Improved detection of air trapping on expiratory computed tomography using deep learning

Area of Science:

• Medical imaging analysis
• Artificial intelligence in healthcare
• Pulmonary medicine

Background:

• Air trapping (AT) on expiratory CT scans indicates early pulmonary dysfunction in cystic fibrosis (CF).
• Current quantitative AT assessment methods are highly variable and limited for disease monitoring.

Purpose of the Study:

• To evaluate a convolutional neural network (CNN) model for quantifying and monitoring AT.
• To compare the CNN model's performance against threshold-based quantitative AT measures.

Main Methods:

• A densely connected CNN was trained using AT segmentation maps from a personalized threshold-based method (PTM).
• Quantitative AT (QAT) from the CNN was compared to PTM QAT in 36 mild CF patients over 24 months.
• Radiographic, spirometric, and clinical data were correlated with CNN QAT values.

Main Results:

• CNN-derived QAT increased from 8.65% to 21.38% over two years, correlating with clinical scores.
• The CNN model showed a systematic drop in Dice coefficient compared to intensity-based measures over time.
• The CNN approach demonstrated greater robustness to variations in expiratory deflation levels than threshold-based methods.

Conclusions:

• The CNN model effectively delineates AT on expiratory CT scans.
• This provides an automated and objective method for assessing and monitoring AT in CF patients.
• The CNN approach shows promise for improved disease management and progression tracking.