Machine Learning Methods for Nodule Detection and Classification

Principal Investigator

Name
Xiaohui Xie

Degrees
Ph.D.

Institution
University of California, Irvine

Position Title
Professor

Email
xhx@uci.edu

About this CDAS Project

Study
NLST (Learn more about this study)

Project ID
NLST-304

Initial CDAS Request Approval
May 11, 2017

Title
Machine Learning Methods for Nodule Detection and Classification

Summary
Low-dose lung CT imaging has been demonstrated to be an effective way for early detection and diagnosis. However, screenings through CT imaging are also associated with high false positive rates, high variability among different clinicians, and over-diagnosis of insignificant lesions. To address these issues, it is important to develop fully automated robust image analysis tools that can increase detection rate and meanwhile reduce false positives.

Inspired by the recent success of using deep convolutional features for object detection and segmentation, we propose end-to-end trained deep convolutional nets for nodule detection and classification in lung CT images. More specifically, we will first develop deep three dimensional convolutional nets for detecting locations of nodules. We will then develop a classifier based on deep nets to separate malignant nodules from benign ones.

If successful, this project can significantly improve the accuracy of automatic lung CT image analysis, improving detection rate while at the same time reducing false positives. Through accurate and efficient computer aided analysis, we hope to significantly reduce the cost of lung cancer screening.

Aims

Specific Aim 1. Develop three dimensional convolutional nets for nodule detection. The challenge is to reduce false positive rates while maintaining high sensitivity. We will try different architecture of 3D convolution and latest machine learning techniques to reduce false positive rates.

Specific Aim 2. Develop deep learning models to classify malignant vs. benign nodules. In addition to image features, we will also include clinical data into the prediction model.

Collaborators

Daniel Kim, UC Irvine