Translating potential improvement in the precision and accuracy of lung nodule measurements on computed tomography scans by software derived from artificial intelligence into impact on clinical practice—a simulation study

Abstract

Abstract Objectives Accurate measurement of lung nodules is pivotal to lung cancer detection and management. Nodule size forms the main basis of risk categorization in existing guidelines. However, measurements can be highly variable between manual readers. This article explores the impact of potentially improved nodule size measurement assisted by generic artificial intelligence (AI)-derived software on clinical management compared with manual measurement. Methods The simulation study created a baseline cohort of people with lung nodules, guided by nodule size distributions reported in the literature. Precision and accuracy were simulated to emulate measurement of nodule size by radiologists with and without the assistance of AI-derived software and by the software alone. Nodule growth was modelled over a 4-year time frame, allowing evaluation of management strategies based on existing clinical guidelines. Results Measurement assisted by AI-derived software increased cancer detection compared to an unassisted radiologist for a combined solid and sub-solid nodule population (62.5% vs 61.4%). AI-assisted measurement also correctly identified more benign nodules (95.8% vs 95.4%); however, it was associated with over an additional month of surveillance on average (5.12 vs 3.95 months). On average, with AI assistance people with cancer are diagnosed faster, and people without cancer are monitored longer. Conclusions In this simulation, the potential benefits of improved accuracy and precision associated with AI-based diameter measurement is associated with additional monitoring of non-cancerous nodules. AI may offer additional benefits not captured in this simulation, and it is important to generate data supporting these, and adjust guidelines as necessary. Advances in knowledge This article shows the effects of greater measurement accuracy associated with AI assistance compared with unassisted measurement.