How scan parameter choice affects deep learning-based coronary artery disease assessment from computed tomography

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Cardiovascular Imaging

Background:

• Deep learning (DL) algorithms show promise for assessing Coronary Artery Disease (CAD) severity from Coronary Computed Tomography Angiography (CCTA) scans.
• Evaluating the robustness of these DL algorithms to variations in Computed Tomography (CT) image formation parameters is crucial for clinical adoption.

Purpose of the Study:

• To assess the impact of common CT image formation parameters on the performance and stability of a DL-based CAD assessment algorithm.
• To identify specific parameters that introduce instability in automated CAD severity grading.

Main Methods:

• Reconstructed 500 CCTA scans using seven different image formation parameter configurations.
• Evaluated a DL algorithm's performance and stability by varying denoising strength, slab combination, and reconstruction kernel.
• Assessed stability by measuring the standard deviation of CAD-RADS grade differences between default and varied configurations.

Main Results:

• The full DL pipeline exhibited slight instability (± 0.226 CAD-RADS) across parameter variations.
• Propagating centerlines from a default configuration to others improved stability (± 0.122 CAD-RADS), especially with varied denoising (± 0.046 CAD-RADS).
• True stack reconstruction (± 0.313 CAD-RADS) and sharper kernels (± 0.150 CAD-RADS) led to more unstable predictions. Clinical performance for ruling out CAD remained high (AUC > 0.937).

Conclusions:

• CT image reconstruction parameters influence the stability of DL-based CAD assessment algorithms.
• Scans reconstructed with the 'true stack' parameter require caution when interpreted by DL methods.
• Underrepresented reconstruction kernels in training data can increase prediction uncertainty.