Evaluation of Deep Learning techniques for mortality prediction and abnormality detection in screening chest imaging
Principal Investigator
Name
Bartosz Borucki
Degrees
M.Sc.
Institution
Interdisciplinary Centre for Mathematical and Computational Modelling, University of Warsaw
Position Title
Head of Artificial Intelligence and Image Analysis in Medical Diagnostics Group
Email
About this CDAS Project
Study
NLST
(Learn more about this study)
Project ID
NLST-625
Initial CDAS Request Approval
Jan 3, 2020
Title
Evaluation of Deep Learning techniques for mortality prediction and abnormality detection in screening chest imaging
Summary
As lung cancer screening with low-dose CT became a state-of-the-art approach there is a growing demand for utilization of all information that can be provided by chest scans. This covers mortality prediction and mortality cause prediction, as well as generic abnormality detection - including early lung cancer detection, classification of other pulmonary diseases and the assessment of other structures covered by chest scans (e.g. heart).
At the same time Deep Learning techniques, significantly based on Convolutional Neural Networks, become successful in better image-based diagnostic processes by themselves or coupled by humans. Thus, the path towards better diagnostic results lies both in the improvement of the automated models, as well as in the quality improvement of low-dose CT images for human or machine perception.
Within this project, we plan to extend our work performed on the basis of lung cancer screening datasets implemented in Poland. We’re focusing on the evaluation of the feasibility of a variety of Deep Learning models for the tasks described above. Significantly it is important to broaden the validation of the methods and compare the results on both internal and external datasets, especially including the referential dataset of NLST.
At the same time Deep Learning techniques, significantly based on Convolutional Neural Networks, become successful in better image-based diagnostic processes by themselves or coupled by humans. Thus, the path towards better diagnostic results lies both in the improvement of the automated models, as well as in the quality improvement of low-dose CT images for human or machine perception.
Within this project, we plan to extend our work performed on the basis of lung cancer screening datasets implemented in Poland. We’re focusing on the evaluation of the feasibility of a variety of Deep Learning models for the tasks described above. Significantly it is important to broaden the validation of the methods and compare the results on both internal and external datasets, especially including the referential dataset of NLST.
Aims
- Evaluation of the feasibility of using Deep Learning techniques for mortality and cause prediction on screening imaging datasets.
- Evaluation of the feasibility of using Deep Learning techniques for abnormality detection in screening imaging datasets.
- Comparison of Deep Learning models performance on internal and external datasets.
Collaborators
Bartosz A. Borucki, ICM, University of Warsaw
Krzysztof S. Nowinski, ICM, University of Warsaw
Norbert Kapinski, ICM, University of Warsaw
Jedrzej Nowosielski, ICM, University of Warsaw
Jakub Zielinski, ICM, University of Warsaw
Wojciech Rosinski, ICM, University of Warsaw
Grzegorz Dudziuk, ICM, University of Warsaw