Random walk-based erasure for medical imaging: augmenting data while preserving diagnostic features

Abstract

This study leverages a revised single‐label CXR14 dataset comprising 91,324 chest radiographs and addresses the dual challenges of limited data availability and model generalizability in medical imaging by means of novel data‐augmentation techniques. In the realm of image processing, we focus primarily on a random‐walk‐based erasure (RWE) algorithm, which we have specifically optimized and simplified for the CXR14 dataset. This approach achieves a controlled occlusion–random feature exploration trade‐off on chest X-ray images, thereby preserving diagnostically critical structures while bolstering model robustness.

The experimental framework employs DenseNet121 with focal loss optimization, incorporating three key innovations: 1) Dual-phase training combining augmented and original data 2) AUC-aware early stopping function for class-imbalanced learning 3) Multi-scale image quality assessment (IQA) using both reference-based (PSNR, SSIM, MS-SSIM) and no-reference metrics (BRISQUE, NIQE). The results demonstrate the superiority of RWE. Compared to the baseline original dataset, the RWE-enhanced dataset achieved a relative improvement of 3.24% in AUC on the full dataset (91,324 images), while the noise-injected dataset showed a relative enhancement of 1.35%. On the smaller dataset (10,000 images), the improvements were 12.5% and 11.7%, respectively. IQA analysis reveals RWE's unique advantage: while trailing in pixel-level fidelity, it better preserves structural semantics critical for diagnosis.

Our investigation uncovers two critical phenomena: 1) Augmentation efficacy inversely correlates with dataset size, suggesting strongest benefits for data-scarce scenarios 2) Randomization strategies significantly impact model stability. The study further proposes a parameter optimization pipeline for medical augmentation, identifying optimal RWE configurations through four-stage hyperparameter tuning.

These findings position RWE as a promising solution for medical AI's data challenges, though future work must address multi-label generalization and anatomical-prior integration. The proposed methodology provides a template for developing domain-specific augmentation strategies that balance physical realism with diagnostic utility, ultimately supporting more reliable and interpretable medical image analysis systems.