Prediction of pathological complete response for breast cancer by post-treatment multi-phase MRI signatures with automatic segmentation

Motivation: Accurate prediction of pathological complete response (pCR) after neoadjuvant chemotherapy (NAC) allows replacing mastectomy with breast-conserving surgery for patients with locally advanced breast cancer. Goal(s): To develop a method of post-NAC pCR prediction with automatic segmentation Approach: A segmentation model was trained on pre-NAC MRI data and applied on post-NAC MRI data. A histogram signature was converted from post-NAC multi-phase images and classified into pCR and non-pCR by a convolutional neural network (CNN). Results: The AUC of post-NAC histogram signature for pCR prediction is 0.869 (95%CI: 0.815-0.913) in the training group (n=199) and 0.815 (95%CI: 0.738-0.877) in the testing group (n=132). Impact: Post-NAC MRI histogram signature based on pre-NAC segmentation model can be used to automatically predict pCR after NAC and assist individualized treatment for locally advanced breast cancer.