Deep learning algorithm for surveillance of pneumothorax after lung biopsy: a multicenter diagnostic cohort study

Area of Science:

• Radiology
• Artificial Intelligence
• Pulmonary Medicine

Background:

• Pneumothorax is a frequent and serious complication of percutaneous lung biopsy.
• Accurate and timely detection of post-biopsy pneumothorax is crucial for patient management.

Purpose of the Study:

• To evaluate the performance of a deep learning algorithm in detecting pneumothorax on chest radiographs (CRs) following percutaneous lung biopsy.
• To compare the algorithm's performance against clinical radiology reports and human radiologists.

Main Methods:

• Retrospective analysis of 1757 post-biopsy CRs from three institutions.
• Commercial deep learning algorithm used for pneumothorax detection.
• Comparison with radiology reports and a reader study involving four radiologists.

Main Results:

• The deep learning algorithm achieved an area under the receiver operating characteristic curve (AUROC) of 0.937, with 70.5% sensitivity and 97.7% specificity.
• The algorithm demonstrated higher sensitivity but lower specificity compared to initial radiology reports.
• In reader studies, the algorithm showed comparable specificity but lower sensitivity than experienced radiologists.

Conclusions:

• The deep learning algorithm accurately identifies pneumothorax in post-biopsy CRs within real-world clinical cohorts.
• This AI tool holds promise as a surveillance aid for the accurate and timely diagnosis of post-biopsy pneumothorax.