Using Deep Learning for Cancer Nodule Detection

—
Principal Investigator
Name
Ashish Gupta
Degrees
Ph.D., M.S.
Institution
Auburn University
Position Title
Associate Professor
Email
azg0074@auburn.edu
Project Information
Study
NLST
Project ID
NLST-300
Title
Using Deep Learning for Cancer Nodule Detection
—

Summary

This project is directed toward developing a clinical decision support system to assist radiologists in the detection and diagnosis of malignant pulmonary nodules in CT scans. We have previously conducted a preliminary study that explored the use of deep learning for training neural network classifiers using the LIDC-IDRI dataset. We plan to use deep learning in this research for the classification of nodule or non-nodule. Candidate nodules identified by the system as nodules will then be assigned a likelihood of malignancy using a separate deep learning model.
We hope to extend this work by using data from the NLST for boosting ensembles of neural network classifiers or via other techniques to increase the effectiveness of the system. The LIDC-IDRI dataset is better suited for training a model to generate the likelihood of malignancy for individual nodules – using the NLST dataset for further training of our models and ensembles will allow for generalizing the likelihood of malignancy scores to entire scans (and subjects). Separately we are implementing an algorithm for filtering and segmentation to generate candidate nodules for the classifiers such that the resulting system would be a proof of concept that the system could be used in practice.
—
Aims

The ultimate objective is to develop a system that delivers the likelihood of malignancy for a patient based entirely on a single CT scan. This could act as second reader for radiologists tasked with detecting and diagnosing lung cancer from CT scans. If a scan is purported to have a high likelihood of malignancy and the radiologist disagrees, then the suspect nodules can be shown to the reading radiologist for a second look. Alternately, if the system detects a low likelihood of malignancy and the radiologist disagrees, the nodules of concern to the radiologist can be compared with the results from the system. Each individual nodule will be given a specific likelihood of malignancy for radiologists to consider. Thus, the system will be designed to simultaneous reduce both type I and type II errors in the process of detecting and diagnosis lung cancer from CT scans.
Collaborators

Richard Gruetzemacher, rrg0013@tigermail.auburn.edu, Doctoral Student at Auburn University (Advisor: Dr. Ashish Gupta)