Pretrained subtraction and segmentation model for coronary angiograms.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Scarcity of annotated medical data hinders AI development in medical imaging.
• Accurate coronary vessel segmentation is crucial for diagnosing conditions like stenosis.

Purpose of the Study:

• Introduce a novel self-supervised learning method for single-frame subtraction and vessel segmentation in coronary angiography.
• Address the challenge of limited annotated samples in AI applications for medical imaging.

Main Methods:

• Pretrain a U-Net model on unannotated coronary angiograms using image-to-image translation.
• Fine-tune the pretrained model on a small set of manually annotated samples.
• Utilize a self-supervised learning approach to leverage large unannotated datasets.

Main Results:

• Achieved comprehensive single-frame subtraction, outperforming existing Digital Subtraction Angiography (DSA) methods.
• Vessel segmentation reached a Dice coefficient of 0.828 with only 40 annotated samples.
• Set a new state-of-the-art benchmark on the XCAD dataset with a Dice coefficient of 0.755.

Conclusions:

• Combining self-supervised pretraining with minimal fine-tuning enables accurate coronary vessel segmentation.
• The developed method provides robust single-frame subtraction for coronary angiography.
• This approach assists physicians in identifying potential vascular stenosis sites.