Characteristics of included studies
| Study, country, software, time period for CT scan | Design, setting, sampling method and sample size | Index test and comparator(s) * | Reader details/reading conditions | Reference standard | Reported outcomes |
| Chamberlin et al USA AI-Rad Companion, prototype VA10A (Siemens Healthineers) January 2018–July 201914 | Retrospective test accuracy study (non-comparative). Random sample from 1 US centre; 117 LDCT scans. | (A) Stand-alone AI | NA | Nodules: Consensus of two expert radiologists | Accuracy for detecting nodules >6 mm (per person and per nodule analysis) |
| Hall et al UK Veolity, version 1.2 (MeVis) Date unclear: LSUT November 2015–July 201723 | Retrospective test accuracy study and MRMC study (fully paired). Consecutive sample from UK-based LSUT; 735 LDCT scans. | (C) Concurrent AI (E) Original unaided reader | (C) MRMC: Two radiographers without prior experience in thoracic CT reporting. (E) Clinical practice (LSUT): Five radiologists with 5–28 years of experience in thoracic imaging (5% double reading). | Clinically significant nodules: Original radiologist reading or consensus of two independent radiologists after reviewing discrepant readings between study radiographers and original radiologists. Cancer: NR | Accuracy for detecting clinically significant lung nodules ≥5 mm; accuracy for detecting malignant nodules; reading time |
| Hsu et al Taiwan ClearRead CT, market version (Riverain Technologies) January–December 201717 | MRMC study (fully paired). Consecutive cases with nodules ≤10 mm or no nodules from one hospital in Taiwan; 150† CT images (57 LDCT from lung cancer screening). | (B) Second-read AI (C) Concurrent AI (D) Unaided reader | MRMC: Three residents in radiology and three experienced chest radiologists. Two reading sessions with 8-week washout period: first unaided reading followed by second-read AI, then concurrent AI; images in random order. | Nodules: Consensus of two thoracic radiologists with >15 years of experience | Accuracy for detecting any nodules (stratified by seniority of readers) |
| Hwang et al South Korea AVIEW Lungscreen (Coreline Soft) April 2017–March 201815 | Before-and-after study (unpaired). Consecutive participants from K-LUCAS (11–14 institutions): 1821 participants (before); 4666 participants (after). | (C) Concurrent AI (E) Original unaided reader | Clinical practice: attending thoracic radiologists from 14 institutions (unpaired design) | Cancer: Review of medical records | Accuracy of detecting and categorising actionable nodules to detect lung cancer (Lung-RADS category ≥3) |
| Hwang et al South Korea AVIEW Lungscreen (Coreline Soft) April 2017–December 201816 | Retrospective analysis of prospective cohort study (non-comparative). 10 424 consecutive participants from K-LUCAS (14 institutions) | (C) Concurrent AI | Clinical practice: 25 radiologists from 14 institutions with 5–38 years of experience; no comparator. | Cancer: Review of medical records; lung cancer diagnosed within 1 year (primary outcome) or any time after LDCT (secondary outcome) | Accuracy of detecting and categorising actionable nodules to detect lung cancer (Lung-RADS category ≥3) |
| Jacobs et al USA/Denmark/ Netherlands Veolity, version 1.5 (MeVis) August 2002–August 200422 | MRMC study (fully paired). Random sample from NLST (baseline and round 1), 40 cases from each Lung-RADS category; 160 LDCT scans. | (C) Concurrent AI (D) Unaided reader | MRMC: Three radiologists with >5 years of experience and four radiology residents. Two reading sessions with ≥2 weeks washout period: Half with AI support and half unaided per session; images in random order. | NA | Shift in Lung-RADS categorisation; Reading time |
| Lancaster et al Russia/Netherlands AVIEW LCS, version 1.0.34 (Coreline Soft) February 2017–201824 | MRMC study (fully paired). Nodule-enriched: ≥1 solid nodule, no lung cancer diagnosed within 2 years from MLCS baseline scan; 283 ultra-LDCT scans. | (A) Stand-alone AI (C) Concurrent AI (D) Unaided reader | (C) MRMC: Three thoracic radiologists with >7 years of experience in lung cancer screening. (D) MRMC: Two different thoracic radiologists with >7 years of experience in lung cancer screening (using other semi-automated volume measurement software). | Nodules: Independent consensus of three experienced radiologists and one IT technologist | Accuracy of nodule volume measurement and categorisation (<100 mm3, ≥100 mm3) |
| Lo et al USA ClearRead CT (first generation, pre-market) (Riverain Technologies) Date unclear: NLST screened from August 2002 to September 200718 | MRMC study (fully paired). Nodule-enriched: 2 normal cases for each case with nodules from NLST and 2 US hospitals; 324 LDCT scans. | (A) Stand-alone AI (C) Concurrent AI (D) Unaided reader | MRMC: 12 general radiologists with 6–26 years of experience. Two reading sessions with a minimum interval of 37 days (mean, 57 days): first unaided; then AI-assisted. | Actionable nodules: Consensus of three expert thoracic radiologist assisted by corresponding NLST or source documentation. Cancer: Histological findings (presence) or long-term follow-up (absence). | Accuracy for detecting actionable nodules; accuracy for detecting malignant nodules; reading time |
| Park et al USA/South Korea VUNO Med-LungCT AI, v.1.0.1 (VUNO) Date unclear: NLST baseline screen from August 200219 | MRMC study (fully paired). Nodule-enriched from NLST baseline screens; 200 LDCT scans. | (A) Stand-alone AI (C) Concurrent AI (D) Unaided reader | MRMC: One radiology resident and four radiologists with 1–20 years of experience. Two reading sessions with 6-week washout period: first unaided, then AI-assisted; images in random order. | Cancer: NR (Lung cancer diagnosed within 1 year in the NLST) | Accuracy of detecting and categorising actionable nodules to detect lung cancer (Lung-RADS category ≥3) Change in Lung-RADS category |
| Singh et al USA ClearRead CT, market version (Riverain Technologies) Date unclear: NLST screened from August 2002 to September 200720 | MRMC study (fully paired). Nodule-enriched: 100 with SSNs and 23 without SSNs from NLST; 123 LDCT scans. | (A) Stand-alone AI (C) Concurrent AI (vessel suppression only) (D) Unaided reader | MRMC: Two radiologists with 5 and 10 years of thoracic CT experience; sequential interpretation of unprocessed CT images alone, then vessel-suppressed images without washout period. | Sub-solid nodules: Consensus of two experienced thoracic radiologists (11 and 27 years of experience) with adjudication of conflicts by a third radiologist | Accuracy for detecting sub-solid nodules ≥6 mm; Change in Lung-RADS category |
| Zhang et al China InferRead CT Lung, market version (Infervision) November to December 201921 | Retrospective test accuracy study and MRMC study (fully paired). Consecutive sample from one hospital in China; 860 LDCT scans. | (C) Concurrent AI (MRMC study) (E) Original unaided reader (clinical practice) | (C) MRMC: One resident with 5 years of experience with supervision by one radiologist with 20 years of experience. (E) Clinical practice: One resident drafted report with supervision by radiologist (14 residents and 15 radiologists). | Nodules: Consensus of two radiologists with 20 and 31 years of experience | Accuracy for detecting any nodules (stratified by types and sizes of nodules) |
*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 was first reviewed by an unaided human reader, then was re-interpreted after analysis by AI-based software was shown; (C) Concurrent AI: CT scan image was reviewed by a human reader assisted by concurrent display of analysis by AI-based software; (D) Unaided reader: CT scan image was reviewed by a human reader without assisted by AI-based software; (E) Original unaided reader: CT scan image was interpreted by a human reader as part of clinical practice, and therefore the reader was different from the human reader who interpret the CT scan image in the reader study.
†The study included mixed populations. Only those who underwent CT scans for screening were included in this systematic review.
AI, artificial intelligence; K-LUCAS, Korean Lung Cancer Screening; LDCT, low-dose CT; LSUT, Lung Screen Uptake Trial; Lung-RADS, Lung CT Screening Reporting & Data System; MLCS, Moscow Lung Cancer Screening; MRMC, multi-reader, multi-case study; NA, not applicable; NLST, National Lung Screening Trial.