Table 2

Limitations of the included studies

Reference and country	Applicability concerns regarding European screening population	Design concerns	Laboratory effect26	Applicability concerns to European screening LDCT reading practice	Applicability concerns—actionable nodule definition	Bias in reference standard	Nodule detection or measurement only	Accuracy reported per nodule only
Chamberlin et al, USA14	Yes— 1 US centre; 55–80 years. Excluded images rejected by software due to slice thickness >3 mm or poor image quality.	Yes— Non-comparative (A)	(A) NA	Yes— Stand-alone AI	Yes— >6 mm	Nodules: Yes— <3 experienced chest radiologist; not blinded to index test	Yes— Detection of actionable nodules	No
Hall et al, UK23	No	Yes— Fully paired CT images but not the same readers with and without AI.	(C) Yes— MRMC	Yes— (C) Two inexperienced radiographers. (E) 5% double reading.	No	Nodules: Yes— Original radiologist report (E) plus radiologist review of scans with additional nodules detected by (C).	Yes— Detection of actionable nodules. Detection of malignant nodules ((C) only).	No
Hall et al, UK23	No		(E) No		No	Cancer: Unclear— Cancers detected either from baseline scan or nodule surveillance.		No
Hsu et al, Taiwan17	Yes— One centre in Taiwan; 31/57 never smoked. Selected cases with nodule≤1 cm or no nodules. Exclusion of cases with severe pulmonary fibrosis, diffuse bronchiectasis, extensive inflammatory consolidation, pneumothorax, and massive pleural effusion. 2.5 mm slice thickness.	No	(A) NA	Yes— (A) Stand-alone AI. (B) (C) (D) Three residents in radiology; three experienced chest radiologists.	Yes— No size cut-off used.	Nodules: Yes— <3 experienced chest radiologists; not blinded to index test.	Yes— Detection of any nodules.	Yes— Per-nodule sensitivity; per-person specificity.
Hsu et al, Taiwan17		No	(B) (C) (D) Yes— MRMC		Yes— No size cut-off used.		Yes— Detection of any nodules.	Yes— Per-nodule sensitivity; per-person specificity.
Hwang et al, South Korea15	Yes— 11 (before) and 14 (after) institutions in Korea (K-LUCAS). Only 145/6,487 (2.2%) women.	Yes— (A) Non-comparative. (C) (E) Unpaired non-randomised; not the same readers with and without AI.	(A) NA	(A) Yes— Stand-alone AI.	(A) Yes— No size cut-off used.	Nodules: Yes— Single radiologist with second-read AI. Cancer: Yes— Medical record review.	(A) Yes— Nodule detection (any, actionable, malignant).	(A) Yes—Per-nodule sensitivity.
Hwang et al, South Korea15			(C) (E) No	(C) (E) No	(C) (E) No		(C) (E) No	(C) (E) No
Hwang et al, South Korea33	Yes— 14 institutions in Korea (K-LUCAS). Only 283/10,424 (2.7%) women.	Yes— Non-comparative (C)	No	No	No	Cancer: Yes— Medical record review.	No	No
Jacobs et al, NL, USA, Denmark22	Yes – US NLST (baseline and round 1). Nodule-enriched; Lung-RADS ≥3 120/160 (75%). Slice thickness 1.0 to 3.2 mm.	No	(C) (D) Yes— MRMC	Yes— Three radiologists with >5 year of experience and four radiology residents (fifth year).	No	NA	NA	NA
Lancaster et al, NL, Russia24	Yes— Ultra-LDCT (≤1 mSv); 50–80 years. Selected participants who had ≥1 solid nodule and did not develop lung cancer in following 2 years.	Yes— Not the same readers with and without AI.	(A) NA	(A) Yes— Stand-alone AI	No	Nodule size and categorisation: Yes— 2/4 consensus panel readers involved in index test.	Yes— Nodule measurement and risk category.	No
Lancaster et al, NL, Russia24		Yes— Not the same readers with and without AI.	(C) (D) Yes— MRMC	(C) (D) No	No		Yes— Nodule measurement and risk category.	No
Lo et al, USA18	Yes— US NLST and 2 US hospitals; nodule-enriched (1:2); 3/178 nodules ≥3 mm.	No	(A) NA	Yes— (A) Stand-alone AI. (C) (D) 12 general radiologists.	No	Nodules: No Cancer: No	Yes— Detection of nodules (actionable, malignant).	Yes— Per-nodule sensitivity; per-person specificity.
Lo et al, USA18		No	(C) (D) Yes— MRMC		No	Nodules: No Cancer: No	Yes— Detection of nodules (actionable, malignant).	Yes— Per-nodule sensitivity; per-person specificity.
Park et al, South Korea, USA19	Yes – US NLST; Nodule and cancer-enriched: Prevalence of Lung-RADS≥3 127/200 (64%); lung cancer prevalence 31/200 (16%).	No	(A) NA	Yes— (A) Stand-alone AI. (C) (D) 1 of 5 readers was a fourth-year radiology resident.	No	Cancer: Unclear— Same-year positive cancer diagnosis (not stated how diagnosed).	No	No
Park et al, South Korea, USA19		No	(C) (D) Yes— MRMC		No		No	No
Singh et al, USA20	Yes— US NLST; enriched for sub-solid nodules: prevalence of sub-solid nodules 100/123 (81%).	No	(A) NA	(A) Yes— Stand-alone AI.	No	Nodules: No	Yes— Detection of actionable nodules.	Yes— Per-nodule sensitivity; Per-person specificity.
			(C) (D) Yes— MRMC	(C) Unclear— AI for vessel suppression.
			(C) (D) Yes— MRMC	(D) No
Zhang et al China21	Yes— One hospital in China (part of NELCIN-B3); general population aged 45–74 years.	Yes— Not the same readers with and without AI.	(C) Yes— MRMC	Yes— (C) One resident supervised by 1 radiologist; (E) 1 of 14 residents supervised by 1 of 15 radiologists.	Yes— No size cut-off.	Nodules: Yes— <3 experienced chest radiologists; not blinded to index test.	Yes— Detection of any nodules.	No
Zhang et al China21		Yes— Not the same readers with and without AI.	(E) No		Yes— No size cut-off.		Yes— Detection of any nodules.	No
Legend	No = Random or consecutive screening LDCT images from heavy current or former smokers aged 50–75 years34 living in Europe; no inappropriate exclusions; ≤2 mm slice thickness.	No = Comparative, fully paired design; same readers with and without AI.	No = CT images assessed in clinical practice.	No = Single reading by experienced chest radiologist with (C) or without AI support (D) or (E).	No = In agreement with BTS35, Lung-RADS36 or EUPS37 guidelines.	No = Nodules: ≥3 blinded, experienced chest radiologists. Cancer: Histopathology after biopsy/excision or 2-year follow-up without cancer diagnosis.	No = Accuracy of detection + risk categorisation + recall for lung cancer diagnosis.	No = Per-person sensitivity and specificity.

Index test and comparators: (A) Stand-alone AI: Analysis of CT scan image by AI-based software without human input; (B) Second-read AI: CT scan image first reviewed by an unaided human reader, then re-interpreted after analysis by AI-based software was shown; (C) Concurrent AI: CT scan image reviewed by a human reader assisted by concurrent display of analysis by AI-based software; (D) Unaided reader: CT scan image reviewed by a human reader without AI-based software assistance; (E) Original unaided reader: CT scan image interpreted by a human reader as part of clinical practice; different to the human reader who interpreted the CT scan image in the reader study.
AI, artificial intelligence; BTS, British Thoracic Society; Categ, categorisation; EUPS, European Position Statement; K-LUCAS, Korean Lung Cancer Screening Project; LDCT, low-dose CT; LSUT, Lung Screen Uptake Trial; Lung-RADS, Lung CT Screening Reporting & Data System; MRMC, multi-reader, multi-case study; NA, not applicable; NELCIN-B3, Netherlands-China Big-3 disease screening; NL, Netherlands; NLST, National Lung Screening Trial.