Abstract Background: In this study we investigate the ability of computer extracted image features (nuclear morphology and texture) from digitized H&E tissue slides to stratify women with lymph node negative (LN-), estrogen receptor positive (ER+) breast cancer (BCa) as low or high risk as determined by Oncotype DX (ODX), a 21 gene-expression assay. Each year, over 120,000 women in the United States (1 million worldwide) are diagnosed with ER+ BCa. Treatment guidelines recommend hormone therapy (HT) plus chemotherapy (CT); however, up to 85% of ER+ BCa patients will not benefit from CT, yet will still suffer its side effects. ODX yields a numeric risk score (RS) ranging from 1-100; RS<18 suggests patients will respond to HT alone while RS>30 indicates need for adjuvant CT. Unfortunately, this test is expensive (>$4000), time-consuming, and involves destructive tissue testing. The goal of this study is to show that quantitative features calculated from H&E images can accurately predict risk stratification as determined by ODX in women with LN-, ER+ BCa, suggesting a histologic image based classifier could serve as a low-cost alternative. Methods: Digitized H&E-stained ER+ BCa tissue sampled from 111 patients (34 high and 77 low-risk as determined by ODX) were obtained from the University of Pennsylvania, the University of Medicine and Dentistry of NJ, and Case Western Reserve University. Regions of cancer were annotated manually by an expert pathologist, and representative fields of view (FOV) were chosen at 20x magnification (2000 by 2000 pixels) for each patient. A selection of nuclear boundaries was annotated manually in each FOV. For each nucleus, a set of 2343 features was extracted, including 21 morphological (size, shape, and boundary) and 2322 texture (Gabor, Local Binary Pattern, Greylevel, and Laws filter features). Using Minimum Redundancy Maximum Relevance (mRMR) feature selection, the 3 features best able to separate low and high ODX risk categories were identified and used to build a supervised Bayesian classifier. Classifier training employed a randomized 3-fold cross-validation scheme; in each trial, two-thirds of the dataset were randomly selected for training, and the remaining one-third employed for independent testing. Classifier performance was evaluated using area under the receiver operating characteristic curve (AUC), positive predictive value (PPV), and negative predictive value (NPV) with respect to low and high ODX risk categorization. Performance metrics were averaged over 100 trials of 3-fold cross-validation (see table). Results: The mRMR method selected one morphological feature (nuclear area) and two Laws-based texture features as being highly discriminating between risk categories. The Bayesian classifier trained with these 3 features yielded high AUC, PPV, and NPV measures with low variance in distinguishing ODX risk categories. The supervised classification results indicate that quantitative image features from H&E-stained histopathology are able to accurately discriminate between low and high risk patients as determined by ODX. Classification PerformancePerformance MetricAverage (100 Trials)Standard DeviationAUC0.870.018PPV0.810.039NPV0.880.017 Citation Information: Cancer Res 2013;73(24 Suppl): Abstract nr P4-03-04.