Adversarial and Correlation-Aware Data Augmentation Framework for Multi-Label Chest X-Ray Image Classification.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Deep learning excels in multi-label chest X-ray (CXR) classification.
• High-performance models typically require extensive, fully-annotated datasets.
• Limited annotations pose a significant challenge for practical CXR classification.

Purpose of the Study:

• To develop an effective framework for multi-label CXR image classification using limited annotations.
• To address the challenge of training high-performance models with scarce data.
• To introduce novel data augmentation techniques for improved model robustness.

Main Methods:

• Proposed an Adversarial and Correlation-Aware Data Augmentation (ACAA) framework.
• Utilized pseudo-labeling on weakly-augmented images.
• Employed Generalized Nesterov Iterative Fast Gradient Sign Method (GNI-FGSM) for image-level adversarial augmentation.
• Introduced feature-space adversarial augmentation and Batch-level Mamba with correlation regularization.

Main Results:

• The ACAA framework demonstrated effectiveness in limited-annotation scenarios.
• Experiments were conducted on large CXR datasets (CheXpert and MIMIC-CXR).
• The proposed methods significantly improved multi-label CXR classification performance.

Conclusions:

• The ACAA framework offers a viable solution for training robust multi-label CXR classification models with limited annotations.
• Adversarial augmentation and correlation-aware techniques are crucial for enhancing model generalization.
• This approach advances the practical application of AI in medical imaging diagnostics.