Lung pattern analysis using artificial intelligence for the diagnosis support of interstitial lung diseases
Options
BORIS DOI
Abstract
Interstitial lung diseases (ILDs) is a group of more than 200 chronic lung disorders characterized by inflammation and scarring of the lung tissue that leads to respiratory failure. Although ILD is a heterogeneous group of histologically distinct diseases, most of them exhibit similar clinical presentations and their diagnosis often presents a diagnostic dilemma. Early diagnosis is crucial for making treatment decisions, while misdiagnosis may lead to life-threatening complications. If a final diagnosis cannot be reached with the high resolution computed tomography scan, additional invasive procedures are required (e.g. bronchoalveolar lavage, surgical biopsy). The aim of this PhD thesis was to investigate the components of a computational system that will assist radiologists with the diagnosis of ILDs, while avoiding the dangerous, expensive and time-consuming invasive biopsies. The appropriate interpretation of the available radiological data combined with clinical/biochemical information can provide a reliable diagnosis, able to improve the diagnostic accuracy of the radiologists.
In this thesis, we introduce two convolutional neural networks particularly designed for ILDs and a training scheme that employs knowledge transfer from the similar domain of general texture classification for performance enhancement. Moreover, we investigate the clinical relevance of breathing information for disease classification. The breathing information is quantified as a deformation field between inhale-exhale lung images using a novel 3D convolutional neural network for medical image registration. Finally, we design and evaluate the final end-to-end computational system for ILD classification using lung anatomy segmentation algorithms from the literature and the proposed ILD quantification neural networks. Deep learning approaches have been mostly investigated for all the aforementioned steps, while the results demonstrated their potential in analyzing lung images.
In this thesis, we introduce two convolutional neural networks particularly designed for ILDs and a training scheme that employs knowledge transfer from the similar domain of general texture classification for performance enhancement. Moreover, we investigate the clinical relevance of breathing information for disease classification. The breathing information is quantified as a deformation field between inhale-exhale lung images using a novel 3D convolutional neural network for medical image registration. Finally, we design and evaluate the final end-to-end computational system for ILD classification using lung anatomy segmentation algorithms from the literature and the proposed ILD quantification neural networks. Deep learning approaches have been mostly investigated for all the aforementioned steps, while the results demonstrated their potential in analyzing lung images.
Date of Publication
2019
Year of graduation
2019
Theses Type
dissertation
Subject(s)
Language(s)
en
Author(s)
Christodoulidis, Stergios |
Faculty/Graduate School
Access(Rights)
open.access
Primary OA Publication
true