Deep Reinforcement Learning with Explicit Spatio-Sequential Encoding Network for Coronary Ostia Identification in CT...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiovascular Imaging

Background:

• Accurate identification of coronary ostia is crucial for automated analysis of coronary arteries in 3D CCTA.
• Existing methods may lack efficiency or accuracy in localizing these critical anatomical landmarks.

Purpose of the Study:

• To propose and evaluate a novel deep reinforcement learning (DRL) framework for precise localization of coronary ostia from 3D CCTA.
• To enhance automated coronary artery tracking and segmentation by improving the initial ostia identification step.

Main Methods:

• A DRL framework employing a spatial-sequential encoding policy network with 2.5D Markovian states and three past histories.
• Training was conducted using a dueling DRL framework on the CAT08 dataset.
• Computational efficiency was measured using floating-point operations (FLOPs).

Main Results:

• The proposed method achieved 2.5M FLOPs, approximately 10 times fewer than 3D box-based methods, indicating superior computational efficiency.
• Accuracy results showed mean errors of 2.22 ± 1.12 mm for the left coronary ostium and 1.94 ± 0.83 mm for the right coronary ostium.
• The DRL framework demonstrated higher efficiency and accuracy compared to existing methods.

Conclusions:

• The developed DRL framework provides an efficient and accurate solution for coronary ostia localization in 3D CCTA.
• This method can be adapted for identifying other anatomical targets and serves as a valuable pre-processing step for coronary artery tracking.