Accelerating research on 3D medical image classification and regression

L. Boulogne

• Promotor: B. van Ginneken and H. van der Heijden
• Copromotor: C. Jacobs
• Graduation year: 2025
• Radboud University

Abstract

This thesis discusses methods for automated classification and regression from 3D medical images, as well as ways to accelerate the development of such methods.

Chapter 2 describes a method for estimating PFT results from inspiration CT scans. This method was developed to also estimate the contribution of each lobe to the total patient-level lung function. It is aimed at improving the assessments of restrictive pulmonary diseases as well as risk assessments of bronchoscopic and surgical lung volume reduction.

Chapter 3 provides a systematic comparison of automatic methods for COVID19 classification from CT scans and provides insights into the added value of individual algorithm components to accelerate the creation of tools for accurate COVID-19 grading. It furthermore proposes adherence of automated systems to the CO-RADS reporting format to increase compatibility with clinical workflow.

Chapter 4 describes a challenge format for training solutions on private data that guarantees reusable training methodologies of challenge solutions. It applies this format to a medical image analysis challenge aimed at identifying severe COVID-19 infections from thoracic CT scans. Severe COVID-19 was defined as death or intubation within one month after the CT scan was made.

Chapter 5 describes a database that can be used for the development of a general-purpose automatic 3D medical image classifier to accelerate future research on 3D medical image classification and regression.

Finally, Chapter 6 summarizes the methodologies, results, and findings presented in this thesis. It furthermore indicates possible directions for future research.