A Unified Framework for Multi-Guidewire Endpoint Localization in Fluoroscopy Images

Area of Science:

• Medical Imaging
• Computer Vision
• Deep Learning

Background:

• Accurate localization of surgical instruments like guidewires is crucial for minimally invasive procedures.
• Existing methods may lack the precision or unified approach needed for complex tasks such as multi-guidewire endpoint localization.

Purpose of the Study:

• To propose a novel unified model, Keypoint Localization Region-based CNN (KL R-CNN), for simultaneous guidewire detection and endpoint localization.
• To introduce a new evaluation metric, AP^PCK, for assessing multi-guidewire endpoint localization performance.

Main Methods:

• KL R-CNN modifies the Mask R-CNN architecture by replacing the mask branch with a specialized keypoint localization branch.
• The model incorporates modified settings to enhance the detail level of keypoint localization results.
• A new metric, AP^PCK, is developed, combining Average Precision (AP) and Percentage of Correct Keypoints (PCK) for intuitive performance evaluation.

Main Results:

• KL R-CNN demonstrates superior performance compared to existing methods, particularly under relaxed thresholds.
• The model achieved a mean AP^PCK of 90.65% with a threshold of 9 pixels.
• This indicates high accuracy and reliability in localizing multiple guidewire endpoints.

Conclusions:

• KL R-CNN establishes state-of-the-art performance in the challenging task of multi-guidewire endpoint localization.
• The model shows significant potential for application in computer-assisted percutaneous coronary intervention.
• The KL R-CNN framework is adaptable for other multi-instrument localization tasks in medical imaging.