Prognostication in patients with idiopathic pulmonary fibrosis using quantitative airway analysis from HRCT: a retrospective study

Nan Yang, Francesco Federico, Stephen M. Humphries, John A. Mackintosh, Christopher Grainge, Helen E. Jo, Nicole Goh, Paul N. Reynolds, Peter Hopkins, Vidya Navaratnam, Yuben Moodley, E. Haydn Walters, Samantha Ellis, Gregory J. Keir, Christopher Zappala, Tamera J. Corte, Ian Glaspole, Athol U. Wells, Guang Yang, Simon WalshEuropean 2025 · DOI

TL;DR

Researchers have developed a novel AI model called SABRE that analyzes high-resolution CT (HRCT) scans to extract quantitative airway biomarkers. These biomarkers provide a more accurate prediction of disease progression and mortality in patients with idiopathic pulmonary fibrosis (IPF) than existing methods, offering a new tool for personalized treatment and clinical trial design.

What Problem Does This Paper Address?

Idiopathic pulmonary fibrosis (IPF) is a devastating lung disease characterized by unpredictable progression, ranging from stability to rapid deterioration. This heterogeneity makes it difficult to prioritize treatment and manage patients effectively. Current methods for predicting disease trajectory, such as lung function tests (e.g., FVC, DLCO) and visual assessment of HRCT scans, have limitations in accuracy and reproducibility. Specifically, while traction bronchiectasis (airway dilation due to fibrosis) is known to be a prognostic indicator, its precise quantification from HRCT has been challenging. This paper aims to address this gap by developing an AI-driven approach to extract quantitative airway features from HRCT that can reliably predict patient outcomes.

Methods

The core of this research is the development of the SABRE (Smart Airway Biomarker Recognition Engine), an AI system designed to quantify airway structures from HRCT scans. SABRE utilizes a fuzzy attention neural network and a smart branch decomposition module. The process involves:

  1. Airway Localization: Identifying the trachea and central points within the 3D HRCT data.
  2. 2D Projection: Projecting the 3D airway tree onto two 2D planes (y-axis and z-axis).
  3. Distance Mapping: Calculating the distance of each airway pixel from the central points.
  4. Branch Segmentation: Segmenting airways into terminal, small, and medium branches based on distance percentiles.
  5. Volume Quantification: Calculating normalized volumes for different airway segments (STermAV, SSmallAV, SMedAV) and the total airway volume (STotalAV). A combined metric, SPAV (Primary Airway Volume), was also derived.

These quantitative airway biomarkers were then used in multivariable analyses (Cox regression) to assess their association with mortality and 12-month disease progression, controlling for traditional measures like FVC, DLCO, composite physiological index (CPI), and visual assessments of disease severity (fibrosis extent, traction bronchiectasis, interstitial lung disease extent).

Datasets

The study utilized several datasets:

  • Training, Validation, and Test Sets for SABRE Development: Data from ATM22 and AIIB23 datasets, which include HRCT scans from healthy volunteers and patients with various pulmonary diseases (nodules, ground-glass opacities, emphysema, lung cancer, COVID-19).
  • External Testing and Clinical Analysis Cohorts:
    • Australian Idiopathic Pulmonary Fibrosis Registry (AIPFR): A national registry of IPF patients.
  • Open Source Imaging Consortium (OSIC): A large dataset of patients with various fibrotic lung diseases.

These datasets were crucial for both developing and validating the SABRE model's performance in a real-world clinical setting.

Key Insights and Ideas

  • Airways Hold Significant Prognostic Power: The study strongly suggests that quantitative analysis of airway structures, beyond just visual assessment of traction bronchiectasis, carries substantial prognostic information for IPF.
  • SABRE Outperforms Traditional Measures: The AI-derived airway biomarkers from SABRE independently predict mortality and progression, even after accounting for established clinical and imaging markers. This indicates SABRE captures signals not present in current assessment tools.
  • Interpretability Matters: While other AI approaches exist, SABRE's strength lies in its ability to provide interpretable quantitative metrics (airway volumes) that can be directly linked to disease mechanisms.
  • Combined Biomarkers Enhance Prediction: Combining SABRE-derived airway volumes with existing measures like DLCO significantly improves prognostic accuracy, highlighting the potential for multi-modal prediction models.
  • Potential for Early Intervention: The ability of SABRE to identify high-risk patients suggests it could facilitate earlier intervention with antifibrotic therapies or closer monitoring before irreversible lung damage occurs.

Key Results

  • Independent Predictors of Mortality and Progression: SABRE-based variables (STotalAV, SPAV, SMedAV, STermAV) were independently associated with mortality and 12-month progression in multivariable analyses, even when controlling for disease severity and traditional lung function measures.
  • Improved Prognostic Utility: Combining SABRE with DLCO resulted in a C-index of 0.752 for mortality prediction, and an AUC of 0.852 for early (year 1) prediction.
  • Risk Stratification: Kaplan-Meier survival curves showed clear stratification into low, medium, and high-risk groups based on SABRE variables, with significantly different survival outcomes. The SABRE-defined high-risk group captured a higher proportion of progression events compared to CPI-based stratification.
  • Superiority over Other AI Tools: SABRE demonstrated superior prognostic utility compared to other AI-based tools like AirQuant and DeepDTA in time-dependent ROC analyses.

Limitations and Future Work

  • HRCT Data Dependency: SABRE relies on volumetric HRCT data, which may limit its application to patients with interspaced CT scans, a common practice in some centers.
  • Need for Standardization: The authors emphasize the need for global standardization of HRCT protocols to ensure consistent data quality and facilitate the adoption of advanced analysis methods.
  • Longitudinal Evaluation: Future research should explore automated longitudinal evaluation of airway volume changes, as these may be sensitive markers of progression.
  • Bronchial Segmentation Challenges: Challenges remain in accurately segmenting airways in regions with extensive honeycombing, requiring further advancements in network architecture and training.
  • Clinical Trial Applications: The authors suggest SABRE could be used for cohort enrichment in clinical trials and as a sensitive imaging end-point to detect treatment effects earlier than traditional lung function measures. Further validation is needed for its potential use as a surrogate end-point.

Summary updated

Abstract

SABRE-based variables capture prognostic signals beyond that provided by traditional measurements, disease severity scores and established AI-based methods, reflecting the progressiveness and pathogenesis of the disease.
BibTeX
@inproceedings{yang2025prognostication,
  author = {Nan Yang and Francesco Federico and Stephen M. Humphries and John A. Mackintosh and Christopher Grainge and Helen E. Jo and Nicole Goh and Paul N. Reynolds and Peter Hopkins and Vidya Navaratnam and Yuben Moodley and E. Haydn Walters and Samantha Ellis and Gregory J. Keir and Christopher Zappala and Tamera J. Corte and Ian Glaspole and Athol U. Wells and Guang Yang and Simon Walsh},
  title = {{Prognostication in patients with idiopathic pulmonary fibrosis using quantitative airway analysis from HRCT: a retrospective study}},
  booktitle = {European},
  year = {2025},
  doi = {10.1183/13993003.00981-2025},
}
Figure