Predict Breast Tumor Response to Chemotherapy Using a 3D Deep Learning Architecture Applied to DCE-MRI Data

Abstract

Purpose: Many breast cancer patients receiving chemotherapy cannot achieve positive response unlimitedly. The main objective of this study is to predict the intra tumor breast cancer response to neoadjuvant chemotherapy (NAC). This aims to provide an early prediction to avoid unnecessary treatment sessions for no responders’ patients. Method and material: Three-dimensional Dynamic Contrast Enhanced of Magnetic Resonance Images (DCE-MRI) were collected for 42 patients with local breast cancer. This retrospective study is based on a data provided by our collaborating radiology institute in Brussels. According to the pathological complete response (pCR) ground truth, 14 of these patients responded positively to chemotherapy, and 28 were not responsive positively. In this work, a convolutional neural network (CNN) model were used to classify responsive and non-responsive patients. To make this classification, two CNN branches architecture was used. This architecture takes as inputs three views of two aligned DCE-MRI cropped volumes acquired before and after the first chemotherapy. The data was split into 20% for validation and 80% for training. Cross-validation was used to evaluate the proposed CNN model. To assess the model’s performance, the area under the receiver operating characteristic curve (AUC) and accuracy were used. Results: The proposed CNN architecture was able to predict the breast tumor response to chemotherapy with an accuracy of 91.03%. The Area Under the Curve (AUC) was 0.92. Discussion: Although the number of subjects remains limited, relevant results were obtained by using data augmentation and three-dimensional tumor DCE-MRI. Conclusion: Deep CNNs models can be exploited to solve breast cancer follow-up related problems. Therefore, the obtained model can be used in future clinical data other than breast images.