DETECT IPN: Real-World Experience with Automated Detection of Incidental Pulmonary Nodules in an All-Comer Population

Detection of incidental pulmonary nodules (IPNs), besides lung cancer screening (LCS) programs, is key in early lung cancer detection and reducing lung cancer mortality. Computer-aided IPN detection was boosted by artificial intelligence (AI), promising important technical support for the radiological workup. In this study, we retrospectively studied the precision of two AI-based, computer-aided IPN detection systems on routine computed tomography (CT) scans at three German radiological centers. Within the 1552 CT datasets included in this study, the two AI-based detection systems detected IPN of any size in ca. 70% of the cases with an average of 2 IPN per CT. IPN of relevant size categories ≥5 mm and ≥8 mm was detected in approximately 60% and 35% of the patients, respectively. Manual IPN review by experienced radiologists revealed a low precision of the detection results of only 20 to 70%, with a good comparability between both systems and high variability between the centers. Besides inter-reader variability and site-specific differences in defining IPNs, homogeneity of the mostly LCS-based AI training datasets, together with the high rate of contrast agent usage and comorbidities under real-world conditions, appear as likely causes. Since detection systems are tuned for high sensitivity, our study supports the general usability of such systems in clinical routines. However, the results emphasize the importance to incorporate more diverse real-world data into AI training datasets to further reduce false positive rates and the inherent danger of unnecessary follow-up interventions.