Jun 2026 08:39
Last Modified:	10 Jun 2026 08:39
URI:	http://eprints.intimal.edu.my/id/eprint/2328

Actions (login required)

View Item