Comparison of two AI-based software tools for detection of Incidental Pulmonary Nodules (IPN) in a University Hospital and a Radiology Practice

<bold>Objective:</bold> To compare the performance of two AI-based software tools for the detection of IPNs in a clinical setting. <bold>Background & Methods:</bold> Recently, AI-tools have been integrated into multiple lung cancer screening (LCS) studies and are currently being considered for use in the clinical routine. During an observation period of one year, we conducted a retrospective analysis of two software tools (S1 and S2) for nodule detection in 766 consecutive clinical thorax CT cases without known history of malignancy in a large university hospital. In 93 of these cases a prior CT examination was available. Within the same observation period, 186 cases with AI detected nodules with a maximum diameter >8mm were included from a radiology practice. Nodules detected by the software tool but not confirmed by the radiologist were classified as false positive. <bold>Results:</bold> At the university hospital, nodules with a mean diameter ≥8 mm were detected in 33% and 38% of cases by S1 and S2, respectively (nodules with a mean diameter ≥5 mm were detected in 40% (S1) and 71% (S2)). In the cases with prior examination, 55% and 49% of nodules were identified as false-positive by S1 and S2, respectively. For true-positive nodules, a volume doubling time of <600 days was shown in 10% of patients for both software tools. In the radiology practice, 77% (S1) and 79% (S2) of nodules >8mm were found to be false-positive. <bold>Conclusion:</bold> While the two tested AI-based lung nodule detection tools may help the radiologist for clinically relevant nodule detection on CT, the high false-positive rate for IPNs >8 mm results in added workload to radiologists and needs to be improved.