← Back to Publications List

An Explainable and Uncertainty-Aware Lightweight Dynamic Attention-Based Deep Learning Architecture for COVID-19 Detection from Chest X-Rays

Students & Supervisors

Student Authors
Moinul Hossain
Bachelor of Science in Computer Science & Engineering, FST
Mohammad Sakib Mahmood
Bachelor of Science in Electrical & Electronic Engineering, FE
Supervisors
Mohammad Alif Arman
Assistant Professor, Faculty, FE
Md. Ashiquzzaman
Associate Professor, Faculty, FE
Nafiz Ahmed Chisty
Associate Professor, Head [undergraduate Program], FE
Muhibul Haque Bhuyan
Professor, Faculty, FE

Abstract

Automated COVID-19 detection from chest X-rays faces four barriers: dataset bias, limited cross-distribution generalization, absent uncertainty quantification, and computational overhead. RobustXAI-COVID addresses all four through five contributions: (i) DAM—a per-sample gating scalar α adaptively fuses channel and spatial attention, surpassing fixed CBAM by +1.10%; (ii) CDRT — weighted joint training on COVIDx CXR-4 and ChestX-ray14 yields +0.68% at zero inference overhead; (iii) UE — Shannon entropy deferral reduces misclassification by 45% on flagged samples; (iv) EGT (core XAI contribution) — Grad-CAM++ alignment loss makes anatomically faithful explanations intrinsic, jointly improving accuracy (+0.50%) and heatmap fidelity; (v) KD from ResNet-18 with hybrid CE+Focal loss achieves +1.32%. On COVIDx CXR-4 (n = 1,000): 95.40% accuracy, 97.00% COVID-19 sensitivity, 0.9531 F1, 0.9901 AUC at 5.6M parameters and 12.1 ms CPU latency.

Keywords

COVID-19 detection chest X-ray dynamic attention cross-dataset robustness uncertainty estimation explainable AI knowledge distillation edge deployment

Publication Details

  • Type of Publication:
  • Conference Name: IEEE Region 10 TENSYMP 2026
  • Date of Conference: 29/06/2026 - 29/06/2026
  • Venue: Penang, Malaysia
  • Organizer: IEEE Malaysia Section
View on Publisher