CorLabelNet: a comprehensive framework for multi-label chest X-ray image classification with correlation guided...

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Deep Learning

Background:

• Deep learning significantly advances multi-label chest X-ray (CXR) classification for clinical diagnosis.
• Existing methods often fail to effectively learn or leverage label correlations and struggle with imbalanced CXR datasets, leading to biased models.

Purpose of the Study:

• To develop a framework that learns label correlations and uses them to guide feature learning and oversampling for improved multi-label CXR classification.
• To address the challenges of learning label correlations and data imbalance in CXR image classification.

Main Methods:

• Incorporated self-attention to capture high-order label correlations from global and local perspectives.
• Proposed a consistency constraint and multi-label contrastive loss to enhance feature learning.
• Developed an oversampling approach leveraging learned label correlations to identify critical seed samples for addressing data imbalance.

Main Results:

• Achieved state-of-the-art performance on the CheXpert and ChestX-Ray14 datasets through rigorous 5-fold cross-validation.
• Demonstrated the effectiveness of learning accurate label correlations for multi-label classification.
• Showcased the benefits of utilizing label correlations for discriminative feature learning and effective oversampling.

Conclusions:

• Learning and utilizing label correlations is crucial for advancing multi-label classification in medical imaging.
• The proposed framework effectively enhances feature learning and addresses data imbalance, leading to superior performance in CXR analysis.
• The methods presented offer significant improvements over existing state-of-the-art approaches in clinical diagnosis using CXR images.