Related Experiment Video
Updated: Jul 23, 2025
03:31
Author Spotlight: Enhancement of Salient Object Detection for Smart Grid Applications
Published on: December 15, 2023
581
Multi-target landmark detection with incomplete images via reinforcement learning and shape prior embedding
Kaiwen Wan1, Lei Li2, Dengqiang Jia3
1School of Data Science, Fudan University, Shanghai, 200433, China.
Medical Image Analysis
|July 13, 2023
Summary
This study introduces a multi-agent reinforcement learning (RL) framework for robust multi-target landmark detection in medical images, even with missing data. The method effectively learns global structure from incomplete images, improving landmark detection accuracy.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Computer Vision
Background:
- Medical images often have limited fields-of-view (FOV), leading to incomplete regions of interest (ROI) and challenges in analysis.
- Learning-based multi-target landmark detection algorithms struggle with varying FOV, often misinterpreting background variations and failing target detection.
Purpose of the Study:
- To propose a novel multi-agent reinforcement learning (RL) framework for simultaneous multi-target landmark detection.
- To enable accurate landmark detection from incomplete or complete medical images by learning implicit global structure.
- To enhance robustness by explicitly incorporating shape models into the RL process for incomplete image analysis.
Main Methods:
- Developed a multi-agent RL framework for simultaneous multi-target landmark detection.
- Trained the framework to learn implicit global structure from both complete and incomplete images.
- Integrated a shape model into the RL process to explicitly leverage global structural information from incomplete data.
Main Results:
- The proposed RL model successfully localized dozens of targets simultaneously and demonstrated robustness with incomplete images.
- Validated on body dual-energy X-ray absorptiometry (DXA), cardiac MRI, and head CT datasets.
- Achieved high accuracy even with up to 80% missing training image proportions (e.g., 2.29 cm average distance error on body DXA) and detected unseen landmarks in missing FOV regions (6.84 mm average distance error on 3D half-head CT).
Conclusions:
- The multi-agent RL framework effectively addresses challenges in medical image analysis caused by limited FOV and incomplete data.
- The integration of shape models further improves the model's ability to exploit global structural information for robust landmark detection.
- The method shows significant promise for clinical applications involving multi-target landmark detection in various medical imaging modalities.
