A machine learning approach for lung nodule classification.
Principal Investigator
Name
Alexander Kagen
Degrees
M.D.
Institution
Mount Sinai Hospital System
Position Title
Associate Professor of Radiology & Site Chair
Email
About this CDAS Project
Study
NLST
(Learn more about this study)
Project ID
NLST-239
Initial CDAS Request Approval
Sep 2, 2016
Title
A machine learning approach for lung nodule classification.
Summary
The landmark National Lung Screening Trial (NLST) is the first randomized controlled study to demonstrate a significant 15-20% decrease in disease-specific lung cancer mortality in participants who underwent screening chest CT [1]. Improving the diagnostic accuracy of lung nodule detection has become a primary focus for radiologists as early detection screening programs are implemented across the nation.
Machine learning approaches are increasingly used as assistive tools in medical image analysis. Computational models such as convolutional neural networks [2], regression trees [3], and support vector machines [4] have demonstrated promising diagnostic performance statistics in the analysis of lung nodules on CT. However, many automated detection algorithms face challenges in achieving high specificity beyond the order of 73.91% – 87.87% [4,5]. A diagnostic tool that decreases the number of false positives will add significant value to improve radiology workflow and to better establish assistive detection technologies as a standard of care.
Recently, "deep" machine learning techniques have dramatically improved the ability to understand images by allowing the algorithms to learn the features not just the decision boundary. We propose to apply these techniques to the medical images in question. The goal of this study is to develop a high-performance machine learning model for the identification of typical and atypical lung nodules on screening CT examinations. Using morphologic features such as nodule shape, size, density, location and margins alongside longitudinal data such as growth rate, we hypothesize that this machine learning model will achieve high accuracy, sensitivity, and specificity in the primary detection of lung nodules and in subsequent classification of benign versus malignant entities.
References:
[1] Kramer BS et al. Lung cancer screening with low-dose helical CT: results from the National Lung Screening Trial (NLST). J Med Screen. 2011;18(3):109-11.
[2] Ciompi F et al. Automatic classification of pulmonary peri-fissural nodules in computed tomography using an ensemble of 2D views and a convolutional neural network out-of-the-box. Med Image Anal. 2015 Dec;26(1):195-202.
[3] Lu L et al. Hybrid detection of lung nodules on CT scan images. Med Phys. 2015 Sep;42(9):5042-54.
[4] Demir Ö, Yılmaz Çamurcu A. Computer-aided detection of lung nodules using outer surface features. Biomed Mater Eng. 2015;26 Suppl 1:S1213-22.
[5] Madero Orozco H et al. Automated system for lung nodules classification based on wavelet feature descriptor and support vector machine. Biomed Eng Online. 2015 Feb 12;14:9.
Machine learning approaches are increasingly used as assistive tools in medical image analysis. Computational models such as convolutional neural networks [2], regression trees [3], and support vector machines [4] have demonstrated promising diagnostic performance statistics in the analysis of lung nodules on CT. However, many automated detection algorithms face challenges in achieving high specificity beyond the order of 73.91% – 87.87% [4,5]. A diagnostic tool that decreases the number of false positives will add significant value to improve radiology workflow and to better establish assistive detection technologies as a standard of care.
Recently, "deep" machine learning techniques have dramatically improved the ability to understand images by allowing the algorithms to learn the features not just the decision boundary. We propose to apply these techniques to the medical images in question. The goal of this study is to develop a high-performance machine learning model for the identification of typical and atypical lung nodules on screening CT examinations. Using morphologic features such as nodule shape, size, density, location and margins alongside longitudinal data such as growth rate, we hypothesize that this machine learning model will achieve high accuracy, sensitivity, and specificity in the primary detection of lung nodules and in subsequent classification of benign versus malignant entities.
References:
[1] Kramer BS et al. Lung cancer screening with low-dose helical CT: results from the National Lung Screening Trial (NLST). J Med Screen. 2011;18(3):109-11.
[2] Ciompi F et al. Automatic classification of pulmonary peri-fissural nodules in computed tomography using an ensemble of 2D views and a convolutional neural network out-of-the-box. Med Image Anal. 2015 Dec;26(1):195-202.
[3] Lu L et al. Hybrid detection of lung nodules on CT scan images. Med Phys. 2015 Sep;42(9):5042-54.
[4] Demir Ö, Yılmaz Çamurcu A. Computer-aided detection of lung nodules using outer surface features. Biomed Mater Eng. 2015;26 Suppl 1:S1213-22.
[5] Madero Orozco H et al. Automated system for lung nodules classification based on wavelet feature descriptor and support vector machine. Biomed Eng Online. 2015 Feb 12;14:9.
Aims
Aims:
(1) To develop deep learning algorithms as an adjunct diagnostic tool for identification and classification of lung nodules from labeled and unlabeled medical imaging data.
(2) To optimize classifiers for predicting the progression of lung nodules into benign or malignant neoplasms using longitudinal data and available pathology correlation.
(3) To implement this technology as a commercial software to assist radiologists in image interpretation.
Collaborators
David Stavens, Ph.D. – Mieon, Inc.
Alexander Kagen, M.D. – Mount Sinai West, Mount Sinai Hospital System