RAFI: Real-Artifact Injection for Robust GI Endoscopy Classification

Deep learning models for gastrointestinal (GI) endoscopy classification achieve high accuracy on clean benchmark images but degrade sharply on real clinical images corrupted by artifacts such as blood, mucus, specular glare, and motion blur. This work introduces RAFI (Real-Artifact Injection), a training-time method that pastes real artifact patches — mined directly from the dataset itself — onto clean images, teaching the model the true distribution of clinical degradation rather than relying on synthetic approximations. Using HyperKvasir (10,662 labelled images; Normal vs. Abnormal), and a strictly controlled single-variable ablation across five strategies with a frozen EfficientNet-B0 backbone (2,562 trainable parameters), RAFI reduces the artifact-induced F1 drop from 0.046 to 0.007 (~85% smaller) with no loss of clean-image accuracy, and improves messy-image F1 by 4.3 percentage points — a uniform gain across blood, glare, mucus, and full-frame degradation. In contrast, conventional synthetic augmentation worsens robustness, empirically demonstrating that real artifacts, not simulated ones, are necessary for genuine robustness. The work also introduces the Lesion Attention Ratio (LAR), a quantitative explainability metric measured against ground-truth lesion masks, showing that an Artifact-Aware Attention Module nearly doubles lesion-localization quality even without improving raw accuracy — an honest ablation result reported alongside the main findings. The full pipeline trains in minutes on a single consumer laptop GPU and runs on CPU at inference, making it reproducible outside institutional compute budgets. This is a research prototype intended for academic study and is not a medical device; it has not been validated for, and must not be used for, clinical diagnosis or patient care.