Added value of artificial intelligence for the detection and analysis of lung nodules on ultra-low-dose CT in an emergency setting

PURPOSE: To analyze the added value of an artificial intelligence (AI) algorithm for lung nodule detection on ultra-low-dose CT's (ULDCT) acquired at the emergency department (ED). MATERIALS AND METHODS: In the OPTIMACT trial 873 patients with suspected non-traumatic pulmonary disease underwent ULDCT at the ED, 870 patients were available for analysis. During the trial clinical reading of the ULDCT's was done by the radiologist on call and clinical relevant incidental lung nodules were reported in the radiology report. All ULDCT's were processed using CIRRUS Lung Nodule AI, a deep learning based algorithm for detection of non-calcified lung nodules >= 6 mm. Three chest radiologists independently reviewed the lung nodules identified during the trial and all marks from the AI algorithm. Each AI mark was accepted or rejected for being a lung nodule. Accepted AI marks were classified as solid, part-solid, non-solid and were volumetrically measured using semi-automatic segmentation software. Incidental lung nodules that (i) were scored as a nodule by at least two of the three chest radiologists and (ii) met the Fleischner criteria for clinically relevant lung nodules were used as reference standard. We assessed differences in proportion of true positive and false positive findings detected during prospective evaluation by the radiologist on call versus a standalone AI reading. RESULTS: During the trial 59 clinical relevant incidental lung nodules in 35/870 (4.0%) patients had been reported. 24/59 of these nodules were scored by at least two chest radiologists and met the Fleischner criteria, leaving 35/59 false positives. In 458/870 (53%) ULDCT's one or more AI marks were found, 1862 marks in total. 104/1862 (5.6%) AI marks were scored as clinically relevant nodules by at least two chest radiologists, leaving 1758/1862 (94%) false positive marks. Overall, 4 times more (104 vs 24) lung nodules were detected with the use of AI, at the expense of 50 times more false positive findings. CONCLUSION: The use of AI on ULDCT in ED patients with pulmonary disease results in the detection of more clinically relevant incidental lung nodules but is limited by the high false positive rate. CLINICAL RELEVANCE: In the ED setting focus lies on the acute presentation. Artificial intelligence aids in the detection of clinical relevant incidental lung nodules but is limited by the high false positive rate.