A Radiomics approach to distinguish benign from malignant nodules on Low Dose CT
Principal Investigator
Name
Anant Madabhushi
Degrees
Ph.D.
Institution
Emory University
Position Title
Professor
Email
About this CDAS Project
Study
NLST
(Learn more about this study)
Project ID
NLST-218
Initial CDAS Request Approval
May 20, 2016
Title
A Radiomics approach to distinguish benign from malignant nodules on Low Dose CT
Summary
There has been recent substantial interest in the use of “radiomics” for computer assisted feature analysis and characterization of nodules on lung CT. The main hypothesis behind these approaches is that radiomic analysis allows for the extraction and identification of subtle image based attributes of the nodule that may not be visually discernible by a human expert but may allow for machine based discrimination of malignant and benign nodules. Our group has been developing new classes of radiomic features for improved discrimination of malignant from benign nodules on CT. For instance, our group has shown that granulomas tend to have lower “energy” values compared to malignant nodules where energy is a reflection of the disruption in local pixel-level gradient architecture on the CT scan. We have also found that cancers tend to have high energy while granulomas comparatively have lower energy values.
We hypothesize that our proposed technology can enable improved discrimination between benign versus malignant lung nodules than has been achievable by clinical and radiomic analysis previously. Additionally, we believe that the machine learning classifier developed as part of this project will have direct clinical utility for lung cancer screening and could serve as a decision support tool in conjunction with screening chest CT scans.
Training the machine learning classifier: The most discriminative radiomic features will be determined via cross-validation and Linear/Quadratic Discriminant Analysis (L/QDA) on the learning set. To mitigate selection and classifier training bias, a three-fold (one fold held-out for testing), patient-stratified, cross-validation scheme will be used and both classification selection strategies will be repeated 1000 times. The sensitivity and the specificity of each feature will be evaluated by Receiver operating characteristic (ROC) curve and AUC analysis.
We hypothesize that our proposed technology can enable improved discrimination between benign versus malignant lung nodules than has been achievable by clinical and radiomic analysis previously. Additionally, we believe that the machine learning classifier developed as part of this project will have direct clinical utility for lung cancer screening and could serve as a decision support tool in conjunction with screening chest CT scans.
Training the machine learning classifier: The most discriminative radiomic features will be determined via cross-validation and Linear/Quadratic Discriminant Analysis (L/QDA) on the learning set. To mitigate selection and classifier training bias, a three-fold (one fold held-out for testing), patient-stratified, cross-validation scheme will be used and both classification selection strategies will be repeated 1000 times. The sensitivity and the specificity of each feature will be evaluated by Receiver operating characteristic (ROC) curve and AUC analysis.
Aims
1) Optimize a machine learning classifiers with radiomic features of nodular morphology on non-contrast lung CT to distinguish benign from malignant nodules.
2) Validate the machine learning classifier developed in Aim 1 to distinguish benign from malignant nodules on an independent test set
Collaborators
Dr. Vamsidhar Velcheti, MD - Taussig Cancer institute, Cleveland Clinic.
Dr. Robert Gilkeson, MD - University Hospitals, Cleveland
Dr. Philip Linden, MD - University Hospitals, Cleveland
Dr. Frank Jacono, MD - Louis Stokes Cleveland VA Medical Center, Cleveland
Dr. Michael Yang, MD - University Hospitals, Cleveland