Deep learning algorithm to predict pathologic complete response to neoadjuvant chemotherapy for breast cancer prior to treatment.

Abstract

600 Background: Pathologic complete response (PCR) to neoadjuvant chemotherapy (NAC) is associated with improved disease-free survival and overall survival in patients with breast cancer. Predicting PCR at the patient level prior to treatment initiation would allow physicians and patients to focus on therapies with the highest likelihood of success and minimize unnecessary toxicities from chemotherapy. We hypothesize that pre-treatment prediction of PCR is possible through a deep neural network algorithm trained on breast MRI imaging obtained prior to treatment. Methods: 126 tumors from patients treated with neoadjuvant chemotherapy for T3 stage breast cancer at a single institution from 2002 to 2006 were analyzed. In total, 3780 MRI slices were included. 3 MRI contrast phases (pre, immediate post, delayed post) from each slice were used as individual inputs to create separate model predictions of PCR. The 2-D CNN was trained over 50 epochs on the training set. The model was tested on the isolated test set (30% of samples). Results: Average model prediction accuracy over the total test set using a single phase of contrast (pre-, immediate post-, or delayed post-) was 90.4%. Concordance adjustment was conducted, with exclusion of slices that had produced discordant predictions across different contrast phase inputs. This resulted in an increase of overall model accuracy to 97.6%. Model accuracy was similar across subsets of age and tumor size despite differences in PCR rates. However, model accuracy was significantly lower in the triple-negative disease group. Conclusions: We demonstrate a deep neural network that accurately predicts PCR based on breast MRI imaging taken prior to NAC initiation. Our findings represent the promise of deep learning algorithms in providing personalized prognostic data for physicians and patients. [Table: see text]