Related Experiment Videos
Computerized analysis of mammographic microcalcifications in morphological and texture feature spaces
1Department of Radiology, University of Michigan, Ann Arbor 48109, USA. chanhp@umich.edu
Medical Physics
|November 4, 1998
Summary
Computerized analysis of mammograms uses genetic algorithms to select features for classifying microcalcifications. Combined texture and morphological features improve accuracy, aiding in early detection of malignant cases.
Area of Science:
- Medical Imaging
- Computer-Aided Diagnosis
- Machine Learning
Background:
- Microcalcifications on mammograms can indicate breast cancer.
- Accurate classification of malignant versus benign microcalcifications is crucial for diagnosis.
- Computerized analysis offers potential for objective and efficient microcalcification assessment.
Purpose of the Study:
- To develop and evaluate computerized methods for extracting and classifying malignant and benign microcalcifications.
- To compare feature selection techniques, specifically genetic algorithms (GA) versus stepwise linear discriminant analysis (LDA).
- To assess the effectiveness of morphological and texture features, individually and combined, for microcalcification classification.
Main Methods:
- Extraction of morphological features (size, contrast, shape) and texture features (spatial gray-level dependence matrices).
- Feature selection using a genetic algorithm (GA) and comparison with stepwise linear discriminant analysis (LDA).
- Classification using linear discriminant classifiers and evaluation via receiver operating characteristic (ROC) analysis (Area under the curve, Az).
Main Results:
- GA-based feature selection yielded results comparable or superior to stepwise LDA.
- Texture features (Az = 0.84) outperformed morphological features (Az = 0.79) for distinguishing microcalcification types.
- Combined texture and morphological features achieved the highest classification accuracy (Az = 0.89), a statistically significant improvement.
- Using averaged discriminant scores across multiple views increased accuracy to Az = 0.93, identifying 50% of benign clusters at 100% sensitivity for malignancy.
Conclusions:
- Combined morphological and texture features significantly enhance computer-aided classification of microcalcifications.
- Genetic algorithms are effective for selecting optimal feature subsets in this context.
- The developed methods show strong potential for improving the accuracy and efficiency of breast cancer diagnosis through mammogram analysis.