Related Experiment Video
Updated: Aug 11, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
2.9K
Surface-GCN: Learning interaction experience for organ segmentation in 3D medical images
Fengrui Tian1,2, Zhiqiang Tian1, Zhang Chen1
1School of Software Engineering, Xi'an Jiaotong University, Xi'an, China.
Medical Physics
|February 4, 2023
Summary
This study introduces Surface-GCN, a Graph Convolutional Network for medical organ segmentation, which uses adjacent slice information and radiologist interaction to improve accuracy. The method achieves state-of-the-art results in prostate and abdominal organ segmentation while reducing manual effort.
Area of Science:
- Medical Imaging
- Computer Vision
- Machine Learning
Background:
- Accurate organ segmentation is crucial for clinical diagnosis but challenging due to unclear boundaries in medical images.
- Leveraging image continuity across adjacent slices can infer segmentation from clearer boundaries.
- Radiologists utilize adjacent slice information to delineate organ boundaries effectively.
Purpose of the Study:
- To develop an organ segmentation model inspired by radiologists' adjacent slice analysis.
- To incorporate a human-machine interactive learning strategy to integrate clinical expertise.
- To improve the accuracy and efficiency of medical organ segmentation in image volumes.
Main Methods:
- Proposed Surface-GCN, an interactive organ segmentation method for medical image volumes based on Graph Convolutional Networks.
- Developed Surface Feature Extraction Network (SFE-Net) supervised by Mini-batch Adaptive Surface Matching (MBASM).
- Designed an automatic segmentation module with Surface Convolution Unit (SCU) and an interactive module to learn from radiologist corrections, reducing clicks.
Main Results:
- Evaluated Surface-GCN on prostate MR and abdominal CT datasets, outperforming state-of-the-art methods.
- Achieved a 94.49% Dice Similarity Coefficient (DSC) for prostate segmentation on the PROMISE12 test dataset.
- Obtained high DSC scores for abdominal organs (e.g., 95% for left kidney, 91% for gallbladder) and reduced interaction clicks by 5-10 for equivalent accuracy.
Conclusions:
- Surface-GCN effectively overcomes medical organ segmentation challenges by imitating radiologist interactions and learning clinical experience.
- The proposed Graph Convolutional Network achieves more accurate segmentation boundaries for single and multiple organs compared to existing methods.
- The interactive approach enhances segmentation accuracy and efficiency, demonstrating the value of integrating human expertise into AI models.
