Machine learning based Computer Aided Diagnosis system for Chest X-Ray images
Principal Investigator
Name
Agnese Marchesi
Degrees
Ph.D.
Institution
Istituto Italiano di Tecnologia (IIT)
Position Title
Post Doctoral Fellow
Email
About this CDAS Project
Study
PLCO
(Learn more about this study)
Project ID
PLCO-371
Initial CDAS Request Approval
Jun 19, 2018
Title
Machine learning based Computer Aided Diagnosis system for Chest X-Ray images
Summary
Chest X-rays are the most common techniques used for the detection of abnormalities and the diagnosis of several pathologies like Pneumonia and Lung cancer. Screening programs are promoted in many parts of the world in order to facilitate the early diagnosis of serious disease. The analysis of screening images is a difficult task even for expert radiologists, due to the large amount of images that have to be analysed, the presence of elements that can be very difficult to detect due to the low contrast of the images, noise and variability affecting the structures of interest like the shape or the size.
Therefore, in order to assist physicians in this difficult task, Computer Aided Detection and Diagnosis (CAD) is a technique often used as a tool to help the diagnosis. The methods developed for CAD systems can be generally grouped in four categories: basic image enhancement methods; stochastic modelling methods; multiscale decomposition methods; and machine learning methods. Among Machine learning methods, the approaches based on supervised learning techniques have received a large share of research in the literature.
Machine learning methods based on Deep learning architectures and specifically on Convolutional Neural Network have been proved to be very effective in several CAD tasks. Their performance is highly connected to the number of images used for the training phase. One of the main problem in using these techniques in medical imaging applications is the high risk of overfitting the data if there is not a proper validation phase. Hence having an adequate number of images may help in avoiding overfitting.
The aim of this project is the research of novel methods for CAD systems based on Deep Learning techniques focusing on Convolutional Neural Networks trained on PLCO data in order to support physicians in the diagnostic process. A transfer learning approach will be implemented at first, in which we will compare different well known architectures trained on computer vision dataset and fine-tuned on the PLCO data. Then, the focus of the project will be on designing a customized Convolutional Neural Network to be trained from scratch directly with the PLCO data. This will allow us to understand which method could be more suitable for the purpose of designing a CAD system in terms of accuracy and robustness.
Therefore, in order to assist physicians in this difficult task, Computer Aided Detection and Diagnosis (CAD) is a technique often used as a tool to help the diagnosis. The methods developed for CAD systems can be generally grouped in four categories: basic image enhancement methods; stochastic modelling methods; multiscale decomposition methods; and machine learning methods. Among Machine learning methods, the approaches based on supervised learning techniques have received a large share of research in the literature.
Machine learning methods based on Deep learning architectures and specifically on Convolutional Neural Network have been proved to be very effective in several CAD tasks. Their performance is highly connected to the number of images used for the training phase. One of the main problem in using these techniques in medical imaging applications is the high risk of overfitting the data if there is not a proper validation phase. Hence having an adequate number of images may help in avoiding overfitting.
The aim of this project is the research of novel methods for CAD systems based on Deep Learning techniques focusing on Convolutional Neural Networks trained on PLCO data in order to support physicians in the diagnostic process. A transfer learning approach will be implemented at first, in which we will compare different well known architectures trained on computer vision dataset and fine-tuned on the PLCO data. Then, the focus of the project will be on designing a customized Convolutional Neural Network to be trained from scratch directly with the PLCO data. This will allow us to understand which method could be more suitable for the purpose of designing a CAD system in terms of accuracy and robustness.
Aims
1) Research of novel methods for CAD systems
2) Deep Learning techniques for CAD systems
Collaborators
Alessio Del Bue, Researcher at Istituto Italiano di Tecnologia (IIT), alessio.delbue@iit.it
Diego Sona, Researcher at Istituto Italiano di Tecnologia (IIT), diego.sona@iit.it