Learning to Quantify Emphysema Extent: What Labels Do We Need?

Area of Science:

• Pulmonary medicine
• Radiology
• Artificial Intelligence in Healthcare

Background:

• Accurate pulmonary emphysema assessment is vital for disease management and lung cancer risk prediction.
• Current visual assessment is time-consuming and has high inter-rater variability.
• Standard densitometry methods are less effective than visual scoring for emphysema quantification.

Purpose of the Study:

• To evaluate machine learning (ML) models for accurate emphysema extent estimation from CT scans.
• To compare ML models trained on emphysema presence versus extent labels.
• To determine if simpler presence labels can achieve performance comparable to extent labels.

Main Methods:

• Utilized four Multiple Instance Learning (MIL) classifiers trained on emphysema presence labels.
• Employed five Learning with Label Proportions (LLP) classifiers trained on emphysema extent labels.
• Evaluated classifier performance on 600 low-dose CT scans from the Danish Lung Cancer Screening Trial.

Main Results:

• ML models trained on emphysema presence labels achieved performance comparable to those trained on extent labels.
• The best MIL and LLP classifiers demonstrated high accuracy, with intra-class correlation coefficients around 0.90.
• Model agreement with human raters was high (78-79%), approaching inter-rater agreement (83%).

Conclusions:

• Machine learning models can accurately quantify emphysema extent from CT scans.
• Training ML models using readily available emphysema presence labels is as effective as using more complex extent labels.
• This approach offers a more efficient and reliable method for emphysema assessment in clinical practice.