Related Experiment Videos
Segmentation of mammograms using multiple linked self-organizing neural networks
J Suckling1, D R Dance, E Moskovic
1Joint Department of Physics, Institute of Cancer Research, London, United Kingdom.
Medical Physics
|February 1, 1995
Summary
This study presents an algorithm for segmenting mammograms into background, pectoral muscle, fibroglandular, and adipose regions. The automated method shows promising results comparable to expert radiologists for fibroglandular region analysis.
Area of Science:
- Medical Imaging
- Computer Vision
- Biomedical Engineering
Background:
- Mammographic analysis requires accurate segmentation of key anatomical regions.
- Automated segmentation can improve efficiency and consistency in mammogram interpretation.
Purpose of the Study:
- To develop and evaluate an algorithm for segmenting mammograms into four main components.
- To assess the performance of the algorithm against expert human segmentation.
Main Methods:
- An algorithm utilizing self-organizing neural networks and texture analysis was developed.
- The algorithm classifies mammographic regions based on statistical texture measures.
- A staged approach incorporating geometric information addresses inter-mammogram variability.
Main Results:
- The algorithm achieved correlation coefficients of 0.74 for area and 0.59 for perimeter when comparing segmented parenchyma to radiologist outlines.
- Normalized overlapping areas between automated and human segmentation yielded a mean of 0.69 +/- 0.12.
Conclusions:
- The developed algorithm demonstrates a viable method for mammographic scene segmentation.
- The automated segmentation shows good agreement with expert radiologist delineations, particularly for the fibroglandular region.