Related Experiment Video
Updated: May 2, 2026
08:25
Combining Eye-tracking Data with an Analysis of Video Content from Free-viewing a Video of a Walk in an Urban Park Environment
Published on: May 7, 2019
8.6K
Self-assessed performance improves statistical fusion of image labels
Frederick W Bryan1, Zhoubing Xu1, Andrew J Asman1
1Electrical Engineering, Vanderbilt University, Nashville, Tennessee 37235.
Medical Physics
|March 6, 2014
Summary
Self-assessed confidence in manual labeling, combined with statistical fusion, significantly improves image segmentation accuracy. This approach offers a robust alternative to expert manual labeling for complex structures like the spinal cord.
Area of Science:
- Medical Imaging
- Computational Anatomy
- Machine Learning
Background:
- Expert manual labeling is the gold standard for image segmentation but is time-consuming and subjective.
- Automated methods exist but are not developed for all essential anatomical structures, such as internal spinal cord structures in MRI.
- Collaborative labeling offers a potential solution by combining automation throughput with expert guidance.
Purpose of the Study:
- To explore the utility of self-assessed confidence in manual labeling within a collaborative labeling framework.
- To investigate the integration of self-assessment with automated assessment of rater performance.
- To develop and evaluate methods for fusing information from multiple raters and self-assessments.
Main Methods:
- A study involving 75 minimally trained undergraduate raters labeling 66 MRI volumes.
- Raters received 15 minutes of training on segmentation tools and techniques.
- A self-assessed quality metric was collected per slice using a confidence bar.
- Volumes were segmented using simple majority voting and statistical fusion, compared against expert segmentations.
Main Results:
- Self-assessed weighted voting outperformed simple majority voting.
- Statistical fusion performance was comparable to self-assessed weighted voting.
- A novel theoretical basis for incorporating self-assessment into statistical fusion was developed.
- Combining statistical assessment with self-assessment significantly improved performance over individual methods.
Conclusions:
- This study provides the first systematic characterization of self-assessed performance in manual labeling.
- Self-assessment and statistical fusion offer complementary benefits for label fusion in medical imaging.
- A new theoretical framework for combining self-assessments with statistical label fusion was presented.
