Aortography Keypoint Tracking for Transcatheter Aortic Valve Implantation Based on Multi-Task Learning.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiovascular Surgery

Background:

• Transcatheter aortic valve implantation (TAVI) is a key treatment for aortic stenosis, but accurate valve positioning is challenging with current imaging.
• Limitations in perioperative imaging necessitate advanced visual assistance systems for TAVI procedures.

Purpose of the Study:

• To develop a multi-task learning algorithm for real-time tracking of anatomical landmarks and critical keypoints.
• To enhance the accuracy and efficiency of valve positioning during TAVI procedures.

Main Methods:

• Proposed an original multi-task learning-based algorithm for landmark tracking and keypoint labeling.
• Designed and tested nine neural networks (MobileNet V2, ResNet V2, Inception V3, Inception ResNet V2, EfficientNet B5) to predict 11 keypoints.
• Evaluated prediction accuracy, time efficiency, and mean absolute error for keypoint localization.

Main Results:

• ResNet V2 and MobileNet V2 architectures demonstrated the best accuracy/time ratio.
• Achieved 97% and 96% accuracy in predicting keypoint labels and coordinates, respectively.
• Reported mean absolute errors of 4.7% and 5.6% for ResNet V2 and MobileNet V2, respectively.

Conclusions:

• Neural networks, particularly ResNet V2 and MobileNet V2, can perform real-time predictions for aortic valve and delivery system localization.
• This AI-driven approach has the potential to significantly improve valve positioning accuracy in TAVI.
• The developed algorithm offers a promising visual assistance tool for enhancing TAVI procedural outcomes.