Assessing emphysema in CT scans of the lungs: Using machine learning, crowdsourcing and visual similarity

Silas Nyboe Ørting

Publikation: Bog/antologi/afhandling/rapportPh.d.-afhandlingForskning

384 Downloads (Pure)

Abstract

Emphysema is a pathology in chronic obstructive pulmonary disease (COPD), a leading cause of death world-wide. Emphysema is characterized by destruction of lung tissue leading to reduced capacity for gas exchange in the lungs. The extent and appe arance of emphysema can be assessed in CT scans of the lungs. Current recommendations for assessing emphysema in CT scans is to use a combination of densitometry and visual assessment. Densitometry is a quantitative method that estimates the amount of lung tissue affected by emphysema, by measuring the amount of voxels in the CT scans with attenuation below a certain threshold. Visual assessment of emphysema by experts provides an assessment of emphysema extent and patterns that has been found to be useful for lung cancer risk prediction. Current densitometry methods are vulnerable to variation in scanners, scan protocols and software implementations, and cannot characterize emphysema patterns. On the other hand, visual assessment requires expert knowledge, is difficult, time-consuming and commonly only provides semi-quantitative estimates of emphysema extent. Machine learning methods that learn from visual assessment could combine the benefits of densitometry and visual assessment, to provide fully automated assessment of emphysema extent and patterns. One of the main issues when applying supervised machine learning in medical image analysis is obtaining labels. Not only can the labeling procedure require medical expertise and be time-consuming and costly, it can also be very difficult, even for experts, to provide accurate labels. This thesis investigates three approaches to reducing the need for labels when training machine learning methods to assess emphysema: weakly supervised learning, crowdsourcing and learning from visual similarity. Weakly supervised machine learning aims at learning from global labels instead of local labels, for example learning from image labels instead of pixel labels. By learning from weak labels, we can reduce the need for medical expertise, reduce the cost of labeling and improve label quality because assigning global labels is generally easier and less costly than assigning local labels. The thesis investigates emphysema detection and quantification within two weakly supervised learning settings, multiple instance learning, where labels are binary, and learning with label proportions, where labels are proportions. Crowdsourcing aims at reducing labeling costs by replacing expert annotators with non-experts. Assessing emphysema in a volumetric CT scan is a complex task requiring expert knowledge and experience. Adapting the task to the crowd setting could enable crowdsourced labels to be used when training machine learning methods. Thereby, allowing experts to focus on interpreting and validating trained models. The thesis provides a survey of how crowdsourcing is used in medical imaging, as well as an investigation into how emphysema assessment can be framed as a task that can be solved by non-expert crowd workers. Learning from visual similarity aims at learning representations from relative comparisons of images. One of the reasons that labels are costly and require expertise, is that labels are often obtained for a specific task and new labels need to be acquired for additional tasks. Visual similarity assessments could provide a more general characterization of visual content in images, than labels obtained for a specific task. By learning from visual similarity it is possible that more general representations can be learned. Additionally, focusing on similarity could allow non-experts to replace experts, since comparing visual similarity requires less expertise than categorizing pathology patterns. The thesis investigates how visual similarity assessments can be obtained and used for training convolutional neural networks to learn representations of chest CT scans.
OriginalsprogEngelsk
ForlagDepartment of Computer Science, Faculty of Science, University of Copenhagen
Antal sider89
StatusUdgivet - 2019

Note vedr. afhandling

PhD ønsker pdf-filen åben tilgængelig.

Citationsformater