Prediction of Pathological Complete Response to Neoadjuvant Chemotherapy Using Multi-scale Patch Learning with Mammography.

Pathological complete response (pCR) indicates the absence of residual tumor in the breast and axillary nodes after neoadjuvant chemotherapy (NAC), which reduces cancerous tumor and improves the prognosis of breast-conserving surgery. To avoid eventual toxicities by NAC and improve the long-term survival outcome, the prediction of pCR using routine breast imaging is an important step to decide the patient treatment. In this paper, we propose a multi-scale patch learning method to predict pCR from pre-NAC mammography, which is widely used for early detection of breast cancer. We use two images (CC and MLO view) of mammography to integrate the texture and shape information of breast tumors in a form of image pyramid with multiple scales. We first extract fixed-sized patches from each pyramid level and concatenate them along the channel dimension to learn multi-scale features of the breast tumor and its surrounding regions. The proposed model achieved better prediction performance (0.803 AUC, 0.75 accuracy, 0.733 sensitivity, and 0.767 specificity) in pCR prediction task than other comparative methods which have been introduced for breast cancer characterization using mammography.