Multidimensional Shape and Color Distributions as a Computational Biomarker for Cancer Pathology
A Milutinovic, MA Reza, DE Breen, FU Garcia. College of Medicine, Drexel University, Philadelphia, PA; College of Engineering, Drexel University, Philadelphia, PA
Background: Imaging, both in 2D and 3D, is now a ubiquitous procedure in medicine, science and engineering. The resulting images contain a wealth of information that to date has only been partially utilized. We are developing a computational biomarker for the characterization of breast carcinomas to predict lymph node status.
Design: Our method is based on computational examination of a routinely applied prognostic panel that uses immunohistochemistry and H&E staining of 50 invasive breast cancer carcinomas with known lymph node status. This panel includes: estrogen and progesterone receptors, MIB-1 (proliferative activity), mutated p53 and HER2/neu. The highest staining marker from the panel was selected for further digital image analysis. Using image processing techniques and geometric analysis, the architectural histologic pattern and the expression of these prognostic markers were analyzed to create feature vectors that numerically characterize each tumor's features. Tumor characteristics including pathologic staging were added to the feature vectors. These feature vectors are fed into an algorithm called Support Vector Machines (SVM), which uses a training set of feature vectors to classify unknown samples.
Results: A total of 50 cases were processed to create the feature vectors. Of the 50 cases, 25 were lymph node positive (N1 and higher) and 25 negative (N0). In iterative testing of the algorithm, where one case is used as an unknown specimen and all others are used for training, the best results were obtained using a Radial Basis Function kernel with a sigma parameter of 35. Of the 50 cases, 19 lymph positive and 16 lymph negative cases were correctly classified (0.76 sensitivity and 0.64 specificity).
Conclusions: 1. SVM analysis can predict lymph node status in the majority of patients with invasive mammary carcinoma. 2. Refinement of the feature vector is currently in development to increase specificity and sensitivity. 3. N0 lymph nodes will be further studied using NSABP protocol B-32 to improve specificity in around 11.4% (detection of occult metastases).
Monday, March 22, 2010 1:00 PM
Poster Session II # 251, Monday Afternoon