Related Experiment Video
Updated: May 2, 2026
09:50
In situ Quantification of Pancreatic Beta-cell Mass in Mice
Published on: June 7, 2010
12.2K
An improvement method for pancreas CT segmentation using superpixel-based active contour
Huayu Gao1,2, Jing Li1,2, Nanyan Shen1,2
1Shanghai Key Laboratory of Intelligent Manufacturing and Robotics, Shanghai University, No. 333 Nanchen Road, Baoshan District, Shanghai, 200444, People's Republic of China.
Physics in Medicine and Biology
|April 12, 2024
Summary
A novel superpixel-based active contour model (SbACM) enhances pancreatic segmentation accuracy by acting as a post-processor for deep learning methods. This approach significantly reduces boundary leakage and improves segmentation speed, offering a cost-effective solution for complex medical imaging challenges.
Area of Science:
- Medical Imaging
- Computer Vision
- Artificial Intelligence
Background:
- Pancreatic segmentation in Computed Tomography (CT) images is challenging due to complex organ shapes and fuzzy boundaries.
- Traditional segmentation methods like Active Contour Models (ACM) struggle with boundary leakage and slow evolution speeds.
- Deep learning methods offer powerful segmentation but can benefit from post-processing for refinement.
Purpose of the Study:
- To propose a Superpixel-based Active Contour Model (SbACM) as a post-processor to improve pancreatic segmentation accuracy.
- To address the limitations of traditional ACMs, specifically boundary leakage and slow contour evolution.
- To enhance the performance of various deep learning segmentation models for pancreas imaging.
Main Methods:
- Developed a SbACM using superpixels to guide narrowband and energy function design for edge adhesion.
- Implemented a multi-scale evolution strategy and dynamic narrowband width to improve contour evolution speed and reduce leakage.
- Applied SbACM as a post-processor to coarse segmentation results from deep learning models (e.g., UNet-based architectures).
Main Results:
- SbACM effectively reduced boundary leakage and improved evolution speed through its superpixel-guided narrowband and dynamic energy functions.
- As a post-processor, SbACM increased Dice Similarity Coefficients (DSC) by an average of 2.35% and a maximum of 9.04% across five UNet-based models.
- SbACM outperformed other enhancement techniques on the nnUNet backbone without increasing model complexity or training time.
Conclusions:
- The proposed SbACM is a convenient and effective post-processor for enhancing deep learning-based pancreatic segmentation.
- SbACM significantly improves segmentation accuracy, particularly for challenging cases with fuzzy and complex edges.
- This method offers a low-cost, high-impact solution for improving medical image segmentation accuracy.
