Related Experiment Video
Updated: Sep 20, 2025
04:25
Author Spotlight: Bridging Gaps in Anatomy and Establishing a Foundation for Algorithmic Studies
Published on: December 15, 2023
2.9K
A Prior Knowledge-Guided, Deep Learning-Based Semiautomatic Segmentation for Complex Anatomy on Magnetic Resonance
Ying Zhang1, Ying Liang1, Jie Ding1
1Department of Radiation Oncology, Medical College of Wisconsin, Milwaukee, Wisconsin.
Summary
This study introduces a deep learning-guided semiautomatic segmentation (DL-SAS) method for abdominal MRI scans. The DL-SAS model significantly improves segmentation accuracy and efficiency for complex structures, reducing manual editing needs.
Area of Science:
- Medical Imaging
- Radiology
- Artificial Intelligence in Medicine
Background:
- Manual segmentation of abdominal organs on MRI is labor-intensive and time-consuming, even with deep learning (DL) advancements.
- Complex anatomical structures present particular challenges for existing autosegmentation techniques.
- Accurate segmentation is crucial for applications like MRI-guided online adaptive radiation therapy.
Purpose of the Study:
- To develop a fast, prior knowledge-guided deep learning semiautomatic segmentation (DL-SAS) method.
- To segment complex abdominal structures on MRI scans efficiently and accurately.
- To reduce the need for extensive manual slice-by-slice editing in abdominal MRI segmentation.
Main Methods:
- Implemented a novel DL-SAS approach using a 2D UNet model that leverages contours from adjacent slices as prior knowledge.
- Trained and tested a generalized DL-SAS model on T2-weighted abdominal MRI scans from 75 patients.
- Compared DL-SAS performance against linear interpolation, rigid propagation, and a 3D DL autosegmentation model using Dice Similarity Coefficient (DSC) and Ratio of Acceptable Slices (ROA).
Main Results:
- The DL-SAS model achieved the highest performance with a slice interval of 1, demonstrating average DSCs ranging from 0.87 to 0.93 across different abdominal organs.
- Performance decreased with increasing slice intervals, indicating the importance of adjacent slice information.
- DL-SAS significantly outperformed other methods (P < .05) with higher ROA values (48%-66%) compared to linear interpolation (31%-57%) and 3D DL autosegmentation (16%-51%) at SI=1.
Conclusions:
- The developed DL-SAS model accurately and efficiently segments complex abdominal structures on MRI.
- This method can serve as an interactive tool or contour editor, complementing full autosegmentation.
- DL-SAS facilitates faster and more accurate segmentation, supporting MRI-guided online adaptive radiation therapy.
