PulmoClass-3DAtt: A Self-Attention Network for Classification of COPD, PRISm and Normal

Area of Science:

• Pulmonary Medicine
• Radiology
• Artificial Intelligence

Background:

• Chronic obstructive pulmonary disease (COPD) is a leading cause of death globally, necessitating accurate diagnostic methods.
• Distinguishing COPD from preserved ratio impaired spirometry (PRISm) is challenging due to shared small-airway pathologies.
• Accurate differentiation is crucial for effective patient management and treatment.

Purpose of the Study:

• To develop and validate a novel deep-learning framework for discriminating between COPD, PRISm, and normal lung function.
• To integrate chest CT imaging features with clinical variables for enhanced diagnostic accuracy.
• To provide a tool for clinical decision support in pulmonary medicine.

Main Methods:

• A retrospective study enrolled 1918 participants, with data split into training (80%) and validation (20%) cohorts.
• A convolutional encoder extracted imaging features from chest CT scans, predicting spirometric parameters.
• The PulmoClass-3DAtt model, utilizing self-attention, classified disease status and predicted spirometric metrics.

Main Results:

• The PulmoClass-3DAtt model achieved high overall performance with an AUC of 0.86 and accuracy of 0.87.
• Specific classification accuracies for COPD, normal, and PRISm were 0.87, 0.88, and 0.87, respectively.
• Spirometric metrics were predicted with Mean Squared Error ranging from 0.03-0.35 and Concordance Correlation Coefficient from 0.56-0.83.

Conclusions:

• The PulmoClass-3DAtt framework effectively integrates multimodal imaging and functional data using a self-attention mechanism.
• This deep-learning model reliably differentiates between COPD, PRISm, and normal lung status.
• The tool offers immediate applicability for clinical decision support and advancing precision pulmonary medicine.