Abstract #0993

Prediction of Breast Cancer Molecular Subtypes Using Conventional Feature Extraction and Two Machine Learning Architectures Based on DCE-MRI

Yang Zhang¹, Siwa Chan², Jeon-Hor Chen^1,3, Daniel Chow¹, Peter Chang⁴, Melissa Khy¹, Dah-Cherng Yeh², Xinxin Wang¹, and Min-Ying Su¹

¹Department of Radiological Sciences, University of California, Irvine, CA, United States, ²Tzu-Chi General Hospital, Taichung, Taiwan, ³E-Da Hospital and I-Shou University, Kaohsiung, Taiwan, ⁴Department of Radiology, University of California, San Francisco, CA, United States

Two different convolutional neural network architectures were applied to differentiate subtype breast cancer based on 5 DCE-MRI time frame images: (1) a conventional serial convolutional neural network; (2) a convolutional long short term memory (CLSTM) Network. In addition, a logistic classifier was trained using morphology and texture features, selected using a random forest algorithm. For CNN, a bounding box based on the automated tumor segmentation was used to create a cropped image of the tumor as network input. A total of 94 cancers were analyzed, including 14 triple negative, 29 HER2-positive, and 51 Hormonal-positive, HER2-negative. Upon 10-fold validation, the differentiation accuracy is 0.81-0.86 using serial CNN, and 0.88-0.95 using the CLSTM.

This abstract and the presentation materials are available to members only; a login is required.

Join Here