Improved Breast Cancer Diagnosis and Prognosis by Computational Modeling and Image Analysis

David E. Axelrod

Dept. Genetics and Cancer Institute of New Jersey, Rutgers University

In order to better understand the progression of breast cancer, four different tumor progression pathways (linear, non-linear, branched, and parallel) were evaluated by comparing the results of computer simulation and clinical observations. Each pathway was described as a compartment model with transitions between compartments, and then as a series ordinary of differential equations or as heuristic graphs. Optimal transition rate values were sought for each model using a random search in combination with a directed search based on the Nelder-Mead simplex method. The model that produced the least root mean squared deviation between simulations and clinically observed co-occurrence frequencies of tumor grades was the parallel model. In this model grades of carcinoma in situ and grades of invasive carcinoma diverge in parallel from a common progenitor. This parallel progression model differs from the traditional pathologist’s view of a linear progression from in situ cancers to invasive cancers.  A prediction of the parallel progression model was tested by quantitative image analysis of human biopsy specimens. Thirty-nine nuclear image features were extracted from 200 nuclei of each of 80 patients with ductal carcinoma in situ whose clinical outcome was known. Linear discriminant analysis was used for dimension reduction. A discriminant function of two nuclear image features was derived that was prognostic for patient survival. Therefore, computer modeling suggested a new parallel pathway of breast tumor progression that is consistent with clinical observations of biopsy specimens and of patient outcome. (Joint work with J.-A. Chapman, W.A. Christens-Barry, H.L. Lickley, N.A. Miller, J. Qian, L. Sontag, Y.Yuan. Supported by NJCCR 1076-CCR-S0 and NIH U56 CA 113004).