Transfer learning approach to predict biopsy-confirmed malignancy of lung nodules from imaging data: A pilot study

Abstract

The goal of this study is to train and assess the performance of a deep 3D convolutional network (3D-CNN) in classifying indeterminate lung nodules as either benign or malignant based solely on diagnostic-grade thoracic CT imaging. While prior studies have relied upon subjective ratings of malignancy by radiologists, our study relies only on data from subjects with biopsy-proven ground truth labels. Our dataset includes 796 patients who underwent CT-guided lung biopsy at one institution between 2012 and 2017. All patients have pathology-confirmed diagnosis (from CT-guided biopsy) and high-resolution CT imaging data acquired immediately prior to biopsy. Lesion location was manually determined using the biopsy guidance CT scan as a reference for a subset of 86 patients for this proof-of-concept study. Rather than training the network without a priori knowledge, which risks over fitting on small datasets, we employed transfer learning, taking the initial layers of our network from an existing neural network trained on a distinct but similar dataset. We then evaluated our network on a held out test set, achieving an area under the receiver operating characteristic curve (AUC) of 0.70 and a classification accuracy of 71%.