Related Experiment Video
Updated: Apr 6, 2026
06:45
Automated Joint Space Detection Improves Bone Segmentation Accuracy
Published on: November 28, 2025
280
Weighting training images by maximizing distribution similarity for supervised segmentation across scanners
Annegreet van Opbroek1, Meike W Vernooij2, M Arfan Ikram2
1Biomedical Imaging Group Rotterdam, Departments of Medical Informatics and Radiology, Erasmus MC - University Medical Center Rotterdam, Rotterdam 3000 CA, The Netherlands.
Medical Image Analysis
|July 27, 2015
Summary
This study introduces a novel transfer learning method for image segmentation, improving accuracy by weighting training data. The approach effectively reduces classification errors, outperforming traditional methods even with diverse imaging sources.
Area of Science:
- Medical Imaging
- Computer Vision
- Machine Learning
Background:
- Supervised machine learning methods for automatic image segmentation require large, representative manually labeled training datasets.
- Acquiring such datasets is challenging due to variations in scanners, scanning parameters, and patient populations.
Purpose of the Study:
- To develop a transfer-learning approach for image segmentation that can utilize heterogeneous training data from different sources.
- To improve segmentation accuracy by addressing the domain shift problem between training and target images.
Main Methods:
- A multi-feature voxelwise classification method using transfer learning.
- Assigning weights to training images based on voxel distribution in feature space to minimize differences between training and target image data distributions.
- Training a weighted classifier using weighted voxels from the training images.
Main Results:
- The proposed weighted classifier significantly outperformed an unweighted classifier across three segmentation tasks (brain-tissue, skull stripping, white-matter lesion segmentation), reducing errors by up to 42%.
- For brain-tissue segmentation and skull stripping, the method surpassed traditional approaches using data from the same study as the target image.
Conclusions:
- The developed transfer-learning method effectively handles heterogeneous training data for improved image segmentation.
- This approach offers a robust solution for segmentation tasks where representative training data is difficult to obtain, enhancing accuracy and reducing classification errors.
