Overview

Schema Version

https://atlas.rsna.org/schemas/2025-11/model.json

Name

Assistive AI for Lung Cancer Screening (concurrent reader-assist system)

Link

https://pubs.rsna.org/doi/10.1148/ryai.230079

Indexing

Keywords: Assistive AI, Lung cancer screening, Low-dose CT, Reader study, PACS integration, Nodule localization

Content: CH, CT, OI, RS

RadLex: RID35737, RID50134, RID50149

Author(s)

Atilla P. Kiraly

Corbin A. Cunningham

Ryan Najafi

Zaid Nabulsi

Jie Yang

Charles Lau

Joseph R. Ledsam

Wenxing Ye

Diego Ardila

Scott M. McKinney

Rory Pilgrim

Yun Liu

Hiroaki Saito

Yasuteru Shimamura

Mozziyar Etemadi

David Melnick

Sunny Jansen

Greg S. Corrado

Lily Peng

Daniel Tse

Shravya Shetty

Shruthi Prabhakara

David P. Nadich

Neeral Beladia

Krish Eswaran

Organization(s)

Google Health Research

Waymo

David Geffen School of Medicine at UCLA

Google

Department of Gastroenterology, Sendai Kousei Hospital, Sendai, Japan

MNES Inc, Hiroshima, Japan

Department of Telemedicine, Northwestern University Feinberg School of Medicine, Chicago, IL

Center for Biological Imaging, New York University–Langone Medical Center, New York, NY

Version

1.0

License

Text: CC BY 4.0

URL: https://creativecommons.org/licenses/by/4.0/

Funding

Supported by Google.

Ethical review

Institutional review board approval was granted for each dataset in their respective locations; Advarra IRB reviewed and granted a waiver for further review of these retrospective reader studies.

Date

Updated: 2024-01-07

Published: 2024-03-13

Created: 2023-03-23

References

[1] Kiraly AP, Cunningham CA, Najafi R, et al.. "Assistive AI in Lung Cancer Screening: A Retrospective Multinational Study in the United States and Japan". Radiology: Artificial Intelligence. 2024;6(3):e230079. 2024-01-01. doi:10.1148/ryai.230079. PMID: 38477661. PMCID: PMC11140517.

Model

Availability

Analysis code: https://github.com/Google-Health/google-health/tree/master/analysis; DICOM processing and PACS-accessible outputs library: https://github.com/Google-Health/google-health/tree/master/ct_dicom. AI models used are licensed to DeepHealth (RadNet) and Apollo Hospitals; contact emails provided in the article.

Clinical benefit

Improved screening specificity without loss of sensitivity in reader studies; reduced unnecessary actionable follow-ups; modest reduction in reading time in Japan.

Clinical workflow phase

Concurrent decision support during image interpretation; workflow integration into PACS; workflow optimization.

Degree of automation

Fully automatic AI assistant requiring no manual nodule selection; outputs are surfaced concurrently within PACS.

Indications for use

Assist radiologists interpreting low-dose chest CT examinations for lung cancer screening to assess case-level suspicion and identify up to three suspicious nodules; evaluated with U.S. Lung-RADS v1.1 and Japan Sendai scoring workflows in hospital PACS environments.

Input

Low-dose chest CT (DICOM) studies (current study series; priors available for readers).

Instructions

AI outputs presented as DICOM images preceded by a title slide; readers can reveal AI results after initial review. Outputs include case-level suspicion category and up to three ROI localizations with nodule-level suspicion text; sagittal views shown adjacent to ROIs.

Limitations

Retrospective enriched datasets; training primarily on North American data may affect generalizability (e.g., subsolid nodules more prevalent in Japan); device- and kernel-specific differences; potential lower sensitivity for minimally invasive adenocarcinomas/subsolid nodules; study did not include full clinical data (e.g., smoking history); less experienced readers not evaluated.

Output

CDEs: RDE1707, RDE1700.5, RDE1702, RDE1700

Description: Case-level cancer suspicion category and up to three localized regions of interest with nodule-level suspicion text, rendered as DICOM overlays/images for PACS viewing. Also provides categorical outputs used to compute management thresholds (e.g., Highly Suspicious).

Recommendation

Use as a concurrent assist within PACS to inform Lung-RADS/Sendai scoring and management decisions; radiologist remains the final arbiter.

Regulatory information

Comment: Models are licensed to clinical partners (DeepHealth/RadNet and Apollo Hospitals); no regulatory authorization details provided in the article.

Reproducibility

Public NLST data used; other datasets under license/research agreements. Analysis and DICOM-processing code released. Stand-alone AI thresholds and detailed system architecture provided in supplemental material (Appendix S2).

Use

Intended: Detection and diagnosis

User

Intended: Subspecialist diagnostic radiologist