Implementation of an Artificial Intelligence-Based Double Read System in Capturing Pulmonary Nodule Discrepancy in CT Studies

Abstract

Studies show that up to 80% of all radiology errors are due to errors in perception. Early detection is critical for good outcomes in patients with primary lung cancer and lung metastasis. However, pulmonary nodules can be easily missed due to their small size. We prospectively applied a machine vision algorithm to CT studies containing lung parenchyma to detect pulmonary nodules, as well as a natural language processing algorithm to the text of the report to identify documentation of pulmonary nodules. Apparent discrepancies in perception – instances where a pulmonary nodule was not reported – were flagged for a secondary review by a radiologist. Four thousand and nine hundred studies were prospectively processed, of which 450 cases with potential discrepancies were detected. Preliminary manual analysis was performed to analyze the base error rate and to optimize thresholds for the machine vision and natural language processing algorithms, and 104 cases were flagged for final review. Of these 104 cases, 50 cases contained undocumented pulmonary nodules. Among these cases, 7 cases were classified as likely to be significant, where report addendums were done and the clinicians notified. We have successfully implemented an automated double read system to detect pulmonary nodule discrepancies, with minimal disruption to the radiology workflow and while keeping personal health information on-premises. This successful implementation demonstrates the viability of using an automated and secure radiology double-read system to improve patient safety in radiology workflows, either at a health system or an independent radiology practice.