Explainable Machine Learning for Estimating the Contrast Material Arrival Time in Computed Tomography Pulmonary...

Area of Science:

• Radiology
• Artificial Intelligence
• Medical Imaging

Background:

• Accurate contrast arrival time (TARR) in CT pulmonary angiography (CTPA) is crucial for diagnosis.
• Predicting TARR using non-invasive features can optimize scan protocols.

Purpose of the Study:

• To develop an explainable machine learning (ML) model for predicting pulmonary artery TARR in CTPA.
• Utilize patient and noncontrast CT features for TARR prediction.

Main Methods:

• Retrospective study of 666 patients undergoing CTPA.
• Employed recursive feature elimination and XGBoost with SHAP for ML modeling.
• External validation was performed to assess model generalizability.

Main Results:

• ML models achieved high performance (AUC > 0.83) in predicting abnormal TARR (<7s or >10s).
• SHAP analysis highlighted vena cava and pulmonary artery measurements as key predictors.
• The models demonstrated strong performance on both testing and external validation sets.

Conclusions:

• An explainable ML algorithm accurately identifies normal and abnormal pulmonary artery TARR.
• This approach facilitates personalized CTPA scan protocols.