Feature-Based Parametric Response Mapping on Thoracic Computed Tomography for Robust Disease Classification in COPD.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Pulmonary Medicine

Background:

• Chronic Obstructive Pulmonary Disease (COPD) phenotyping relies on CT scans to identify emphysema and functional small airways disease (fSAD).
• Traditional methods face challenges with noise and interpretability, potentially impacting diagnostic accuracy and multi-center study reliability.

Purpose of the Study:

• To develop an interpretable, feature-based Deep Parametric Response Mapping (PRM D) method for precise spatial detection and quantification of emphysema and fSAD.
• To enhance noise robustness in CT scan analysis for improved COPD phenotyping.

Main Methods:

• A deep learning approach combining wavelet scattering convolution networks and machine learning was developed using paired inspiratory-expiratory CT scans from 8,972 participants.
• PRM D extracts translation-invariant features and uses subspace learning for voxel-level classification of emphysema and fSAD.
• The method was validated against conventional PRM, assessing voxel-wise agreement, correlation with pulmonary function, and noise robustness.

Main Results:

• PRM D achieved 95% voxel-wise agreement with standard PRM (r = 0.98) and demonstrated superior noise robustness.
• PRM D showed stronger correlations with FEV1 for both emphysema (r = -0.54) and fSAD (r = -0.51) compared to standard PRM (r = -0.42).
• Under high noise, standard PRM overestimated disease by ~15%, while PRM D limited the error to <5%.

Conclusions:

• PRM D offers an interpretable, feature-driven, and noise-resilient alternative to traditional PRM for classifying emphysema and fSAD.
• This method enhances the reliability of CT-based COPD phenotyping, particularly for multi-center studies and low-dose imaging applications.