Explainable Cryobiopsy AI Model, CRAI, to Predict Disease Progression for Transbronchial Lung Cryobiopsies with Interstitial Pneumonia

Wataru Uegami, Ethan N. Okoshi, Kris Lami, Yuichiro Nei, Mutsumi Ozasa, Kensuke Kataoka, Yuka Kitamura, Yasuo Kohashi, Lee Cooper, Hidenori Sakanashi, Yuji Saito, Yasuhiro Kondoh, Junya FukuokamedRxiv 2025 · DOI

TL;DR

Researchers have developed CRAI, an explainable ensemble AI model that predicts disease progression in patients with Interstitial Lung Disease (ILD) using small, minimally invasive cryobiopsy samples. By extracting 18 pathologically significant features, CRAI outperformed most human pathologists in identifying which patients would face rapid lung function decline and shorter survival times.

What Problem Does This Paper Address?

Interstitial Lung Disease (ILD) is a complex umbrella term for disorders that cause scarring (fibrosis) of the lungs. Predicting how a patient’s disease will progress is notoriously difficult. While surgical lung biopsies (SLB) are the "gold standard" for diagnosis, they are highly invasive. Transbronchial lung cryobiopsy (TBLC) is a safer, less invasive alternative, but it produces smaller tissue samples that are harder for pathologists to interpret.

Compounding this is the issue of inter-observer variability: even expert pathologists often disagree on a diagnosis or prognosis when looking at the same slide. There is a dire need for a standardized, objective way to analyze these smaller cryobiopsy samples to identify "progressive pulmonary fibrosis" early enough to intervene with life-saving antifibrotic drugs.

Methods

The authors developed CRAI (Cryobiopsy AI), an ensemble model designed specifically for the unique challenges of TBLC. Unlike "black-box" models that jump straight from image to prediction, CRAI uses a multi-step, explainable pipeline:

  1. Feature Extraction: Seven specialized deep learning modules (including U-Net for segmentation and YOLOv7 for lymphocyte detection) analyze the tissue at various magnifications (1×, 5×, 20×).
  2. Histological Findings: These modules extract 18 specific features, such as the presence of "honeycombing," collagen density, elastic fiber distribution, and the ratio of fibrotic to cellular areas.
  3. Downstream Prediction: These 18 features are fed into an XGBoost classifier to predict whether a patient’s disease will be "progressive" or "non-progressive."

By breaking the analysis down into human-understandable histological features, the model provides a "rationale" for its predictions, making it more trustworthy for clinical use.

Datasets

The study utilized four distinct datasets to train and validate the model:

  • TGH-SLB: Surgical lung biopsy images used to supplement training data for tissue patterns.
  • TGH-TBLC: The primary cohort of 135 cryobiopsy cases used for model development and internal cross-validation.
  • HRH-PFT: An external test set of 30 cases used to evaluate predictions against Pulmonary Function Test (PFT) data.
  • HRH-survival: An external test set of 48 cases used to evaluate the model’s ability to predict actual patient survival.

Key Insights and Ideas

  • Separating the "Noise": TBLC samples often contain large bronchovascular bundles (stroma) that can distract from the actual lung parenchyma where the disease lives. CRAI includes a specific "Lung-Stroma Model" to distinguish between these regions, a crucial step for accurate analysis of small samples.
  • The "Aha" Moment in Explainability: Using SHAP (SHapley Additive exPlanations) analysis, the researchers found that the AI's predictions correlated with specific biological markers. For instance, the presence of elastic fibers and nodular collagen were strong indicators of progression, while lymphoid follicles were associated with more stable, non-progressive disease.
  • AI vs. Human: The study compared CRAI against 19 experienced pathologists. Only 5 of the 19 pathologists were able to statistically differentiate between progressive and non-progressive cases as effectively as the AI.

Key Results

  • Lung Function Decline: In internal validation, the group CRAI predicted as "progressive" showed a significantly steeper decline in lung function (annual Δ%FVC of -5.198) compared to the "non-progressive" group (-1.293).
  • Survival Prediction: In the external test cohort, patients classified by CRAI as "progressive" had significantly shorter survival times (p=0.034).
  • Consistency: While pathologists showed low agreement (a Fleiss’ kappa of 0.293), CRAI provided a consistent, reproducible assessment across all cases.

Limitations and Future Work

The authors acknowledge a few hurdles:

  • Sample Size: TBLC is a relatively new technique, meaning the available datasets are still small compared to those used in cancer research.
  • Histology-Only Focus: Currently, CRAI only looks at tissue slides. The authors suggest that future versions should integrate clinical data (like age and smoking history) and radiological data (CT scans) for a more holistic "multimodal" prediction.
  • Deployment: The system is currently undergoing a business transfer process and is not yet available for widespread clinical use.

In summary, CRAI represents a significant step toward "precision pathology" for lung disease, turning small, difficult-to-read biopsies into powerful prognostic tools.

Summary updated

Abstract

Abstract Background Interstitial lung disease (ILD) encompasses diverse pulmonary disorders with varied prognoses. Current pathological diagnoses suffer from inter-observer variability,necessitating more standardized approaches. We developed an ensemble model AI for cryobiopsy, CRAI, an artificial intelligence model to analyze transbronchial lung cryobiopsy (TBLC) specimens and predict patient outcomes. Methods We developed an explainable AI model, CRAI, to analyze TBLC. CRAI comprises seven modules for detecting histological features, generating 19 pathologically significant findings. A downstream XGBoost classifier was developed to predict disease progression using these findings. The model’s performance was evaluated using respiratory function changes and survival analysis in cross-validation and external test cohorts. Findings In the internal cross-validation (135 cases), the model predicted 105 cases without disease progression and 30 with disease progression. The annual Δ%FVC was −1.293 in the non-progressive group versus −5.198 in the progressive group, outperforming most pathologists’ diagnoses. In the external test cohort (48 cases), the model predicted 38 non-progressive and 10 progressive cases. Survival analysis demonstrated significantly shorter survival times in the progressive group (p=0.034). Interpretation CRAI provides a comprehensive, interpretable approach to analyzing TBLC specimens, offering potential for standardizing ILD diagnosis and predicting disease progression. The model could facilitate early identification of progressive cases and guide personalized therapeutic interventions. Funding New Energy and Industrial Technology Development Organization (NEDO) and Japanese Ministry of Health, Labor, and Welfare.
BibTeX
@misc{uegami2025explainable,
  author = {Wataru Uegami and Ethan N. Okoshi and Kris Lami and Yuichiro Nei and Mutsumi Ozasa and Kensuke Kataoka and Yuka Kitamura and Yasuo Kohashi and Lee Cooper and Hidenori Sakanashi and Yuji Saito and Yasuhiro Kondoh and Junya Fukuoka and Junya Fukuoka},
  title = {{Explainable Cryobiopsy AI Model, CRAI, to Predict Disease Progression for Transbronchial Lung Cryobiopsies with Interstitial Pneumonia}},
  year = {2025},
  doi = {10.1101/2025.08.06.25333181},
  howpublished = {medRxiv},
}
Figure