Citation

Aravinda, C. V. and Emerson Raja, Joseph and Alasmari, Sultan (2025) A Neural Architecture Search-Driven Quantum Patch Attention Framework for Skin Disease Recognition and Classification With XAI Vision Transformers. IEEE Access, 13. pp. 197312-197328. ISSN 2169-3536

Text A_Neural_Architecture_Search-Driven_Quantum_Patch_Attention_Framework_for_Skin_Disease_Recognition_and_Classification_With_XAI_Vision_Transformers.pdf - Published Version Restricted to Repository staff only Download (2MB)

Abstract

Accurate early differentiation of morphologically similar viral exanthems, such as chickenpox, measles, and monkeypox, remains a significant diagnostic challenge in dermatology. To address this, we present Quantum Patch Attention Vision Transformer guided by Neural Architecture Search (QPAViT-NAS), a hybrid quantum-classical deep learning framework, designed to enhance both accuracy and transparency in skin disease classification. The architecture integrates a Vision Transformer (ViT) backbone with a quantum-inspired feature encoding layer, leveraging quantum superposition principles to amplify subtle textural differences in dermoscopic images. A key innovation lies in the NAS-optimized patch attention module, which dynamically prioritizes diagnostically salient skin regions during classification, enabling fine-grained decision-making at the sub-image level. Trained on a curated dataset of 4,800 high-resolution dermoscopic images spanning four classes (normal skin, chickenpox, measles, monkeypox), the model achieved 90.2% accuracy (±1.3%) under 5-fold cross-validation, while reaching 84.62% accuracy (95% CI: [83.8, 85.4]) on the independent held-out test set. The framework also attained superior sensitivity (89.7%) and specificity (91.5%) in distinguishing measles from monkeypox, a pair notorious for clinical confusion. To ensure translational relevance, we incorporate Explainable AI (XAI) tools (Grad-CAM, SHAP, LIME) to visualize attention maps and validate model alignment with dermatological expertise. Additionally, ResNet-50-based embedding analysis quantifies inter-class visual similarity, revealing latent decision boundaries that correlate with pathological features (e.g., vesicle morphology, erythema patterns). The framework’s computational efficiency (12 ms/inference on Kaggle GPU) supports integration into point-of-care workflows, while its modular design allows adaptation to emerging skin conditions. By bridging quantum machine learning with interpretable vision systems, QPAViT-NAS advances the paradigm of “glass-box” AI in healthcare, offering precision and diagnostic insight. These findings underscore the potential of hybrid architectures to transform teledermatology and AI-augmented clinical decision making.

Item Type:	Article
Uncontrolled Keywords:	Clinical decision support, dermatoscopic imaging, explainable AI, neural architecture search, patch-wise attention, quantum machine learning, Viral skin diseases
Subjects:	R Medicine > RC Internal medicine > RC71-78.7 Examination. Diagnosis
Divisions:	Faculty of Engineering and Technology (FET)
Depositing User:	Nor Afiqah Mohd Adnan
Date Deposited:	02 Dec 2025 06:01
Last Modified:	12 Dec 2025 13:55
URII:	http://shdl.mmu.edu.my/id/eprint/14934

Downloads

Edit (login required)