CoroSAM: adaptation of the Segment Anything Model for interactive segmentation in Coronary angiograms

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiovascular Diagnostics

Background:

• X-ray Coronary Angiography (XCA) is vital for assessing coronary artery disease and morphology.
• Accurate segmentation of major coronary vessels is essential for automated geometric analysis but challenging due to complex anatomy.
• Existing methods struggle with the intricacies of coronary artery visualization.

Purpose of the Study:

• To introduce CoroSAM, an adapted LiteMedSAM Foundation Model for interactive coronary artery segmentation in XCA images.
• To enhance domain-specific feature extraction using parameter-efficient fine-tuning.
• To improve the accuracy and efficiency of automated coronary artery analysis.

Main Methods:

• Utilized Convolutional Adapter layers within TinyViT blocks for efficient, domain-specific feature extraction.
• Implemented a point-based prompting strategy encoding vessel endpoints and branch points.
• Evaluated performance using 5-fold cross-validation on the ARCADE dataset and zero-shot testing on XCAD and DCA1 datasets.

Main Results:

• CoroSAM achieved superior performance on the ARCADE test set (Dice=0.87, Precision=0.86, Recall=0.89) with fewer trainable parameters.
• Demonstrated significant improvements over alternative Adapter configurations.
• Showcased robust transferability and competitive zero-shot performance on external datasets (XCAD: Dice=0.82; DCA1: Dice=0.73).

Conclusions:

• CoroSAM accurately delineates major coronary vessels with minimal user intervention through specialized adapters and point prompts.
• The architecture enables efficient inference on standard hardware, making it practical for clinical applications.
• This framework effectively balances segmentation accuracy and computational efficiency for routine analysis.