Deep learning for classifying fibrotic lung disease on high-resolution computed tomography: a case-cohort study

Area of Science:

• Radiology
• Artificial Intelligence
• Pulmonology

Background:

• High-resolution CT (HRCT) is crucial for diagnosing fibrotic lung disease.
• Idiopathic pulmonary fibrosis (IPF) can be diagnosed via HRCT without biopsy if findings match usual interstitial pneumonia (UIP).
• International guidelines (ATS/ERS/JRS/ALAT, Fleischner Society) define HRCT criteria for IPF diagnosis.

Purpose of the Study:

• To evaluate a deep learning (DL) algorithm for automated classification of fibrotic lung disease on HRCT.
• To assess the algorithm's performance against international diagnostic guidelines and expert radiologists.
• To determine the algorithm's utility in centers with limited thoracic imaging expertise.

Main Methods:

• A dataset of 1157 anonymized HRCT scans with fibrotic lung disease was used for training, validation, and testing.
• The DL algorithm was trained using 2011 ATS/ERS/JRS/ALAT guidelines and subsequently retrained using Fleischner Society criteria.
• Algorithm performance was evaluated on test sets against the majority vote of specialist thoracic radiologists, measuring accuracy, prognostic accuracy, and interobserver agreement (κw).

Main Results:

• The DL algorithm achieved 76.4% accuracy on test set A and 73.3% on test set B, outperforming 66% of 91 radiologists.
• Interobserver agreement between the algorithm and expert consensus was good (κw=0.69), exceeding that of 62% of radiologists.
• The algorithm demonstrated comparable prognostic discrimination between UIP and non-UIP patterns versus expert radiologists.

Conclusions:

• A DL algorithm can classify fibrotic lung disease on HRCT with human-level accuracy.
• This AI approach offers a reproducible, near-instantaneous, and potentially low-cost method for disease classification.
• The technology could enhance diagnostic capabilities in resource-limited settings and aid clinical trial patient stratification.