Early Prediction of Pathologic Response to Neoadjuvant Chemotherapy in Advanced Breast Cancer using Convolutional Neural Network Algorithm

Objectives: The aim of this study was to predict the efficacy for predicting pathologic response to neoadjuvant chemotherapy (NAC) in advanced breast cancer with PET/CT and MRI. Convolutional Neural Network (CNN) was introduced and compared with conventional methods.

Methods: The study consisted of 56 patients with advanced breast cancer. All patients received three cycles of NAC before curative surgery. Baseline and interim images of PET/CT and MRI were conducted before and after first cycle of NAC. The SUV0, MTV0, TLG0 was determined as SUVmax, MTV, and TLG of PET data taken at the baseline, and the SUV1, MTV1, TLG1 was also obtained from interim images in the same way. The ADCmean taken at diffusion image of baseline MRI is defined as ADC0, and the ADCmean of interim is defined as ADC1. The percent change of these parameters between baseline and interim images were also measured (ΔSUV, ΔMTV, ΔTLG, ΔADC). Subgroup analysis was performed for HER2 negative and triple negative group according to molecular subtype. Data set for deep learning were cropped from baseline (PET0, MRI0) and interim (PET1, MRI1) images. The CNN learned using training images and labels. The data set was assigned as dataset (80%) and a test dataset (20%). Regarding pathologic response, Miller and Payne system was used based on tumor necrosis. ROC curve analysis was used for assessing the performance of distinguishing responders and non-responders.

Results: The median value of age was 49 years (range, 26-66). The pathologic responses of the patients were 6 responders (11%) and 50 non-responders (89%). The ΔSUV showed the highest area under the curve 0.805 (95% CI, 0.677 to 0.899). In HER2 negative subtype, ΔSUV showed highest area under curve of 0.879 (95% CI, 0.722 to 0.965). When CNN was applied, area under curve for ROC curve analysis was also improved except ADC0 (SUV0 : PET0 = 0.652 : 0.890, SUV1 : PET1 : 0.687 : 0.976, ADC0 : MRI0 = 0.703 : 0.593, ADC1 : MRI1 = 0.537 : 0.677).

Conclusions: CNN has been shown the possibility of predicting pathologic response to NAC in patients with advanced breast cancer.