Modular Machine Learning Methods for Computer-Aided Diagnosis of Breast Cancer
Abstract
The purpose of this study was four fold: 1) Identify subsets of the training data of breast cancer features using both a priori information and unsupervised learning methods. 2) Build local models for breast cancer prediction for each subset of the training data using supervised learning methods. Evaluate the performance of the local models on the training data relative to a single, global, supervised learning model and to current clinical practice. 3) Combine the local models to form a global, modular model and evaluate the performance of the modular model on the evaluation data set relative to a single, global, supervised learning model and to current clinical practice. 4) Develop an ensemble classifier combining three sources of data (image processing, radiologist-extracted mammographic findings, and patient history) for the task of computer-aided diagnosis of breast microcalcification clusters. We developed the world's largest database of over 4400 cases containing radiologist-extracted mammographic findings, patient age, and biopsy outcome, and we used this data to develop modular, global CAD models using different machine learning algorithms applied to the entire database.
Document Details
- Document Type
- Technical Report
- Publication Date
- May 01, 2005
- Accession Number
- ADA438319
Entities
People
- Jonathan Jesneck
- Joseph Y. Lo
Organizations
- Duke University