A Soft Labeling Approach to Develop Automated Algorithms that Incorporate Uncertainty in Pulmonary Opacification on...

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Medicine
• Computational Pathology

Background:

• Automated algorithm training relies on binary labels, which cannot capture uncertainty.
• Pulmonary opacification quantification on chest CT is crucial for diagnosis and monitoring.
• Existing labeling methods present challenges in handling inter-user variability and labeling burden.

Purpose of the Study:

• To propose and evaluate a soft labeling methodology for quantifying pulmonary opacification and percent well-aerated lung (%WAL) on chest CT.
• To address the limitations of hard labeling by incorporating uncertainty in segmentation.
• To reduce the overall labeling burden in medical image analysis.

Main Methods:

• Retrospective analysis of 760 COVID-19 chest CT scans from five international centers.
• Creation of pixel-wise labels for over 27,000 axial slices, classifying three opacification patterns.
• Quantification of %WAL and inter-user hard label variability using Shannon entropy.
• Implementation of a soft labeling and modeling cycle and comparison with hard labeling performance metrics.

Main Results:

• Hard labels showed significant inter-user variability, with only 3.37% pixel agreement among radiologists.
• The soft labeling approach significantly increased point-wise accuracy from 60.0% to 84.3% (p=0.01) for opacification prediction.
• The soft label model showed strong correlation for %WAL prediction (R=0.900), outperforming the hard label model (R=0.856).

Conclusions:

• Soft labeling enhances automated quantification and classification accuracy of pulmonary opacification on chest CT.
• The developed soft labeling methodology shows broad applicability for various pulmonary opacification contexts.
• This work provides a foundation for future advancements in medical image analysis using soft labeling techniques.