Development of CADx for pulmonary nodules
Principal Investigator
Name
Timor Kadir
Degrees
DPhil, Oxon
Institution
Mirada Medical Ltd.
Position Title
Chief Scientist
Email
About this CDAS Project
Study
NLST
(Learn more about this study)
Project ID
NLST-143
Initial CDAS Request Approval
Jul 15, 2015
Title
Development of CADx for pulmonary nodules
Summary
A key step in the screening workflow is the stratification of Pulmonary Nodules in Chest CT. Indeterminate pulmonary nodules, seen quite frequently in screening and as incidental findings, either require further short-term investigation, if sufficiently suspicious, or otherwise long-term follow-up imaging, e.g. further CTs at 6, 12 and 18 months to assess for growth/stability.
Mirada Medical has a research collaboration with the University of Oxford to investigate a prototype Computer Aided Diagnosis (CADx) application based on Quantitative Texture Analysis to assist the reading radiologist in correctly stratifying nodules within screening and non-screening workflows.
We have so far developed a system capable of distinguishing malignant nodules from benign nodules with an Area Under the Curve (AUC) of up to 0.93 from a single contrast CT. This is achieved using a suite of lung texture features fed into a machine learning system, trained on a mixture of publicly-available data (e.g. the LIDC-IDRI database) and local patient data from the Oxford University Trust Hospitals. The Mirada-Oxford team won the SPIE LUNGx lung nodule identification challenge 2015, having used a model trained on LIDC-IDRI data to calculate likelihood-of-malignancy scores for the test nodules released in the completion.
For the next phase in this project, more screening data is sought, to allow the team to assess the applicability of these texture measures and trained models to large lung cancer screening populations. The incidence rate of primary lung cancer is low for any given study, and consequently access to the complete set of primary cancers released with the NLST data would be valuable for training the discriminative algorithms, along with a representative sample of benign lung nodules.
Mirada Medical has a research collaboration with the University of Oxford to investigate a prototype Computer Aided Diagnosis (CADx) application based on Quantitative Texture Analysis to assist the reading radiologist in correctly stratifying nodules within screening and non-screening workflows.
We have so far developed a system capable of distinguishing malignant nodules from benign nodules with an Area Under the Curve (AUC) of up to 0.93 from a single contrast CT. This is achieved using a suite of lung texture features fed into a machine learning system, trained on a mixture of publicly-available data (e.g. the LIDC-IDRI database) and local patient data from the Oxford University Trust Hospitals. The Mirada-Oxford team won the SPIE LUNGx lung nodule identification challenge 2015, having used a model trained on LIDC-IDRI data to calculate likelihood-of-malignancy scores for the test nodules released in the completion.
For the next phase in this project, more screening data is sought, to allow the team to assess the applicability of these texture measures and trained models to large lung cancer screening populations. The incidence rate of primary lung cancer is low for any given study, and consequently access to the complete set of primary cancers released with the NLST data would be valuable for training the discriminative algorithms, along with a representative sample of benign lung nodules.
Aims
1) Validate existing proprietary lung nodule characterization techniques on screening population data.
2) Develop and test new characterization methods particularly tuned to lung cancer screening.
3) Optimise lung nodule segmentation and tracking across multiple time points.
Collaborators
Dr. Lyndsey Pickup, Mirada Medical Ltd.
Dr. Julien Willaime, Mirada Medical Ltd.
Dr. Mark Gooding, Mirada Medical Ltd.
Dr. Amalia Cifor, Mirada Medical Ltd.
Dr. Djamal Boukerroui, Mirada Medical Ltd.