Histological proven AI performance in the UKLS CT lung cancer screening study: Potential for workload reduction

Area of Science:

• Radiology
• Artificial Intelligence
• Oncology

Background:

• Lung cancer screening using computer tomography (CT) involves significant radiologist workload.
• Artificial intelligence (AI) offers potential for automating the interpretation of CT scans, particularly for identifying negative cases.
• Validation of AI performance against established standards is crucial for its clinical implementation in lung cancer screening.

Purpose of the Study:

• To validate a commercial AI software's performance in the UK lung cancer screening (UKLS) trial dataset.
• To compare AI performance against human readers and gold-standard histological lung cancer outcomes.
• To estimate the potential reduction in CT-reading workload achievable with AI as a first-reader.

Main Methods:

• 1252 UKLS baseline CT scans were independently assessed by AI and human readers.
• AI performance was evaluated against an EU reference standard and gold-standard histological outcomes.
• Misclassification rates were analyzed, and CT-reading workload reduction was calculated based on AI-identified negative scans.

Main Results:

• AI demonstrated a negative predictive value (NPV) of 92.0% compared to the reference standard, outperforming human reads in reducing misclassification.
• Against gold-standard histological outcomes, AI detected all 31 baseline lung cancers with an NPV of 99.8%, with only one false negative due to a size threshold.
• An estimated maximum CT-reading workload reduction of 79% was calculated.

Conclusions:

• AI shows significant potential as a first-reader in lung cancer screening to reduce CT interpretation workload.
• Implementing AI for ruling out negative CT scans does not appear to lead to missed lung cancers.
• AI's high accuracy and workload reduction potential support its role in optimizing lung cancer screening programs.