A 3D-CNN model with CT-based parametric response mapping for classifying COPD subjects.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Pulmonology

Background:

• Chronic obstructive pulmonary disease (COPD) is characterized by lung parenchymal abnormalities.
• Current COPD assessment relies on pulmonary function tests and computed tomography (CT).
• Accurate classification of COPD severity and type is crucial for effective management.

Purpose of the Study:

• To introduce a novel deep learning-based classification method for COPD.
• To integrate parametric-response mapping (PRM) with a 3D convolutional neural network (CNN) for COPD grouping.
• To evaluate the performance of the proposed 3D-CNN and PRM (3D-cPRM) method in classifying COPD.

Main Methods:

• Extracted functional small airway disease percentage (fSAD%) and emphysema percentage (Emph%) using image registration.
• Utilized a 3D CNN model with fSAD% and Emph% as input parameters.
• Integrated the 3D CNN with the PRM method to create the 3D-cPRM classification system.
• Employed gradient-weighted class activation mapping (Grad-CAM) to visualize key discriminative regions.

Main Results:

• The 3D-cPRM achieved 89.3% classification accuracy and 88.3% sensitivity in five-fold cross-validation.
• The proposed 3D-cPRM outperformed 2D models and traditional 3D CNNs.
• Grad-CAM analysis identified class-discriminative regions in the upper and middle lung lobes, correlating with elevated fSAD% and Emph%.

Conclusions:

• The 3D-cPRM method effectively represents parenchymal abnormalities in COPD.
• The classification accuracy of 3D-cPRM is comparable to established 2D PRM models.
• This deep learning approach shows potential for enhancing CT-based COPD diagnosis.