Coronary artery segmentation in non-contrast calcium scoring CT images using deep learning

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiovascular Imaging

Background:

• Accurate coronary artery segmentation is vital for diagnosing heart conditions.
• Non-contrast CT scans offer a less invasive imaging option but present segmentation challenges due to low visibility of fine structures.
• Existing methods for non-contrast CT often yield high recall but low precision, limiting their application beyond calcium scoring.

Purpose of the Study:

• To develop and evaluate a deep learning algorithm for segmenting coronary arteries in multi-vendor, ECG-gated, non-contrast cardiac CT images.
• To introduce a novel semi-automatic framework for generating Ground Truth (GT) data using image registration, aiming for increased efficiency and data diversity.
• To propose a new method for manual mesh-to-image registration for creating a robust test-GT dataset.

Main Methods:

• A deep learning algorithm was developed using an AutoML framework.
• A semi-automatic GT generation process leveraging image registration was implemented for efficient data creation.
• A novel manual mesh-to-image registration technique was employed to establish a high-quality test-GT dataset for evaluation.

Main Results:

• The proposed deep learning model demonstrated significantly more accurate delineation of coronary arteries compared to the training GT.
• The segmentation performance achieved Dice and clDice metrics comparable to interrater variability.
• The semi-automatic GT generation framework proved efficient in creating large, diverse datasets for model training.

Conclusions:

• The developed deep learning algorithm effectively segments coronary arteries in non-contrast cardiac CT, addressing a significant research gap.
• The novel semi-automatic GT generation method enhances efficiency and data diversity, leading to well-generalizing models.
• This approach holds promise for improving the medical assessment of heart pathologies using less invasive CT imaging.