Deep learning for pneumothorax diagnosis: a systematic review and meta-analysis

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Radiology
• Diagnostic Performance Evaluation

Background:

• Deep learning (DL), a subset of artificial intelligence (AI), is increasingly used to assist in pneumothorax diagnosis.
• No prior meta-analysis has evaluated the diagnostic performance of DL for pneumothorax detection.

Approach:

• A comprehensive search of electronic databases was conducted up to September 2022 to identify relevant studies.
• Meta-analysis using a hierarchical model was performed to calculate pooled sensitivity, specificity, and area under the curve (AUC) for DL and physician diagnoses.
• The Prediction Model Study Risk of Bias Assessment Tool was adapted to assess the risk of bias in the included studies.

Key Points:

• DL models achieved a total AUC of 0.97, comparable to physicians.
• Pooled sensitivity for DL was 84% and for physicians was 85%.
• Pooled specificity for DL was 96% and for physicians was 98%.
• A significant proportion of studies (57%) exhibited a high risk of bias.

Conclusions:

• The diagnostic performance of DL models for pneumothorax is similar to that of human physicians.
• The high risk of bias in the majority of studies necessitates cautious interpretation of results.
• Further research is crucial to advance AI in pneumothorax diagnosis and address identified limitations.