UNSUPERVISED DOMAIN ADAPTION WITH ADVERSARIAL LEARNING (UDAA) FOR EMPHYSEMA SUBTYPING ON CARDIAC CT SCANS: THE MESA...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Healthcare
• Pulmonology

Background:

• Emphysema quantification is established on full-lung high-resolution CT (HRCT) scans.
• Adapting these tools to cardiac CT scans is difficult due to lower resolution and degraded textural patterns.
• Cardiac CT scans cover approximately 70% of lung volume, making them valuable for population studies.

Purpose of the Study:

• To develop a deep learning domain-adaptation framework for emphysema labeling on cardiac CT scans.
• To leverage a pre-existing lung texture pattern (LTP) dictionary from HRCT scans for cardiac CT analysis.
• To enable robust, large-scale longitudinal studies of emphysema progression using cardiac CT data.

Main Methods:

• Proposed an original deep learning framework utilizing convolutional neural networks (CNNs).
• Employed supervised lung texture classification on synthetic cardiac images.
• Utilized adversarial learning to differentiate between real and synthetic cardiac images, combined with classification tasks.

Main Results:

• The framework successfully labeled emphysema regions on real cardiac CT scans from the MESA cohort (N = 15,357).
• Image features from adversarial training maintained labeling accuracy on synthetic scans.
• Lung texture pattern (LTP) histogram signatures from 4,315 longitudinal cardiac CT scan pairs showed high consistency over time and across scanner generations.

Conclusions:

• The proposed deep learning framework effectively adapts emphysema quantification tools to cardiac CT scans.
• This method enables robust labeling of emphysema texture patterns, overcoming image quality challenges.
• The approach facilitates large-scale, longitudinal studies for improved understanding of emphysema disease progression.