Jae Ho Sohn

MD, MS

UCSF

Researcher & Resident Physician

NLST (Learn more about this study)

NLST-297

Apr 4, 2017

Application of Deep Learning to Combine Clinical and Imaging Data to Localize, Characterize, and Prognosticate Lung Cancer Patients

Lung cancer is the leading cause of cancer death in the United States. Despite aggressive research for diagnosis and treatment, overall survival for lung cancer has remained nearly flat in the past decade. National lung cancer screening trial followed by implementation of low dose screening CT program has dramatically improved detection of lung cancer at early stage. Unfortunately, increased sensitivity came at the cost of increased false positives, likely resulting in increased radiation exposure due to numerous CT follow ups, more biopsies of benign nodules, and increased diagnosis of indolent lung cancers that does not impact patient survival. Hence, there is important clinical need to improve diagnostic and prognostic accuracy of lung cancer.

Traditional methods of lung cancer diagnosis and prognosis research often relied on manual feature engineering to extract what the investigator thought were important imaging features to predict patient outcome. These were manually annotated and combined with categorial variables from clinical data to create a prediction model, often with a variation of generalized linear models or Cox proportional hazards models. Unfortunately, image analysis portion was inevitably limited by the subjectivity of data annotator and manual feature engineering task which always resulted in information loss.

Recent advancements in machine learning with artificial neural network, also known as deep learning, has brought about exciting improvement in computer vision. Starting in 2016, computer vision has consistently surpassing human experts in classifying objects on the ImageNet large scale visual recognition challenge (ILSVRC). Numerous cross-pollination projects between machine learning and radiology have begun, with the goal of improving radiologists' diagnostic accuracy and improve workflow efficiency. Unfortunately, a majority of these interdisciplinary projects remain in immature or experimental stage at this point, in part due to paucity of large-scale available medical data such as the NSLT data.

With this project, we plan to fully utilize the NLST data to create a convolutional neural network to automatically extract relevant features from imaging data and then combine this with a separate regression network that brings together clinical data. This will be followed by a top layer that accounts for all of these features to predict useful information for the patient, including characterization of lung nodule and prognostication for the patient. In the process, we also plan to optimize and utilize a previously developed lung nodule localizer. Ultimately, our algorithm will bring a fresh perspective and improvement to existing diagnosis and staging systems for lung cancer.