Related Experiment Videos
Comparison of supervised MRI segmentation methods for tumor volume determination during therapy
M Vaidyanathan1, L P Clarke, R P Velthuizen
1Department of Radiology, University of South Florida, Tampa, USA.
Magnetic Resonance Imaging
|January 1, 1995
Summary
Two pattern recognition methods, k-nearest neighbor (kNN) and semi-supervised fuzzy c-means (SFCM), accurately segment brain magnetic resonance images (MRI) for tumor volume analysis. SFCM shows promise for tracking tumor changes during therapy.
Area of Science:
- Medical Imaging
- Computational Biology
- Pattern Recognition
Background:
- Accurate quantitative estimation of brain tumor volume is crucial for treatment monitoring.
- Existing segmentation methods may lack reproducibility and accuracy, particularly for dynamic volume changes.
- Multispectral pattern recognition offers potential for improved image segmentation in neuro-oncology.
Purpose of the Study:
- To evaluate two multispectral pattern recognition methods, k-nearest neighbor (kNN) and semi-supervised fuzzy c-means (SFCM), for segmenting brain MRI.
- To compare the reproducibility and accuracy of kNN and SFCM against reference methods (seed growing and manual segmentation) for tumor volume estimation.
- To assess the utility of these methods for measuring tumor volume changes during therapy.
Main Methods:
- Supervised k-nearest neighbor (kNN) and semi-supervised fuzzy c-means (SFCM) algorithms were applied to segment magnetic resonance imaging (MRI) data.
- Segmentation results were compared against a seed growing method and manual segmentation on contrast-enhanced T1-weighted images.
- Intra- and inter-observer reproducibility was rigorously assessed for all segmentation techniques.
Main Results:
- SFCM demonstrated superior reproducibility (6% intra-observer, 4% inter-observer) compared to kNN (9% intra-observer, 5% inter-observer) and seed growing (6% intra-observer, 17% inter-observer).
- Multispectral methods consistently yielded smaller absolute tumor volumes than reference methods.
- Results indicated patient case-dependency, highlighting the need for advanced segmentation approaches.
Conclusions:
- SFCM shows potential for reliable relative measurements of tumor volume during therapy, though further validation is needed.
- Minimally supervised or unsupervised methods are essential for robust tumor volume measurements in clinical practice.
- Multispectral pattern recognition techniques offer improved reproducibility for brain tumor segmentation in MRI.