Elizabeth J. Sutton MD, Presenter: Nothing to Disclose

Brittany Dashevsky MD, DPhil, Abstract Co-Author: Nothing to Disclose

Jung Hun Oh PhD, Abstract Co-Author: Nothing to Disclose

Harini Veeraraghavan, Abstract Co-Author: Nothing to Disclose

Elizabeth A. Morris MD, Abstract Co-Author: Nothing to Disclose

Joseph Owen Deasy PhD, Abstract Co-Author: Nothing to Disclose

Aditya Prakash Apte PhD, Abstract Co-Author: Nothing to Disclose

Girard Gibbons BA, Abstract Co-Author: Nothing to Disclose

Breast cancer subtypes have been classified based on tumor genotype variation and are indicators of disease free and overall survival. Using texture features extracted from magnetic resonance imaging (MRI) and a machine learning method, we investigated whether imaging characteristics could differentiate breast cancer subtypes.

This retrospective study received institutional review board approval and need for informed consent waived. 178 women with invasive ductal carcinoma (IDC) and preoperative breast MRI were identified. Immunohistochemistry surrogates defined subtypes, and the distribution was: estrogen and progesterone receptor positive (ERPR+; n=95, 53.4%), HER2 receptor positive (HER2+; n=35, 19.6%) and triple negative (TN; n=48, 27.0%). Clinical and pathologic data were collected. Tumors were contoured on the fat-suppressed T1-weight pre- and three post-contrast images. Shape-, texture- and histogram-based features were extracted using in-house software (Computational Environment for Radiological Research). Support vector machine (SVM), a frequently used machine learning technique for classification problems, was used to identify significant image features and build a robust model to predict each IDC subtype. 

SVM identified significant clinical, pathologic and imaging features. When the top 9 features were incorporated, the predictive model distinguished IDC subtypes with an overall accuracy of 83.4%. The model’s accuracy for each subtype was 89.2% (ERPR+), 63.6% (HER2+) and 82.5% (TN). The nine features were: nuclear grade, tumor volume, presence of multi-centric disease, three texture features, and three histogram-based features. For these features, statistical analysis was performed using Kruskal-Wallis test. For all the 9 features, there was a statistically significant difference between ERPR+, HER2+ and TN subtypes with p < 0.0001.

We have developed a machine learning-based predictive model using texture features extracted from MRI that can distinguish IDC subtypes with significant predictive power. 

We were able to leverage computer-derived MRI phenotypic image-based biomarkers that reflect the genetic variability of different breast cancer subtypes, which are associated with different outcomes.

Sutton, E, Dashevsky, B, Oh, J, Veeraraghavan, H, Morris, E, Deasy, J, Apte, A, Gibbons, G, Classification of Breast Cancer Subtypes Using MRI Texture Features.  Radiological Society of North America 2014 Scientific Assembly and Annual Meeting, - ,Chicago IL. http://archive.rsna.org/2014/14011966.html