Fully Automated Anatomy Labeling for Intracardiac Echocardiography Using Deep Learning.

Area of Science:

• Cardiology
• Medical Imaging
• Artificial Intelligence

Background:

• Intracardiac echocardiography (ICE) is vital for guiding electrophysiologic (EP) procedures.
• A significant learning curve exists for mastering ICE interpretation.
• Automated tools are needed to enhance efficiency and accuracy in EP procedures.

Purpose of the Study:

• To develop and validate an automated deep learning algorithm for cardiac anatomy detection in ICE images.
• To assess the algorithm's precision and recall in identifying key anatomic structures.
• To evaluate the potential of the algorithm as an educational or navigational aid.

Main Methods:

• Deep learning model trained on 196,768 ICE images from 605 procedures across two institutions.
• Algorithm designed to automatically detect and label anatomic structures within the right atrium.
• Performance evaluated using precision and recall metrics for 21 distinct anatomic structures.

Main Results:

• The algorithm achieved >70% precision and recall for 15 out of 21 identified anatomic structures.
• Accurate identification of cardiac anatomy was demonstrated.
• Instances of mislabeling one structure for another were infrequent.

Conclusions:

• A fully automated deep learning algorithm can reliably detect cardiac anatomy from ICE images.
• This algorithm shows promise as an educational tool for trainees.
• The tool can potentially serve as a real-time navigation aid for ICE operators during EP procedures.