Improving the Generalizability of Deep Neural Networks by Teaching them Lung Cancer Pathophysiology

点击翻译按钮获取中文摘要

PROJECT SUMMARY Diagnostic and treatment approaches for non-small cell lung cancer (NSCLC) have evolved over the last decade from primarily empirical methodologies to objective strategies that rely on clinical characteristics of the patient and morphological features of the nodule. Following recommendations by the United States Preventive Service Task Force (USPSTF), high-risk individuals are screened yearly with low-dose computed tomography (LDCT) as this provides high sensitivity with acceptable specificity for lung cancer. However, the introduction of LDCT as the primary screening modality for lung cancer has increased detection rates of indeterminate pulmonary nodules that then require invasive investigation. This decreases the quality of life for at-risk individuals through repeated follow-ups and procedures, and greatly increases anxiety over what usually turns out to a benign nodule. In this proposal, we aim to improve upon these outcomes by determining the features that convolutional neural networks (CNNs) utilize when classifying lung nodules as either or benign. We will also determine if providing CNNs with pre-specified histologic image features known to be associated with lung cancer improves their ability to generalize to novel images outside the image set used to train them. The central hypothesis of this proposal is that increasing the attention of a CNN on LDCT image features that are accepted as being pathophysiologically relevant will improve its generalizability to novel images and thus its ability to accurately distinguish between malignant versus benign nodules. In the F99 Aim of this proposal, we will address this hypothesis by utilizing LDCT images from the National Lung Screening Trial (NLST) together with concept activation vectors to determine which parenchymal and tumor-specific features are used by CNNs to classify lung nodules. In the K00 aim, we will determine if endophenotypes extracted from the COPDgene LDCT image set can be used to improve CNN generalizability. Completion of these aims will lead to an increased understanding of the morphologic biomarkers of lung cancer inherent in LDCT images of the lung that are most important for accurate diagnosis. This will have potential application to the improvement of CNN classification performance in other medical domains. In addition, by adhering to the training program outlined in this proposal I will gain high levels of expertise in image biomarkers, early cancer pathogenesis and detection, genetic networks, and genomics. These will collectively serve as a solid foundation for my future career as an independent biomedical investigator.