A Computer-Aided Diagnosis System for Breast Cancer Combining Mammography and Proteomics
Annual summary rept. 1 May 2006 30 Apr 2007
DUKE UNIV MEDICAL CENTER DURHAM NC
Pagination or Media Count:
This study investigated a computer-aided diagnosis system for breast cancer by combining the following three data sources mammogram films, radiologist-interpreted BI-RADS descriptors, and proteomic profiles of blood sera. We implemented under 100-fold cross-validation various classification algorithms, including Bayesian probit regression, iterated Bayesian model averaging, linear discriminant analysis, artificial neural networks, as well as a novel method of decision fusion. The top-performing classifier, decision fusion achieved AUC 0.85 0.01 on the calcification data set and 0.94 0.01 on the mass data set. Decision fusion had a slight performance gain over the ANN and LDA p 0.02, but was comparable to Bayesian probit regression. Decision fusion significantly outperformed the other classifiers p 0.001. The blood serum proteins detected lesions moderately well AUC 0.82 for normal vs. malignant and normal vs. benign but failed to distinguish benign from malignant lesions AUC 0.55, suggesting they indicate a secondary effect, such as inflammatory response, rather than a role specific for cancer.
- Medicine and Medical Research
- Computer Programming and Software