https://biometry.nci.nih.gov/cdas/approved-projects/2062/

Principal Investigator
Name
Hadi Nia
Degrees
Ph.D.
Institution
Boston University
Position Title
Assistant Professor
Email
htnia@bu.edu
Project Information
Study
NLST
Project ID
NLST-471
Title
A novel approach to lung cancer detection and prediction using biomechanics-enhanced artificial intelligence
Summary
Deep learning, the leading machine learning tool in medical image analysis, was named as one of the 10 breakthrough technologies in 2013. The majority of conventional research on lung tumor analytics is focused on the tumor segmentation in CT images. The main goals of the current methodologies are to utilize efficient artificial intelligence and image processing algorithms to automate tumor detection and process large amounts of imaging data rapidly. This approach suffers from two limitations: the best achievable accuracy is the human accuracy since the models are calibrated – or “trained” as in machine learning terminologies – by images processed by radiologists; (ii) the existing algorithms act as a black box and lack intuition, a major obstacle for FDA approval. In the proposed project, we will utilize biophysical and computational modeling to enhance artificial intelligence algorithms to process low dose CT data to go beyond the mere “tumor segmentation” objective. Specifically, our goals are to improve early detection and (ii) predict key outcomes including incidence and overall survival, beyond the prediction ability of trained pathologists. The biophysical model provides the personalized map of the mechanical properties of the lung from the input CT images. The biophysical model is also capable of predicting the progression of chronic obstructive pulmonary disease (COPD), a major risk factor for lung cancer, based on a network model, which has been previously validated and published. By utilizing the additional biophysics-based information that the computational model provides and feeding them along with the original CT images of the patient, we hypothesize that the prediction power of the deep learning algorithm will be improved significantly.

Aims
Aim 1: Develop personalized computational model based on a network model of the parenchyma
Our group has shown through multiple publications that the physical microenvironment plays an important role in cancer development and progression as well as parenchymal destruction with advancing COPD. The information on lung mechanical microenvironment will potentially improve the prediction of lung cancer incidence and progression based on the following two evidence: mechanical abnormalities such as increased stiffness and high mechanical stresses promote the growth, invasiveness and migration of cancer cells. (ii) the progression of COPD, a major risk factor for lung cancer, can be predicted by computational models based on biophysical first principles. In this Aim, we will develop a personalized computational model that takes the CT images of the lung as input, and will provide the mechanical tensional map of the lung. For patients with COPD, the computational model will also be capable of taking the CT images of the patient in an early stage COPD, and providing a CT-like image of the same patient at a later stage of COPD. These additional personalized and high level information along with the original CT image of the patient will be fed to the deep learning framework (Aim 2) to predict lung cancer incidence and overall survival.

Aim 2: Develop biomechanics-enhanced artificial intelligence framework for spatio-temporal prediction of lung cancer
We will develop 3D multichannel deep learning algorithms to predict spatio-temporal incidence and overall survival in patients with lung cancer. We will use convolutional neural networks (CNN) combined with recurrent neural networks as supervised deep learning approaches. The proposed framework takes the CT images complemented by the biophysical maps from the computational model (Aim 1). The computational model will provide high-level information on physical properties of the lung as well as the the progression of COPD (in the subset of patients with existing COPD) as an input to the deep learning framework. The output of the deep learning framework will be overall survival, spatio-temporal incidence of lung cancer, and potentially the invasion pattern of the lung cancer based on the histological data. To test the efficacy of the proposed method, we will compare the prediction power of the proposed frame work (biophysical model in conjunction with classical deep learning) to the prediction power of the classical deep learning without the biophysical input information.

Collaborators
Hadi T. Nia, Ph.D. – Boston University
Béla Suki, Ph.D. – Boston University
Ramin Oftadeh, Ph.D. – Massachusetts Institute of Technology
George Washko, M.D. – Brigham and Women's Hospital, Harvard Medical School
Raúl San José Estépar, Ph.D. - Brigham and Women's Hospital, Harvard Medical School