Harnessing deep learning to detect bronchiolitis obliterans syndrome from chest CT

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Pulmonary Medicine

Background:

• Bronchiolitis Obliterans Syndrome (BOS) is a significant post-lung transplant airway disease.
• Diagnosis conventionally relies on pulmonary function tests (PFTs) due to CT imaging limitations.
• Deep neural networks (DNNs) have not been previously applied for BOS detection in CT scans.

Purpose of the Study:

• To develop and train a DNN for detecting BOS in CT scans.
• To address the challenge of low-data scenarios in medical AI research.
• To improve early detection and management of BOS.

Main Methods:

• A DNN was trained using a co-training method optimized for low-data environments.
• An auxiliary task was incorporated to enhance disease sensitivity and reduce anatomical feature influence.
• The DNN predicted sequential CT scans from BOS patients, evaluated on 75 post-transplant patients (26 with BOS).

Main Results:

• The DNN achieved a ROC-AUC of 0.90 for BOS detection, correlating with disease stage (0.88-0.94).
• Comparable performance was observed on standard- and high-resolution CT scans.
• The DNN successfully predicted BOS in at-risk patients (ROC-AUC 0.87) and identified air-trapping/bronchiectasis.

Conclusions:

• The DNN approach shows promise for improved BOS diagnosis and early management.
• Detection from standard-resolution scans at any respiratory phase enhances accessibility.
• Overfitting mitigation techniques are vital for DNNs in low-data medical settings.