Validation of AI-Based Detection of Idiopathic Pulmonary Fibrosis in Serial Chest Radiographs: A Retrospective Longitudinal Study

Retrospective Evaluation of AI-Based Early Detection of Reticular Opacity in Longitudinal Chest Radiograph Sequences in Patients With Idiopathic Pulmonary Fibrosis

Idiopathic pulmonary fibrosis (IPF) is a chronic, progressive fibrotic lung disease of unknown cause with a median survival of only 3-5 years after diagnosis. Early detection and timely initiation of antifibrotic therapy may improve outcomes, but diagnosis is frequently delayed. Chest radiography (CXR) is widely accessible and cost-effective but has limited sensitivity for early interstitial opacity (IO), so radiologists may miss or delay documentation of relevant findings.

This retrospective, single-center, observational cohort study evaluates whether an artificial-intelligence algorithm (VUNO Med-Chest X-ray) can detect interstitial opacity earlier than radiologists in the historical chest radiograph series of patients who were diagnosed with IPF. The cohort was identified via a April 2025 registry screening of patients carrying an IPF diagnosis at Chung-Ang University Hospital. For each patient, the date of the first AI-detected IO (using a pre-specified score cutoff) is compared with the date of the first radiologist-reported mention of interstitial/reticular opacity, across all chest radiographs obtained before the IPF diagnosis date, within a 15-year retrospective imaging window anchored to the April 2025 screening date (January 2010-April 2025). The study also explores patient characteristics that modify this lead-time difference and whether longitudinal AI IO-score trajectories are associated with mortality.

Study Overview

• Status: Completed
• Conditions: Idiopathic Pulmonary Fibrosis; Interstitial Lung Disease
• Intervention / Treatment: Device: VUNO Med-Chest X-ray

Detailed Description

Background:

IPF is an interstitial lung disease with a median survival of 3-5 years following diagnosis. Because interstitial opacity on CXR is subtle in early disease, opportunistic AI-based detection during CXR obtained for unrelated indications (emergency, gastroenterology, cardiology, etc.) could substantially shorten the time to diagnosis. Prior AI-CXR research has largely validated single-timepoint detection performance for findings such as pneumothorax, nodules, and pleural effusion; few studies have quantified how much earlier an AI system can detect interstitial opacity compared with radiologist reporting across a patient's full longitudinal CXR history.

Objectives:

1. Quantify the lead-time advantage of AI (VUNO Med-Chest X-ray) versus radiologist reporting for first detection of interstitial opacity in the retrospective CXR series of IPF patients.
2. Compare the proportion of early detections (greater than 180 days before diagnosis) between AI and radiologists.
3. Explore patient-level effect modifiers of the AI-radiologist lead-time difference (follow-up duration, number of CXRs, prior emphysema/pneumonia history, prior CT availability, CPFE/COPD mention, pulmonology/allergy visit history).
4. Examine whether longitudinal AI IO-score trajectory patterns (progressive, oscillating, persistently high) are associated with mortality using Cox proportional-hazards modeling.

Design and methods:

Retrospective, single-center, observational cohort study conducted at Chung-Ang University Hospital (Seoul, Republic of Korea), reported in accordance with STROBE reporting guidelines. The cohort was identified via an April 2025 registry screening of patients carrying an IPF diagnosis at Chung-Ang University Hospital. All available frontal (PA or AP) chest radiographs obtained before each patient's IPF diagnosis date, within a 15-year retrospective imaging window anchored to the April 2025 screening date (January 2010-April 2025), are analyzed with VUNO Med-Chest X-ray (interstitial opacity, consolidation, nodule/mass) and compared against the corresponding radiology reports. Vital status (mortality) was ascertained as of the April 2025 screening date, with no follow-up beyond that cutoff.

Index test: VUNO Med-Chest X-ray interstitial opacity (IO) score, using a pre-specified cutoff (0.35). The AI-detected date is defined as the date of the earliest pre-diagnosis CXR meeting this cutoff.

Comparator: The radiologist-detected date is defined as the earliest chest-radiograph report date containing terminology consistent with reticular pattern / interstitial opacity(reticular opacity) / IO. A sensitivity analysis separately accounts for reports where a chest CT was obtained within +/- 3 months, given the potential influence of CT findings on radiologist reporting. Of 175 enrolled patients with a confirmed IPF diagnosis and an available pre-diagnosis CXR series, 166 comprise the primary paired analysis cohort (AI and radiologist detection dates both available before diagnosis).

• Study Type: Observational
• Enrollment (Actual): 175

Contacts and Locations

Study Locations

• South Korea
  • Seoul, South Korea
    • Chung-Ang University Hospital

Participation Criteria

Eligibility Criteria

• Ages Eligible for Study: Adult; Older Adult
• Accepts Healthy Volunteers: No

• Sampling Method: Non-Probability Sample

Study Population

Adults aged 19 years or older carrying a final diagnosis of idiopathic pulmonary fibrosis (IPF; ICD-10 J84.1 or clinical diagnosis) at Chung-Ang University Hospital, identified via an April 2025 registry screening, with a digital chest radiograph series available before the diagnosis date within the 15-year retrospective imaging window (January 2010-April 2025).

Description

Inclusion Criteria:

• IPF diagnosis on record at Chung-Ang University Hospital as of the April 30, 2025 registry screening, based on imaging findings, pathology results, and clinical information as determined by a pulmonology specialist
• Age greater than or equal to 19 years at IPF diagnosis
• Confirmed diagnosis of IPF (by clinician or multidisciplinary discussion, including CT and/or biopsy)
• Two or more frontal (PA or AP) chest radiographs obtained before the diagnosis date
• DICOM images available and analyzable by VUNO Med-Chest X-ray
• Date of initial IPF diagnosis available

Exclusion Criteria:

• Only non-frontal chest radiograph views available (e.g., lateral view only)
• One or fewer analyzable chest radiographs
• Missing initial diagnosis date
• No radiology report data available for comparison

Study Plan

How is the study designed?

Number of groups / cohorts

1

Cohorts and Interventions

Group / Cohort

Intervention / Treatment

IPF-diagnosed cohort; Patients with a final diagnosis of idiopathic pulmonary fibrosis at Chung-Ang University Hospital, identified via an April 2025 registry screening, whose historical chest radiograph series obtained before diagnosis (within a 15-year retrospective window, January 2010-April 2025) were retrospectively analyzed by both the AI algorithm and radiology reports.

Device: VUNO Med-Chest X-ray; Retrospective, offline application of the AI-based chest radiograph analysis software VUNO Med-Chest X-ray (VUNO Inc., Seoul, Korea) to archival chest radiographs obtained before each patient's IPF diagnosis. The software outputs scores for interstitial opacity(reticular opacity), consolidation, and nodule/mass; interstitial opacity(reticular opacity) score, applying a pre-specified cutoff, is used for the primary and secondary analyses. The AI analysis is performed solely for research purposes and does not inform clinical care.

What is the study measuring?

Primary Outcome Measures

Outcome Measure	Measure Description	Time Frame
Paired lead-time difference (radiologist first-mention date minus AI first-detection date, days)	Delta = radiologist_detected_date - ai_detected_date, in days. Delta greater than 0 indicates AI detected interstitial opacity earlier than the radiologist; Delta = 0 indicates same-day detection; Delta less than 0 indicates the radiologist detected it earlier. Analyzed in the paired cohort (n=166) using the Wilcoxon signed-rank test (zero differences excluded, two-sided), with effect size reported as the Hodges-Lehmann estimate and bootstrap 95% CI (4,000 resamples). Reported measures: median Delta (IQR), Hodges-Lehmann estimate (95% CI), p-value, and the proportional breakdown of AI-earlier / same-day / radiologist-earlier pairs (n, %).	From first available chest radiograph to IPF diagnosis date (retrospective, up to 15 years, anchored to April 2025 registry screening)

Secondary Outcome Measures

Outcome Measure	Measure Description	Time Frame
Proportion with AI-earlier detection among discordant pairs	Exact binomial sign test (null hypothesis p=0.5, two-sided); reported as the AI-earlier proportion (n/N, %) and p-value, assessed among the subset of the paired cohort with a non-zero lead-time difference (n=73).	From first available chest radiograph to IPF diagnosis date (retrospective, up to 15 years, anchored to April 2025 registry screening)
Proportion detecting more than 180 days before diagnosis - AI vs. Radiologist	Exact McNemar test plus continuity-corrected asymptotic McNemar test on a 2x2 paired table; reported measures: each proportion (n, %, 95% CI), number of discordant pairs, and p-value. Assessed in the full enrolled cohort (n=175).	From first available chest radiograph to IPF diagnosis date (retrospective, up to 15 years, anchored to April 2025 registry screening)
Proportion detecting within 180 days before diagnosis - AI vs. Radiologist	Exact McNemar test plus continuity-corrected asymptotic McNemar test on a 2x2 paired table; reported measures: each proportion (n, %, 95% CI), number of discordant pairs, and p-value. Assessed in the full enrolled cohort (n=175).	From first available chest radiograph to IPF diagnosis date (retrospective, up to 15 years, anchored to April 2025 registry screening)
Sensitivity analysis of the primary lead-time comparison using alternative zero-handling methods	Wilcoxon signed-rank test repeated with three zero-handling methods (wilcox: zeros excluded; pratt: zeros included in ranking then removed; zsplit: zero ranks split); statistic and p-value reported for each method. Assessed in the primary paired analysis cohort (n=166).	From first available chest radiograph to IPF diagnosis date (retrospective, up to 15 years, anchored to April 2025 registry screening)

Other Outcome Measures

Outcome Measure	Measure Description	Time Frame
Kaplan-Meier Time-to-Detection Curves for AI versus Radiologist	Kaplan-Meier curves comparing time from first available chest radiograph to first detection, separately for AI and radiologist, with log-rank test comparing the two curves.	From first available chest radiograph to IPF diagnosis date (retrospective, up to 15 years, anchored to April 2025 registry screening)
Time From AI First Detection to Eventual IPF Diagnosis	Descriptive Kaplan-Meier analysis of the time elapsed between the AI's first positive detection and the patient's pulmonology/allergy-confirmed IPF diagnosis.	From first available chest radiograph to IPF diagnosis date (retrospective, up to 15 years, anchored to April 2025 registry screening)
Predictors of the Primary Lead-Time Difference (Regression Coefficients)	Exploratory, hypothesis-generating analysis of candidate patient-level and process-level variables potentially associated with the primary lead-time difference (e.g., patient demographics, follow-up duration, CXR imaging frequency, comorbidity history, and specialist visit patterns), using univariable and multivariable OLS regression with HC3 robust standard errors as the primary model, with quantile regression (median, tau=0.5) as a robustness check. The final covariate set will be determined based on pre-specified clinical rationale at the time of analysis. Reported measures: regression coefficients (β, days), 95% CI, and p-value for each retained predictor.	From first available chest radiograph to IPF diagnosis date (retrospective, up to 15 years, anchored to April 2025 registry screening)

Collaborators and Investigators

• Sponsor: Chung-Ang University Hospital
• Collaborators: VUNO Inc.
• Investigators: Principal Investigator: Kyoungmin Moon, Chung-Ang University Hospital; Principal Investigator: Yoona Hwang, VUNO Inc.

Study record dates

Study Major Dates

• Study Start (Actual): April 30, 2025
• Primary Completion (Actual): April 30, 2025
• Study Completion (Actual): April 30, 2025

Study Registration Dates

• First Submitted: July 2, 2026
• First Submitted That Met QC Criteria: July 13, 2026
• First Posted (Actual): July 20, 2026

Study Record Updates

• Last Update Posted (Actual): July 20, 2026
• Last Update Submitted That Met QC Criteria: July 13, 2026
• Last Verified: July 1, 2026

More Information

Terms related to this study

• Keywords: Artificial Intelligence; Deep Learning; Idiopathic Pulmonary Fibrosis; Interstitial Lung Disease; Retrospective Cohort Study; Chest Radiograph; Lead Time; STROBE
• Additional Relevant MeSH Terms: Respiratory Tract Diseases; Lung Diseases; Pulmonary Fibrosis; Idiopathic Pulmonary Fibrosis; Lung Diseases, Interstitial
• Other Study ID Numbers: 2507-022-19585

Plan for Individual participant data (IPD)

• Plan to Share Individual Participant Data (IPD)?: NO
• IPD Plan Description: This is a retrospective study using data collected under an IRB-approved waiver of informed consent. Individual participant data were not collected with participant consent for sharing with third parties, and no data-sharing infrastructure or de-identification protocol for external release has been established.

Drug and device information, study documents

• Studies a U.S. FDA-regulated drug product: No
• Studies a U.S. FDA-regulated device product: No