Multi-Label Chest X-Ray Image Classification With Single Positive Labels.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Deep learning for Chest X-ray (CXR) classification requires large, fully annotated datasets, which are difficult to obtain.
• Existing methods struggle with noisy labels and the cost of data acquisition.
• Weakly supervised learning, specifically Single Positive Multi-label Learning (SPML), offers a solution by annotating only one positive label per image.

Purpose of the Study:

• To address the challenges of SPML in CXR image classification (SPML-CXR).
• To propose a novel Multi-level Pseudo-label Consistency (MPC) framework to improve classification accuracy under weak supervision.
• To mitigate the issue of false negative labels introduced by simple SPML solutions.

Main Methods:

• Developed a weak-to-strong consistency framework using pseudo-labeling and consistency regularization.
• Introduced Image-level Perturbation-based Consistency (IPC) with Random Elastic Deformation (RED) to recover mislabeled positive labels.
• Incorporated Feature-level Perturbation-based Consistency (FPC) and Batch-level Transformer-based Correlation (BTC) regularization for expanded perturbation and sample relationship exploration.

Main Results:

• The proposed MPC framework demonstrated significant effectiveness in SPML-CXR tasks.
• Experiments on CheXpert and MIMIC-CXR datasets validated the performance improvements.
• The framework successfully handled the challenges of limited annotations and noisy labels in CXR classification.

Conclusions:

• The MPC framework offers a robust solution for weakly supervised multi-label classification of CXR images.
• This approach reduces the reliance on large, fully annotated datasets, making medical image analysis more accessible.
• The study highlights the potential of advanced deep learning techniques in improving diagnostic accuracy from medical imaging.