CT-Based Deep Learning Enables Chemotherapy Response Prediction for Pancreatic Ductal Adenocarcinoma: A Pilot Study

Abstract

Abstract Background Predicting response to chemotherapy can lead to its optimization. Furthermore, prediction using a noninvasive approach is clearly preferable to an invasive approach. This study aimed to predict in vitro chemosensitivity assay results by combining computed tomography (CT) images and deep learning (DL) to optimize chemotherapy for pancreatic ductal adenocarcinoma (PDAC) Methods We collected the dataset of preoperative CT images and the histoculture drug response assay (HDRA) of 33 patients undergoing surgery for PDAC at our facility. We trimmed small patches from the entire tumor area, using the data augmentation technique, and obtained 10,730 patches. We established various prediction labels for 5-fluorouracil (FU), gemcitabine (GEM), and paclitaxel (PTX). We built a predictive model using a residual convolutional neural network and used 3-fold cross-validation. Results Of the 33 patients, effective response to FU, GEM, and PTX by HDRA was observed in 19 (57.6%), 11 (33.3%), and 23 (88.5%) patients, respectively. The average accuracy and the area under the receiver operating characteristic curve (AUC) of the model for predicting the effective response to FU were 93.4% and 0.979, respectively. In the prediction of GEM, the models demonstrated high accuracy (92.8%) and AUC (0.969). Likewise, the model for predicting response to PTX had a high performance (accuracy 95.9%, AUC 0.979). Conclusions Our CT-patch-based DL model exhibited high predictive performance in projecting HDRA results. Our study suggests that the DL approach could possibly provide a noninvasive means for the optimization of chemotherapy.