Zijian Zhou1, David E. Rauch1, Jong Bum Son1, Benjamin C. Musall1, Nabil A. Elshafeey2, Jason B. White3, Mark D. Pagel4, Stacy Moulder3, and Jingfei Ma1
1Imaging Physics, The University of Texas MD Anderson Cancer Center, Houston, TX, United States, 2Breast Imaging, The University of Texas MD Anderson Cancer Center, Houston, TX, United States, 3Breast Medical Oncology, The University of Texas MD Anderson Cancer Center, Houston, TX, United States, 4Cancer Systems Imaging, The University of Texas MD Anderson Cancer Center, Houston, TX, United States
1Imaging Physics, The University of Texas MD Anderson Cancer Center, Houston, TX, United States, 2Breast Imaging, The University of Texas MD Anderson Cancer Center, Houston, TX, United States, 3Breast Medical Oncology, The University of Texas MD Anderson Cancer Center, Houston, TX, United States, 4Cancer Systems Imaging, The University of Texas MD Anderson Cancer Center, Houston, TX, United States
Using convolutional and recursive neural networks on pre-treatment DCE
and DWI, the deep learning ensemble can predict pathologic complete response to
neoadjuvant systemic therapy for triple-negative breast cancer patients.
Figure 1. Illustration of the deep learning ensemble developed for pathologic
complete response (pCR) prediction for the triple-negative breast cancer
cohort. The ensemble took the pre-treatment DCE and DWI as input. Two
convolutional neural networks extracted features from the DCE and DWI,
respectively. The sequences of features were then input to the two recursive
neural networks, respectively. Outputs of the recursive neural networks were
concatenated and used for pCR or non-pCR prediction.
Figure 3. Receiver operating characteristic curve (blue) of the prediction using the deep
learning ensemble. Using the pre-treatment DCE and DWI, the ensemble achieved
the best accuracy of 69%, with the sensitivity of 75% for pCR patients and
specificity of 63% for non-pCR patients. The area under the curve (AUC) was
0.68.