Computerized Identification of Normal Mammograms
Abstract
The purpose of this concept-award project is to develop an automated method to identify normal mammograms, that is those without breast disease. This is a new paradigm in computer-aided diagnosis (CAD), since all other CAD schemes identify breast cancer. We are relying on the natural pattern of glandular tissue in the normal breast, which radiates out from the nipple. Breast cancer disturbs this pattern. We have developed a database of approximately 20,000 regions of interest (ROIs) of normal breast tissue and 100 regions containing a portion of a breast cancer. Each region was automatically extracted from a mammogram that was reduced in size. These ROIs were used to train an artificial neural network called a self-organizing map (SOM) to learn the mammographic pattern of normal breast tissue. SOMs are self-learning classifiers that categorize input data into a number of distinct classes. The SOM was able to train to classify normal and abnormal ROIs. However, to date our method correctly identified only 10% of the normal cases as normal, and 92% abnormal cases as not normal. Our target was to achieve between 25% and 50% recognition of normal mammograms. As a result of this study, several new ideas were generated that could improve the performance of the technique.
Document Details
- Document Type
- Technical Report
- Publication Date
- Oct 01, 2005
- Accession Number
- ADA448474
Entities
People
- Robert Nishikawa
Organizations
- University of Chicago