The primary objective of this project is to improve the diagnosis of breast cancer and reduce unnecessary benign breast biopsies. An artificial neural network computer-aided 'advisor' will be constructed to assist radiologists in deciding which solid breast masses warrant immediate biopsy and which can be safely observed. Up to 80% of breast biopsies result in a benign diagnosis because benign and malignant masses often appear similar at mammography and sonography. We have previously developed computer models - artificial neural networks - for predicting breast cancer based on mammogram features and the patient's medical history. Models built with only mammography findings and the patient's age demonstrate good preliminary results, potentially eliminating 22% of benign biopsies while identifying all cancers. The principal innovative features of this project are: (1) the focus on solid breast masses by incorporating high-resolution ultrasound features, and (2) the use of a novel, standardized lexicon devised by the American College of Radiology for describing breast ultrasound findings. The goal of the proposed project is to maintain near perfect sensitivity (>98%) - similar to the accuracy of the present clinical practice of short-interval follow-up for 'probably benign' lesions - while theoretically reducing the number of biopsies of benign breast masses by half. Mammogram and ultrasound features will be established for a large retrospective database of approximately 1000 biopsy-proven cases. Artificial neural networks will be constructed with 'supervised' training to predict which masses are very likely benign and which are suspicious for malignancy. The computer predictive models will be tested on a prospective validation database of approximately 1350 cases with known biopsy results. Performance will be evaluated in terms of fraction of benign biopsies avoided while maintaining greater than 98% sensitivity. An interobserver variability study will determine whether the computer model reduces interpretation error and inconsistency between observers. Assuming performance approaching that of the preliminary model, a decision aid based on this model will be ready for prospective clinical trials at the conclusion of this project.
| Singh, Swatee; Maxwell, Jeff; Baker, Jay A et al. (2011) Computer-aided classification of breast masses: performance and interobserver variability of expert radiologists versus residents. Radiology 258:73-80 |
| Mazurowski, Maciej A; Habas, Piotr A; Zurada, Jacek M et al. (2008) Training neural network classifiers for medical decision making: the effects of imbalanced datasets on classification performance. Neural Netw 21:427-36 |
| Jesneck, Jonathan L; Lo, Joseph Y; Baker, Jay A (2007) Breast mass lesions: computer-aided diagnosis models with mammographic and sonographic descriptors. Radiology 244:390-8 |
| Jesneck, Jonathan L; Nolte, Loren W; Baker, Jay A et al. (2006) Optimized approach to decision fusion of heterogeneous data for breast cancer diagnosis. Med Phys 33:2945-54 |
