Weakly- and Semisupervised Probabilistic Segmentation and Quantification of Reverberation Artifacts.
Alex Ling Yu Hung1, Edward Chen2, John Galeotti1,2
1Department of Biomedical Engineering, Carnegie Mellon University, Pittsburgh, PA, USA.
BME Frontiers
|October 18, 2023
Summary
This study introduces a new algorithm for segmenting ultrasound artifacts caused by needles. The method improves image analysis by accurately separating artifact pixels from tissue pixels.
Area of Science:
- Medical Imaging
- Computer Vision
- Machine Learning
Background:
- Reverberation artifacts from needles degrade ultrasound image quality.
- These artifacts are challenging for current computer vision algorithms.
- Artifact boundaries are ambiguous, causing expert labeling disagreements.
Purpose of the Study:
- To develop a weakly- and semisupervised algorithm for segmenting needle and reverberation artifacts.
- To separate tissue pixel values from superimposed artifacts.
- To model artifact intensity decay and minimize human labeling error.
Main Methods:
- A three-part learning-based framework was employed.
- A probabilistic segmentation network generated soft labels from human inputs.
- A transform function and subsequent network generated final artifact masks.
Main Results:
- The algorithm differentiates between artifact-free and artifact regions.
- It accurately models the intensity fall-off within artifacts.
- Performance was compared favorably against other segmentation algorithms.
Conclusions:
- The method achieves state-of-the-art artifact segmentation performance.
- It provides a new standard for estimating per-pixel artifact contributions.
- The algorithm enhances downstream medical image analysis tasks.
