Related Experiment Video
Updated: Mar 3, 2026
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
3.6K
gcDLSeg: integrating graph-cut into deep learning for binary semantic segmentation
Hui Xie1, Weiyu Xu1, Ya Xing Wang2
1Department of Electrical and Computer Engineering, The University of Iowa, Iowa City, IA 52242, USA.
Biomedical Optics Express
|March 2, 2026
Summary
This study integrates graph-cut segmentation with deep learning (DL) for improved computer vision. The novel approach enables end-to-end learning, achieving optimal segmentation accuracy and robustness.
Area of Science:
- Computer Vision
- Medical Image Analysis
- Machine Learning
Background:
- Binary semantic segmentation is crucial in computer vision.
- Graph-cut methods offer global optimality but lack deep integration.
- Deep learning (DL) methods have revolutionized segmentation performance.
Purpose of the Study:
- To integrate graph-cut segmentation within a deep learning network for end-to-end learning.
- To address the challenge of backpropagation through the combinatorial graph-cut algorithm.
- To leverage the strengths of both graph-cut and DL for enhanced segmentation.
Main Methods:
- Developed a novel residual graph-cut loss function.
- Introduced a quasi-residual connection for enabling gradient backpropagation.
- Integrated graph-cut energy minimization with DL-optimized image features.
Main Results:
- Achieved promising segmentation accuracy on chronic wound and pancreas cancer datasets.
- Demonstrated improved robustness against adversarial attacks.
- Enabled effective feature learning guided by the graph-cut segmentation model.
Conclusions:
- The proposed integrated approach successfully combines graph-cut and DL for binary semantic segmentation.
- The novel loss and connection facilitate end-to-end training and globally optimal inference.
- This method offers a robust and accurate solution for medical image segmentation tasks.
