A CADx tool improves lung nodule risk stratification when compared to British Thoracic Society guidelines on routine...

The aim of this study was to evaluate whether a computed tomography (CT) image-based lung nodule artificial intelligence (AI) classifier can improve lung nodule risk classification and clinical management decisions relative to British Thoracic Society (BTS) 2015 guidelines.

This is a retrospective single-site case-control study of incidental lung nodules identified from routine clinical CT scans. An AI-based classification tool (RevealAI-Lung) was used to compute a malignancy similarity index (mSI) of lung nodules from CT scans of patients with known definitive diagnoses. For comparison to current clinical best practices, nodules were retrospectively classified for follow-up using the BTS 2015 guidelines by a re-review of all cases. The mSI was used to adjust BTS recommendations (msi > 0.9: upgrade, mSI < 0.1: downgrade, and 0.1 < mSI < 0.9: maintain), and performance of the mSI reclassification against BTS guidelines was compared.

A total of 100 control patients and 100 cancer patients were analysed from an UK National Health Service (NHS) referral population who had proven diagnosis (benign or malignant). Forty-five of 100 cancer patients would have received interval scans before eventual malignant diagnosis. mSI would have upgraded 44% (n=20/45) of these cases and downgraded none, reducing delayed cancer diagnoses from 45% to 25% (P<0.001). A total 100 control patients would have received followup scans before eventual benign diagnosis. mSI would have downgraded 7% (n=7/100) of these cases to no follow-up and upgraded 7% (n=7/100) from an interval scan to an immediate positron emission tomography (PET)-CT.

mSI significantly improved nodule classification in incidental lung nodules when combined with current best practice BTS guidelines.