Grid-Enabled Quantitative Analysis of Breast Cancer
Abstract
The tong-term goat of this research is to improve breast cancer diagnosis, risk assessment, response assessment, ana patient care via the use of large-scale, multi-modality computerized image analysis. The central hypothesis of this research is that large-scale image analysis for breast cancer research will yield improved accuracy and reliability when optimized over multiple features and large multi-modality databases. In the first year of research, we designed a pilot study utilizing large scale parallel Grid computing harnessing nationwide infrastructure for medical image analysis. Also, using a 256-CPU high-throughput computing cluster, dimension reduction techniques were applied to ultrasound, full-field digital mammography, and DCE-MRI breast CADx feature spaces. Results indicated the ability to rival or exceed traditional breast CA Ox performance. Building on this success, during the second year, we investigated methods for using unlabeled ("truth-unknown") data. Often, there are practical difficulties in assembling large, labeled (histo-pathology) breast image data sets, while unlabeled data may be abundant This is problematic for conventional CADx schemes reliant on supervised classifiers trained using labeled data only. We proposed using unlabeled breast image data to enhance breast CADx. We hypothesize that unlabeled data information call act as a 'regularizing" factor aiding classifier robustness. After conducting experiments using previously collected data sets. encouraging na results were found indicating unlabeled data can improve CA Dx classifier performance.
Document Details
- Document Type
- Technical Report
- Publication Date
- Oct 01, 2010
- Accession Number
- ADA551856
Entities
People
- Andrew R. Jamieson
Organizations
- University of Chicago