Evaluating reinforcement learning agents for anatomical landmark detection.

Area of Science:

• Medical Image Analysis
• Artificial Intelligence
• Machine Learning

Background:

• Manual annotation of anatomical landmarks in medical scans is time-consuming and error-prone.
• Accurate landmark detection is crucial for various medical image analysis applications.

Purpose of the Study:

• To evaluate novel deep reinforcement learning (RL) strategies for precise and robust automatic anatomical landmark localization.
• To investigate the efficacy of fixed- and multi-scale search strategies with hierarchical action steps.

Main Methods:

• Trained artificial RL agents using deep Q-network (DQN) architectures to navigate 3D medical images and identify landmarks.
• Employed coarse-to-fine search strategies, including fixed-scale, multi-scale, and novel hierarchical action steps.
• Evaluated agents on fetal head ultrasound (US), adult brain MRI, and cardiac MRI datasets.

Main Results:

• The developed RL agents surpassed state-of-the-art supervised and RL methods in landmark detection performance.
• Multi-scale search strategies demonstrated superior performance over fixed-scale agents, especially in challenging imaging conditions (e.g., cardiac MRI).
• Novel hierarchical action steps accelerated the landmark searching process by 4-5 times.

Conclusions:

• Deep reinforcement learning offers a powerful and efficient approach for automated anatomical landmark detection in medical imaging.
• Multi-scale and hierarchical search strategies enhance the robustness and speed of RL-based landmark localization.
• This work provides a foundation for more advanced automated analysis in medical imaging applications.