Hugo Aerts

PhD

Harvard-DFCI

Director CIBL

NLST (Learn more about this study)

NLST-208

Apr 26, 2016

AI for Lung Nodule Characterization

Lung cancer is one of the leading causes of cancer deaths worldwide. The long-term survival rate of lung cancer remains low due to the aggressive and heterogeneous nature of lung tumors. Moreover, most of the patients diagnosed with lung cancer, fall in advance stage categories (40% Stage IV, 30% Stage III), which limits the treatment and/or therapy impact. Therefore early stage detection of lung cancer could improve the long-term survival rate. National lung cancer screening trial (NLST) has shown significant (20%) reduction in Lung cancer mortality using low dose computed tomography. However, the resulting high false positive rates and the requirement of highly skilled staff to manually interpret the screen results limit the broad clinical applications of the screening. The goal of this project is to evaluate and compare different AI techniques to quantify the nodule phenotype. We want to apply a large suite of radiomic features that are able to automatically quantify phenotypic traits in a data driven fashion and evaluate its performance. Furthermore, we want to investigate deep learning techniques (Convolutional neural networks (CNNs), to characterize the nodules and normal tissue. And the end we want to compare and combine both techniques. These investigations could enhance the efficiency and efficacy of lung cancer screening and could be a potential step forward towards the clinical implementations.

1) Evaluate radiomic metrics derived from nodules and other tissues.

2) Evaluate deep learning metrics extracted from nodules and other tissues

3) Compare radiomics and deep learning techniques.