Enhancing Robustness in Gastroscopy Analysis With Poisson–Gaussian Noise-Based Adversarial Augmentation

Abstract

Gastroscopy is a key diagnostic procedure for gastric cancer, but its image quality is often degraded by complex noise from illumination, anatomy, and device variability, which can impair computer-aided diagnosis (CADx) systems. To address these challenges, we propose a Poisson-Gaussian noise-based Adversarial Augmentation (PGAA) method that simulates realistic photon-limited and sensor-readout noise through imperceptible perturbations, improving model robustness against clinically relevant noise. An accompanying Adaptive Weight Normalization (AWN) mechanism adaptively scales perturbations according to local noise sensitivity, preserving lesion boundaries and mucosal textures. To ensure a fair and comprehensive comparison, experiments were conducted using both single-step Fast Gradient Sign Method (FGSM) and iterative Projected Gradient Descent (PGD) adversarial perturbations, as well as across different backbone networks, including ConvNeXt V2-B. PGD was adopted as a theoretically consistent and stronger baseline to verify robustness beyond simple gradient-based attacks. Results on both in-house and public datasets demonstrated consistent improvements across recall, accuracy, F1-score, and AUC metrics. On the in-house dataset, the proposed PGAA-AWN achieved an AUC of 95.5%, with improvements of up to 2.4% in recall, 4.9% in accuracy, and 4.3% in F1-score over the baseline, while also showing the highest performance on the public Kvasir dataset. Although optimized for gastroscopy, the proposed frameworks are modality-agnostic and can be extended to other noise-prone medical imaging modalities such as CT, MRI, and X-ray, contributing to the development of robust and interpretable medical AI systems. © 2013 IEEE.