Related Experiment Video
Updated: Aug 14, 2026
07:15
Machine Learning Algorithms for Early Detection of Bone Metastases in an Experimental Rat Model
Published on: August 16, 2020
Robustness of computerized lesion detection and classification scheme across different breast US platforms
Karen Drukker1, Maryellen L Giger, Charles E Metz
1Department of Radiology MC2026, University of Chicago, 5841 S Maryland Ave, Chicago, IL 60637, USA. kdrukker@uchicago.edu
Radiology
|November 24, 2005
Summary
This study shows a promising automated method for detecting and classifying breast lesions in ultrasound images, performing consistently across different equipment. The computerized system accurately distinguishes actual lesions and identifies potential cancers.
Area of Science:
- Medical Imaging
- Artificial Intelligence in Healthcare
- Breast Cancer Diagnostics
Background:
- Accurate detection and diagnosis of breast lesions using ultrasonography (US) are crucial for effective breast cancer management.
- Computerized methods offer potential for improving the efficiency and accuracy of breast US image analysis.
- Variability in US equipment can impact the performance of diagnostic algorithms.
Purpose of the Study:
- To evaluate the performance of an automated computerized method for lesion detection and diagnosis in breast ultrasonographic images.
- To assess the method's consistency across images acquired with US equipment from two different manufacturers (Philips and Siemens).
Main Methods:
- Utilized two independent clinical breast US databases comprising images from Philips (1740 images) and Siemens (151 images) equipment.
- Employed a computerized scheme for lesion detection, feature calculation, and classification into distinct categories.
- Evaluated two classification tasks: lesion vs. false positives and cancer vs. other lesions, using receiver operating characteristic (ROC) and free-response ROC analyses.
Main Results:
- Area under the ROC curve (A(z)) values for distinguishing actual lesions from false positives ranged from 0.87 to 0.95.
- For cancer detection, A(z) values ranged from 0.80 to 0.86, with no significant difference across testing protocols.
- Performance differences between databases were attributed to database size, not equipment type, indicating robustness.
Conclusions:
- The fully automated computerized method for breast lesion detection and classification demonstrates promising performance.
- The system exhibits robustness and consistent performance across different ultrasonographic equipment, suggesting broad applicability.
- This technology holds potential for enhancing diagnostic accuracy in breast ultrasonography.