Negligible impact of perifissural nodules in an AI-first reader workflow from UK lung screening trial.
Authors
Affiliations (13)
Affiliations (13)
- Research Institute for Diagnostic Accuracy, Groningen, The Netherlands.
- Department of Public Health, Erasmus Medical Center Rotterdam, Rotterdam, The Netherlands.
- Department of Respiratory Medicine, University of Groningen, University Medical Center Groningen, Groningen, The Netherlands.
- Department of Epidemiology, University of Groningen, University Medical Center Groningen, Groningen, The Netherlands.
- Roy Castle Lung Cancer Research Programme, Department of Molecular and Clinical Cancer Medicine, University of Liverpool, Liverpool, UK.
- Medimaps Group SA, Geneva, Switzerland.
- Department of Radiology, Gelre Ziekenhuizen, Apeldoorn, The Netherlands.
- Department of Medicine and Surgery (DiMeC), Scienze Radiologiche, University of Parma, Parma, Italy.
- Department of Radiology, University of Massachusetts Memorial Health, University of Massachusetts, Chan Medical School, Worcester, MA, USA.
- Coreline Soft, Seoul, Republic of Korea.
- Department of Respiratory Medicine, Amsterdam University Medical Center, Amsterdam, The Netherlands.
- Research Institute for Diagnostic Accuracy, Groningen, The Netherlands. [email protected].
- Faculty of Medical Sciences, University of Groningen, Groningen, The Netherlands. [email protected].
Abstract
To evaluate the effect of perifissural nodules (PFNs) on radiologist workload within an AI-first reader workflow for lung cancer screening, given that AI cannot morphologically classify benign PFNs measuring ≥ 100 mm<sup>3</sup>. One thousand two hundred fifty baseline low-dose CT scans from the UK Lung Screening (UKLS) Trial were analyzed. A commercially available AI software automatically identified all nodules with solid components ≥ 100 mm³ per the NELSON 2.0-European Position Statement (EUPS) guideline. Three readers independently performed PFN classification, with a senior radiologist with over 20 years of experience performing an arbitration read for the final reference classification (typical PFN, atypical PFN, or non-PFN). Histological outcomes for all fissure-attached nodules were reviewed to confirm benignity. The proportion of participants where a benign typical PFN was the sole finding of nodule presence ≥ 100 mm³ was calculated, representing the extra workload for radiologists to review. A total of 1252 participants (mean age, 68.5 ± 4.0 years; 928 men [74%]) were analyzed. AI detected 838 nodules with solid components ≥ 100 mm³ in 431 (34%) participants. 57 nodules in 49 (3.9%) participants were classified as typical PFNs by the reference standard. Only 24 of 1252 participants (1.9%) had a typical PFN ≥ 100 mm³ as the sole finding that added extra workload. No typical PFNs (0/57) were malignant. The impact of typical PFNs on the maximum achievable radiologist workload reduction in an AI-first reader workflow is negligible, with only 1.9% of participants requiring additional radiologist review triggered solely by these benign nodules. Question In an AI-first lung cancer screening workflow, do typical PFNs ≥ 100 mm<sup>3</sup> create a significant bottleneck for radiologist workload? Findings In the UKLS trial, typical PFNs ≥ 100 mm³ were rare, creating negligible extra workload (1.9% of participants), and none were malignant (0/57). Clinical relevance The concern that PFN morphology creates a bottleneck in AI-first screening workflows is unfounded. Our findings support the feasibility of volume-based AI triage, allowing radiologists to focus on other false positives without being overwhelmed by PFNs.