Computer-Aided Mammography Using Automated Feature Extraction for the Detection and Diagnosis of Breast Cancer.
Abstract
We developed artificial neural network (ANN) techniques to predict breast lesion malignancy and invasion based on mammographic features extracted by radiologists and by computerized image processing techniques. We incorporated the radiologist impression as an input to the malignancy-predicting ANN, which outperformed the radiologists. We developed a semi-automated technique for extracting and characterizing breast mass margins, and incorporated those features into an ANN to predict malignancy. In preparation for developing ANNs for feature extraction, we explored the underlying behavior of the previous ANNs by examining their error surfaces in weight space. Finally we developed a novel ANN which predicts invasion among malignant breast lesions based on BI-RADS mammographic findings and patient age. This ANN performed well with Az of 0.91 + or - 0.03. Together these four studies provided important new information which will be crucial toward developing a complete system for computer-aided diagnosis of breast cancer.
Document Details
- Document Type
- Technical Report
- Publication Date
- Oct 01, 1996
- Accession Number
- ADA321765
Entities
People
- Joseph Y. Lo
Organizations
- Duke University Hospital