MSc Thesis Defense by Silas Nyboe Ørting – Københavns Universitet

MSc Thesis Defense by Silas Nyboe Ørting


Automatic estimation of emphysema extent in low-dose CT scans of the lungs.


Precise and accurate estimates of emphysema extent and sub-type is of interest to the clinical community, both to better understand the development and progression of emphysema, but also as a tool in the search for genetic associations. Emphysema is part of Chronic Obstructive Pulmonary Disease (COPD), which is a leading cause of death world wide, and the lack of effective treatments for COPD have prompted research into genetic associations of COPD.

Visual assessment of emphysema can be used to quantify the extent and sub-type of emphysema, but is time-consuming and suffers from inter-observer variability. Visual assessment typically provides information about emphysema at the scan level for sub-types, and at the level of large sub-regions for extent. Automated methods matching visual assessment by experts are in themselves interesting, but the possibility of finer-grained assessment even more so.

We consider the problem of training a classifier to predict emphysema extent in smaller patches (instances) from information on extent in larger regions (bags). This kind of learning problem occurs in multiple instance learning and learning from label proportions (LLP), where instance classifiers are trained from bag level labels. An existing LLP method, cluster model selection (CMS), is adapted to the problem of predicting emphysema extent and it is evaluated on CT scans from the Danish Lung Cancer Screening Trial. The implementation of CMS and an associated feature extraction pipeline is made publicly available .

Empirical evaluation shows that CMS can use bag label proportions to learn a weighting of the feature space that matches the problem domain. We also show that CMS can learn from both binary bag labels and proportional bag labels. The evaluation also indicates that predicting emphysema extent in a region is considerably more difficult than predicting presence of emphysema in region.

Supervisor: Jens Petersen

Co-supervisor: Marleen de Bruijne

Censor: Rasmus Reinhold Paulsen