This proposal will develop and evaluate an effective model to detect cancerous lung nodules to improve low-dose CT lung cancer screening efficiency and reduce false positives through rigorous analysis of chest CT images. We propose (a) unsupervised novel deep learning approach to identify cancerous lung nodules and differentiate them from benign nodules on chest CTs, (b) supervised statistical learning image analysis to understand how patient-specific information, lung function and severity of symptoms from clinical reports affect detection of lung cancer by means of CT imaging; and (c) integrate individual patient profiles in the nodule models with patient characteristics, medical history and CT images and assess the added values of epidemiological and personal history information for profiling.

Previous image analysis of lung imaging data have not fully utilized 3D quantitative features. Deep learning computational models, composed of multiple processing layers, fit nonlinear models for data representation. These data abstractions have dramatically improved isual object recognition applications. We propose to analyze CT image data using radiomics approach by extracting quantitative variables from the radiological features such as intensity, shape, texture, and wavelets. We leverage a 3D convolution operation for full contextual analysis of CT scans. 3D convolution operation will be used as part of a unified deep neural network detection mechanism, designed to improve sensitivity and specificity. Our central hypothesis is that a deep learning technology for object recognition that utilizes a full stack of CT scans and their 3D context is sufficiently accurate for detecting cancerous lesions on chest CTs. We also hypothesize that patient-specific information, the severity and combination of lung cancer symptoms can considerably improve statistical power of findings from CT images. We will test this central hypothesis to attain our overall objective by completing the following two specific aims:

Aim 1: (a) Develop and assess a novel deep learning method to detect cancerous nodules on chest CT scans in unsupervised learning fashion and differentiate them from benign nodules. (b) Correlate NLST chest CT scans with annotated pulmonary non-calcified nodules from the trial with patient-specific information and findings from clinical reports such pulmonary function test in three cohort of patients with squamous cell, adenocarcinoma, small cell, and metastatic lung cancer in the supervised statistical learning fashion.

Aim 2: Evaluate our developed nodule assessment models on an external validation dataset. For, this aim, we will use images that are being compiled into a imaging atlas for lung cancer screening and detection at Vanderbilt University. This data repository has accumulated images from CT screening and can be used to evaluate the test characteristics of deep learning algorithms for identifying images that associate with lung cancer risk.

The anticipated outcome for the proposed project is an automatic method that can detect cancerous lesions on CT scans with high sensitivity and specificity. We expect the successful completion of this project will have a significant positive impact on the effectiveness of lung cancer screening programs and early detection of lung cancer through comprehensive rigorous analysis of NLST chest scans in unsupervised and supervised fashion.

Saeed Hassanpur, PhD Dartmouth College
Fenghai Duan, PhD Brown University
Eugene Demidenko, PhD Dartmouth College
Matthew Schabbath, PhD Moffitt Cancer Center
Denise Aberle, MD University of California - Los Angeles
Kristen Anton, MS Dartmouth College
Robert Gillies, MD Moffitt Cancer Center
Luigi Marchionni, PhD Johns Hopkins University
Luca Cinquini, PhD Jet Propulsion Laboratory
Christopher Amos, PhD Dartmouth College