Explainable deep-learning-based ischemia detection using hybrid O-15 H₂O perfusion positron emission tomography and...

Area of Science:

• Cardiovascular imaging analysis
• Artificial intelligence in medicine
• Medical diagnostics

Background:

• Flow-limiting coronary artery disease (CAD) diagnosis relies on complex imaging and clinical data.
• Current diagnostic methods can be labor-intensive and require expert interpretation.
• Need for accurate, efficient, and interpretable tools for CAD detection.

Purpose of the Study:

• To develop and evaluate an explainable deep-learning (DL) classifier for identifying flow-limiting CAD.
• To integrate O-15 H2O perfusion positron emission tomography computed tomography (PET/CT) and coronary CT angiography (CTA) imaging data.
• To assess the performance of the DL classifier against invasive coronary angiography.

Main Methods:

• A combined image-and data-based DL model was developed using 35 clinical, CTA, and PET variables from 138 individuals.
• The DL model utilized polar map images and numerical data, incorporating explainability features.
• Performance metrics included accuracy, AUC, F1 score, sensitivity, specificity, precision, net benefit, and Cohen's Kappa, compared to invasive coronary angiography.

Main Results:

• The DL model achieved a median accuracy of 0.8478 and AUC of 0.8481.
• Key performance indicators (ACC, AUC, F1S, SEN, SPE, PRE) demonstrated strong diagnostic capability.
• The DL model showed comparable performance to clinical reading, with no statistically significant differences.

Conclusions:

• The developed explainable DL model is a feasible and effective tool for detecting CAD.
• The model successfully integrates multimodal imaging and clinical data for improved diagnostic accuracy.
• The explainable nature of the DL model allows for highlighting important data findings in an interpretable manner.