Use Of Deep Learning To Predict Tumor Response To Neoadjuvant Therapy In Pancreatic Adenocarcinoma: Pure And Hybrid Modelling

Abstract

Presenter: Michael Watson MD | Carolinas HealthCare System Background: Neoadjuvant chemotherapy is being increasingly utilized for the treatment of pancreatic adenocarcinoma and is associated with improved survival. However, determining response to neoadjuvant therapy is difficult with preoperative imaging. New advances in artificial intelligence and deep learning techniques allow for novel analysis of images. Our hypothesis is that a deep learning model can be used to predict tumor response to neoadjuvant therapy. Methods: We identified patients with pancreatic adenocarcinoma undergoing neoadjuvant therapy that underwent planned Whipple’s procedure between November 2009 and January 2018. Surgical pathology report was reviewed to determine tumor treatment effect, with grade 0-2 defined as treatment response and grade 3-4 defined as no treatment response. Only patients with adequate imaging and pathologically determined treatment effect were included for analysis. Computed tomography (CT) scans after neoadjuvant therapy and prior to surgery were reviewed. All axial slices of the tumor were obtained with 5mm spacing. These slices were converted to an image file for analysis. These images were used to create a deep learning model to predict tumor response to neoadjuvant therapy. Image augmentation using Gaussian blurring was used to increase the number of images for model creation and validation. Accuracy of the model was determined by area under the curve (AOC) and Brier statistics. Results: 91 patients appropriate for image analysis (had appropriate CT imaging and tumor treatment effect documented by pathology report) were identified. Images were divided based on treatment response (333 images) or no treatment response (443 images) and image augmentation was used to artificially increase the image sample size. A “training set” of images consisting of 80% of patient images were used for creation of the deep learning model. Images from 73 patients were run through a 5-layer Convolutional Neural Network (CNN) with a flattened layer and two fully connected layers (LeNet model). Model initial training and validation accuracy were achieved to 100% with a loss function of < 0.02. The “testing” set of images consisted of images from 20% of patients. Images from the remaining 16 patients were then processed through the internally validated model. The model had an AOC of 0.7383 (p<0.001) and Brier statistic of 0.2347. Using the same procedure, a model was created with patients found to have greater than 10% reduction in CA 19-9 level after neoadjuvant chemotherapy (n=58). This improved the model with AOC improved to 0.7846 (p<0.0001) with Brier statistic of 0.1735. Conclusion: A deep learning model can be used to predict response to neoadjuvant chemotherapy for patients with pancreatic adenocarcinoma. Model accuracy is further improved when combined with another indicator of tumor response (hybrid model). With further model improvement and increased patients for analysis, an application can be created to predict tumor response to neoadjuvant chemotherapy so as to safely obviate the need for subjective clinical interpretation.