Multi-Label Local to Global Learning: A Novel Learning Paradigm for Chest X-Ray Abnormality Classification

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computer-Aided Diagnosis

Background:

• Deep neural networks (DNNs) excel at chest X-ray classification but often train all abnormalities simultaneously.
• Existing curriculum learning methods may not be optimal for disease diagnosis due to varying abnormality complexity.

Purpose of the Study:

• To propose a novel curriculum learning paradigm, multi-label local to global (ML-LGL), for improved multi-label chest X-ray classification.
• To enhance DNN training by incorporating clinical knowledge for abnormality prioritization.

Main Methods:

• The ML-LGL approach iteratively trains DNNs on progressively complex sets of abnormalities, starting from fewer (local) to more (global).
• Abnormality priority is determined using three novel clinical knowledge-leveraged selection functions.
• A dynamic loss function is employed during training on curated image sets.

Main Results:

• The ML-LGL paradigm demonstrated superior performance compared to baseline methods on PLCO, ChestX-ray14, and CheXpert datasets.
• The method achieved results comparable to state-of-the-art techniques.
• ML-LGL showed improved initial stability during DNN model training.

Conclusions:

• The proposed ML-LGL curriculum learning strategy offers a promising advancement for multi-label chest X-ray classification.
• This approach has potential applications in enhancing automated diagnostic tools for radiologists.