Automated pneumothorax triaging in chest X-rays in the New Zealand population using deep-learning algorithms.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Pneumothorax detection on chest X-rays (CXRs) is critical for patient care.
• Automated triaging of CXRs can improve efficiency in clinical settings.

Purpose of the Study:

• To develop and evaluate artificial intelligence (AI) models, specifically convolutional neural networks (CNNs), for automated pneumothorax classification and segmentation on CXRs.
• To assess the interpretability and potential bias of the best-performing AI model.

Main Methods:

• Utilized the CANDID-PTX dataset (19,237 CXRs) for training and testing CNN models.
• Employed metrics such as AUC-ROC, sensitivity, specificity, and Dice coefficients for performance evaluation.
• Conducted interpretability analysis using Grad-CAM heatmaps and simulated a triage workflow.

Main Results:

• The best-performing model achieved high accuracy in pneumothorax classification (AUC-ROC 0.94, sensitivity 0.93, specificity 0.95).
• Segmentation performance showed a TP-Dice coefficient of 0.69.
• Triage simulation demonstrated a significant reduction in reporting delay (from 9.8 days to 1.0 day) with high classification accuracy (sensitivity 0.95, specificity 0.95).
• Interpretability analysis confirmed the model's logic aligns with radiologists' understanding, with minimal bias.

Conclusions:

• AI models can automate pneumothorax detection with clinically relevant accuracy.
• Implementation of these AI models as triaging tools has the potential to substantially decrease reporting delays for critical findings.