Deep Learning Classification of Usual Interstitial Pneumonia Predicts Outcomes

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Medicine
• Pulmonary Medicine

Background:

• Computed tomography (CT) is crucial for noninvasive usual interstitial pneumonia (UIP) diagnosis.
• Limitations in visual assessment of CT scans necessitate enhanced image analysis techniques.
• Improved diagnostic accuracy is vital for managing interstitial lung diseases.

Purpose of the Study:

• To develop an explainable deep learning algorithm using multiple instance learning (MIL) for UIP prediction from CT.
• To validate the MIL algorithm's performance in independent clinical cohorts.
• To assess the algorithm's ability to predict patient survival and lung function decline.

Main Methods:

• An MIL algorithm was trained on a pooled dataset (n=2,143) and validated in three independent cohorts (n=127, n=239, n=979).
• Performance was evaluated using receiver operating characteristic analysis with histologic UIP as ground truth.
• Cox proportional hazards and linear mixed-effects models analyzed associations between MIL predictions, mortality, and forced vital capacity (FVC) decline.

Main Results:

• The MIL algorithm demonstrated improved accuracy for histologic UIP classification compared to visual assessment in two cohorts (AUCs 0.77 and 0.79 vs. 0.65 and 0.71).
• MIL-UIP classifications were significant predictors of mortality in independent cohorts (unadjusted HRs 3.1 and 3.64, P < 0.001).
• Patients classified as UIP positive by MIL showed a significantly greater annual FVC decline (-88 ml/yr vs. -45 ml/yr, P < 0.01), adjusted for fibrosis extent.

Conclusions:

• Computerized assessment using MIL effectively identifies clinically significant UIP features on CT.
• This AI-driven approach can enhance confidence in radiologic assessments for interstitial lung disease.
• The MIL method holds potential for earlier and more precise diagnosis of UIP, improving patient management.