Related Experiment Video
Updated: Jan 11, 2026
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
Weakly-supervised ultrasound image segmentation with elliptical shape prior constraint.
Changyan Wang1, Yehua Cai2, Ruyi Yang3
1The SMART (Smart Medicine and AI-based Radiology Technology) Lab, School of Communication and Information Engineering, Shanghai University, Shanghai, China; Shanghai Institute of Advanced Communication and Data Science, Key Laboratory of Specialty Fiber Optics and Optical Access Networks, Shanghai University, Shanghai, China.
This study introduces Elliptical Shape Prior Constraint Multiple Instance Learning (ESPC-MIL), a novel weakly supervised method for ultrasound image segmentation. ESPC-MIL improves accuracy for elliptical shapes, matching fully supervised methods with less annotation.
Area of Science:
- Medical Imaging
- Computer-Aided Diagnosis
- Machine Learning
Background:
- Accurate pixel-level segmentation of ultrasound (US) images is crucial for computer-aided disease screening and diagnosis.
- Weakly supervised methods, particularly multiple instance learning (MIL), offer potential to reduce radiologist workload by requiring less labeled data.
- Radiologists use elliptical annotations in US examinations, providing valuable prior information for segmentation tasks.
Purpose of the Study:
- To propose a novel weakly supervised method, Elliptical Shape Prior Constraint MIL (ESPC-MIL), for pixel-level segmentation of ultrasound images.
- To leverage elliptical shape prior information within the MIL framework to enhance segmentation accuracy, especially for elliptical anatomical structures.
- To improve edge segmentation and localization accuracy through global supervision using elliptical shape priors.
Main Methods:
- Developed ESPC-MIL, integrating an elliptical shape prior constraint into the MIL framework.
- Utilized elliptical shape priors to generate more accurate foreground/background candidate regions for MIL.
- Applied elliptical shape prior information for global supervision to refine edge segmentation.
Main Results:
- ESPC-MIL achieved state-of-the-art results on four diverse US image datasets (Achilles tendon, median nerve, breast tumor), with Dice scores ranging from 0.748 to 0.876.
- The method demonstrated performance comparable to fully supervised segmentation approaches, significantly reducing annotation effort.
- ESPC-MIL showed a more pronounced performance improvement for segmenting objects with approximately elliptical shapes compared to those with complex shapes.
Conclusions:
- ESPC-MIL is an effective weakly supervised method for pixel-level segmentation of ultrasound images, particularly for elliptical structures.
- The integration of elliptical shape priors enhances segmentation accuracy and reduces annotation requirements, making it a valuable tool for quantitative US image analysis.
- The developed method shows promise for automating quantitative analysis in medical imaging, aiding in disease screening and diagnosis.
