Deep hiearchical multi-label classification applied to chest X-ray abnormality taxonomies

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Chest X-rays (CXRs) are vital for diagnosis, driving research in computer-aided diagnosis (CAD).
• Usable CAD systems require high accuracy and predictions aligned with clinical knowledge.
• Existing hierarchical methods have limitations in performance and handling data complexities.

Purpose of the Study:

• To develop a deep hierarchical multi-label classification (HMLC) approach for enhanced CXR CAD.
• To improve model performance by strategically training with conditional and unconditional probabilities.
• To demonstrate the efficacy of HMLC in managing missing or incomplete labels in medical imaging.

Main Methods:

• Implemented a novel deep HMLC model for CXR abnormality detection.
• Employed a two-stage training strategy: initial conditional probability modeling followed by unconditional probability refinement.
• Developed a numerically stable cross-entropy loss function for unconditional probabilities.
• Evaluated on the Prostate, Lung, Colorectal and Ovarian (PLCO) and PadChest datasets.

Main Results:

• Achieved a mean AUC of 0.887 on the PLCO dataset, the highest reported for this dataset.
• Reported significant improvements on the PadChest dataset: 1.2% AUC and 4.1% average precision over flat classifiers.
• Demonstrated superior performance in handling incompletely labeled data compared to existing methods.

Conclusions:

• The proposed HMLC approach offers a significant advancement for CXR CAD systems.
• The method provides accurate, clinically relevant predictions and robust performance with incomplete data.
• This work represents a valuable step forward for the clinical application of AI in radiology.