New image features for automatically discrimination of patients with/without lung cancer
Principal Investigator
Name
Consuelo Gonzalo Martin
Degrees
Ph..
Institution
UNIVERSIDAD POLITECNICA DE MADRID
Position Title
Professor
Email
About this CDAS Project
Study
NLST
(Learn more about this study)
Project ID
NLST-373
Initial CDAS Request Approval
Nov 14, 2017
Title
New image features for automatically discrimination of patients with/without lung cancer
Summary
Every day, thousands of medical images are recorded, and the time needed to analyze and interpret these images requires a great human effort and cost, delaying the diagnosis and treatment of different diseases. In this regard, the availability of tools and models that automate, or at least speed up, the analysis of these images would generate significant savings and improve the efficiency of health systems.
Lung cancer is responsible for one in five cancer deaths worldwide (1.6 million deaths, 19.4% of all cancer deaths) and is the leading cause of cancer deaths in men in 87 countries and women in 26 countries. Because of its high mortality (the overall ratio of mortality to incidence is 0.87), lung cancer is considered as one of the more critical diseases, therefore is necessary to research and develop new methodologies that allow to facilitate and accelerate its diagnoses and monitoring.
Given the variability in the images used to diagnose the different kind of lung cancer, their automatic discrimination is an open research challenge.
In order to prioritize the images that must be analyzed in depth to diagnose which type of cancer these patients suffer from, an important task is, in a first step, to determine with very high precision whether patients suffer from cancer or not. And then, to give some indicator of malignity that allows establishing a reliable ranking of the images to determine the order they should be analyzed by physicians.
In this regard, an exhaustive study of new textural and contextual image features is proposed, that allows grouping the patients with lung cancer and the ones without, with an accuracy higher than 98%. Moreover, a tool to help physicians to delineate cancer nodules will be developed in the framework of this project. To this end, the new NLST data: non-lung cancer and AJJCC lung cancer stage, will be used. For the generation of these models a large number of annotated images are need. The exact amount of images needed for this study cannot be estimated a priori.
The methodology proposed to address this challenge has been divided in two stages:
1. A global analysis of the lung isolated images. In this stage, the use of atlas is proposed to identify and separate left and right lungs. Different global spectral, textural and shape features will be used to characterize the whole lung. Supervised and unsupervised algorithms will be investigated to generate robust models that allow a first binary clustering of lungs with and without nodules.
2. In the second stage, only the identified lungs with nodules will be analyzed locally. For that, an over-segmentation of the lungs will be carried out using supervoxels algorithms. Another set of features, including new textural and contextual features, will be investigated in order to select the optimal subset that allow an accurate discrimination of nodules with a different level of malignity.
This study will be contacted as part of the iASiS project (H2020, Grant No. 727658) - http://project-iasis.eu/.
Lung cancer is responsible for one in five cancer deaths worldwide (1.6 million deaths, 19.4% of all cancer deaths) and is the leading cause of cancer deaths in men in 87 countries and women in 26 countries. Because of its high mortality (the overall ratio of mortality to incidence is 0.87), lung cancer is considered as one of the more critical diseases, therefore is necessary to research and develop new methodologies that allow to facilitate and accelerate its diagnoses and monitoring.
Given the variability in the images used to diagnose the different kind of lung cancer, their automatic discrimination is an open research challenge.
In order to prioritize the images that must be analyzed in depth to diagnose which type of cancer these patients suffer from, an important task is, in a first step, to determine with very high precision whether patients suffer from cancer or not. And then, to give some indicator of malignity that allows establishing a reliable ranking of the images to determine the order they should be analyzed by physicians.
In this regard, an exhaustive study of new textural and contextual image features is proposed, that allows grouping the patients with lung cancer and the ones without, with an accuracy higher than 98%. Moreover, a tool to help physicians to delineate cancer nodules will be developed in the framework of this project. To this end, the new NLST data: non-lung cancer and AJJCC lung cancer stage, will be used. For the generation of these models a large number of annotated images are need. The exact amount of images needed for this study cannot be estimated a priori.
The methodology proposed to address this challenge has been divided in two stages:
1. A global analysis of the lung isolated images. In this stage, the use of atlas is proposed to identify and separate left and right lungs. Different global spectral, textural and shape features will be used to characterize the whole lung. Supervised and unsupervised algorithms will be investigated to generate robust models that allow a first binary clustering of lungs with and without nodules.
2. In the second stage, only the identified lungs with nodules will be analyzed locally. For that, an over-segmentation of the lungs will be carried out using supervoxels algorithms. Another set of features, including new textural and contextual features, will be investigated in order to select the optimal subset that allow an accurate discrimination of nodules with a different level of malignity.
This study will be contacted as part of the iASiS project (H2020, Grant No. 727658) - http://project-iasis.eu/.
Aims
Ob1.- Develop an automatically method to determine with very high precision whether patients suffer from cancer or not.
Ob2.- Provide some new indicator(s) of malignity that allows establishing a reliable ranking of the images
Ob3.- Implement a tool for facilitating physicians information for diagnosis
Collaborators
From Hospital Universitario Puerta de Hierro Majadahonda, Madrid, Spain
María Torrente Regidor. M.P.H.
From Center of Biomedical Technology, Universidad Politécnica de Madrid
Angel García Pedrero. Post-doc researcher
César Ortiz Toro: Pre-doc student
Miriam Jiménez: Undergraduated student