Related Experiment Video
Updated: Jan 10, 2026
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
3.3K
Small-object-sensitive deep reinforcement learning for fully automatic 3D vessel segmentation in medical images.
Jingliang Zhao1, Xianyang Lin1, An Zeng1
1School of Computer Science, Guangdong University of Technology, Guangzhou, People's Republic of China.
Biomedical Physics & Engineering Express
|November 25, 2025
Summary
This study introduces a new deep reinforcement learning method for 3D vessel segmentation in medical images. The approach improves accuracy and efficiency for better clinical navigation and evaluation.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Computer Vision
Background:
- Accurate 3D vessel segmentation is crucial for clinical applications like intraoperative navigation and postoperative evaluation.
- Fully automatic segmentation faces challenges due to the imbalanced proportion of vessels in medical images, often leading to target loss.
- Existing methods struggle with precise segmentation of small or partially obscured vessels.
Purpose of the Study:
- To present a fully automatic 3D vessel segmentation method using small-object-sensitive deep reinforcement learning (DRL) and convolutional neural networks (CNNs).
- To enhance DRL-based detection by improving information acquisition, prioritizing complete region detection, and correcting errors using inter-slice dependencies.
- To achieve high accuracy and topological integrity in 3D vessel segmentation for improved clinical utility.
Main Methods:
- A hybrid approach combining DRL for initial bounding box detection and CNN for precise segmentation.
- Three key improvements to the DRL network: random receptive field expansion for robust detection, recall-priority reward for complete region extraction, and utilization of inter-slice vascular spatial dependencies for error correction.
- Validation on a dataset of 100 computed tomography angiography (CTA) scans.
Main Results:
- The proposed method achieved a Dice coefficient of 93.75% for vessel segmentation accuracy.
- Demonstrated superior performance compared to baseline and other automatic 3D vessel segmentation algorithms.
- Showcased advantages in positioning accuracy, segmentation accuracy, and operational efficiency.
Conclusions:
- The developed method offers a robust and accurate solution for fully automatic 3D vessel segmentation.
- The improvements in DRL enhance the detection of small vessels and maintain topological integrity.
- The method's accuracy, efficiency, and ease of application make it suitable for clinical practice.
