Abstract P4-09-11: Computer extracted features of tumor grade from H&E images predict oncotype DX risk categories for early stage ER+ breast cancer

Abstract

Abstract Introduction: The ODx test is a 21 gene assay that is currently employed for separating Estrogen Receptor positive (ER+) breast cancer patients into low (L) and high (H) risk of recurrence categories, helping clinicians decide if adjuvant chemotherapy is appropriate. In this study, we sought to explore whether computer extracted features pertaining to tumor grade (nuclear pleomorphism, tubule count, mitotic index) in conjunction with a machine learning classifier were predictive of the corresponding ODx risk category for ER+ breast cancer patients. Design: First, 2000x2000 pixel sub-regions of digitized H&E slides at 40x are processed to both identify and segment epithelial and stromal nuclei using a combination of watershed and deep learning (DL). 247 nuclear features consisting of architecture, shape, and texture features were extracted from these segmentations. Subsequently, the mitotic and tubule related features were extracted at each nuclei candidate using DL detectors. The input to this process was a binary mask computed by thresholding a blue ratio transformed image using Otsu's method. The identified regions were analyzed using DL to determine if a nucleus is a part of a tubule, and/or if it is mitotic. Finally, all of these features were combined, evaluated using Ranksum feature ranking, and then used to generate predictive models using four different supervised machine learning classifiers - random forest, support vector machine, linear discriminant analysis, and a neural network – via a 3-fold cross validation scheme. The classifiers were evaluated by their ability to distinguish between the four different classification tasks presented above using the area (AUC) under the Receiver Operating Characteristic (ROC) curve: 1) L ODx and L mBR grade vs. H ODx and H mBR grade (L-L vs. H-H), 2) L ODx vs. H ODx, 3) L ODx vs. Intermediate (T) and H ODx, 4) L and T ODx vs. H ODx. Results: The highest performing features were consistently mitosis, epithelial architectural, and tubule features. Classification accuracy ranged from 0.61 (L vs. T and H) to 0.97 (L-L vs. H-H) (Table 1). These features were able to provide the highest level of classification utility for the most distinct cases (L-L vs. H-H) and had less classification accuracy with classification problems involving more difficult T cases. Number top 10 Features in each categoryNumber top 10 Features in each categoryNumber top 10 Features in each categoryExperimentMax AUCMitosisTubuleEpithelial ArchitectureL-L vs. H-H (N=36)0.97315L vs. H (N=72)0.77505L vs. T and H (N=125)0.61208L and T vs. H (N=125)0.75505Table 1: Maximum AUC, and best features used to obtain those results. Conclusion: Computer derived features pertaining to nuclear architecture and mitotic index were predictive of ODx risk categories. Additional independent validation of these findings is needed in a separate test set. Citation Format: Whitney JR, Romeo-Bucheli D, Janowczyk A, Ganesan S, Feldman M, Gilmore H, Madabhushi A. Computer extracted features of tumor grade from H&E images predict oncotype DX risk categories for early stage ER+ breast cancer [abstract]. In: Proceedings of the 2017 San Antonio Breast Cancer Symposium; 2017 Dec 5-9; San Antonio, TX. Philadelphia (PA): AACR; Cancer Res 2018;78(4 Suppl):Abstract nr P4-09-11.