Abstract
Perceptrons are typically trained to minimize mean square error (MSE). In computer-aided diagnosis (CAD), model performance is usually evaluated according to other more clinically relevant measures. The purpose of this study was to investigate the relationship between MSE and the area (Az) under the receiver operating characteristic (ROC) curve and the high-sensitivity partial ROC area (0.90A′z). A perceptron was used to predict lesion malignancy based on two mammographic findings and patient age. For each performance measure, the error surface in weight space was visualized. Comparison of the surfaces indicated that minimizing MSE tended to maximize Az, but not 0.90A′z.
| Original language | English |
|---|---|
| Pages (from-to) | 99-109 |
| Number of pages | 11 |
| Journal | Computers in Biology and Medicine |
| Volume | 32 |
| Issue number | 2 |
| DOIs | |
| State | Published - 2002 |
| Externally published | Yes |
Funding
Perceptrons, like more complicated backpropagation artificial neural networks, are typically trained to minimize mean square error (MSE). In computer-aided diagnosis (CAD) applications, model performance is usually evaluated according to other more clinically relevant measures from receiver operating characteristic (ROC) analysis. The purpose of this study was to investigate the relationship between MSE and the area ( A z ) under the ROC curve and the partial ROC area ( 0.90 A z ′ ) under the high sensitivity portion of the ROC curve. A perceptron was used to predict whether or not breast lesions were malignant based on two mammographic findings and patient age. For each performance measure, the error surface in weight space was visualized. Comparison of the surfaces indicated that minimizing MSE tended to maximize A z , but not 0.90 A z ′ . If it is important to maximize 0.90 A z ′ , then predictive models trained to minimize MSE may provide inferior solutions. Mia K. Markey is a doctoral candidate in Biomedical Engineering at Duke University. She received her B.S. from Carnegie Mellon University. She was awarded a Dissertation Research Award from the Susan G. Komen Breast Cancer Foundation. Her research interest is the application of machine learning techniques to problems in biology and medicine. Her current work is in computer-aided diagnosis of breast cancer, with a focus on the development of decision aids to reduce the number of benign breast biopsies. Joseph Y. Lo received his B.S.E. in 1988 and Ph.D. in 1993, both in Biomedical Engineering from the Duke University School of Engineering. He is now an Assistant Research Professor in the Department of Radiology at the Duke University Medical Center and the Department of Biomedical Engineering at Duke University School of Engineering. His research interests involve image processing and artificial intelligence techniques applied to medical imaging, in particular the computer-aided diagnosis of breast cancer. Rene Vargas-Voracek received his B.S. degree in Electrical Engineering from the National University of Mexico, Mexico City, in 1989 and the M.S. and Ph.D. degrees in Electrical Engineering from Duke University, Durham North Carolina in 1991 and 1996, respectively. He is currently a Research Associate at the Division of Imaging, Department of Radiology at Duke University Medical Center. His research interests include medical image and signal processing, information theory and statistical pattern classification and estimation. Georgia D. Tourassi , Ph.D. is an Assistant Research Professor in the Department of Radiology at Duke University Medical Center. She earned a B.S. in Physics from the University of Thessaloniki in Greece and the Ph.D. in Biomedical Engineering from Duke University. Dr. Tourassi is a member of the Institute of Electrical and Electronics Engineers (IEEE), the International Society for Optical Engineering (SPIE), the American Association of Physicists in Medicine (AAPM), and the Radiological Society of North America (RSNA). She is also associate editor in Radiology . Her research interests include applications of artificial intelligence in computer-aided medical diagnosis and medical image processing. Carey E. Floyd Jr. graduated from Eckerd College in 1976, earned the Ph.D. in Physics from Duke University in 1981, and is currently Professor of Radiology and Biomedical Engineering at Duke where he directs the Digital Imaging Research Division of Radiology. His research interests include computer assisted medical diagnosis and digital radiographic imaging.
| Funders | Funder number |
|---|---|
| National Cancer Institute | R29CA075547 |
| Susan G. Komen |
Keywords
- Breast cancer
- Computer-aided diagnosis
- Error surface
- Neural network
- Perceptron
