Artificial Intelligence in Functional Respiratory Imaging
11 december 2019
UAntwerp - Campus Drie Eiken - Building O - Auditorium O7 - Universiteitsplein 1 - 2610 WILRIJK (route: UAntwerpen, Campus Drie Eiken
Prof W. De Backer
PhD defence Maarten Lanclus - Faculty of Medicine and Health Sciences (Presentation in English)
Current medical practice within pulmonology focuses on outdated spirometry methods such as pulmonary function tests. Seeing these tests are bulk measures, they don’t show any regional information of the lungs, and they are subject to significant intrinsic variability and uncertainty. This eventually leads to false diagnoses and non-optimized treatment in patients. Furthermore, disease progression and effects of medical interventions are usually studied on a population level, which makes it hard to draw conclusions for patient-specific cases.
FLUIDDA has developed Functional Respiratory Imaging (FRI), which combines computational fluid dynamics and biomedical imaging techniques. FLUIDDA's proprietary FRI technology allows for visualization and quantification of regional lung structures and lung function, enabling characterization of lung health with more sensitive endpoints. FRI has been used extensively to both map disease progression, as well as study effects of medical interventions on a population level, but extrapolation of those results to patient-specific strategies was not established yet.
In this work, the possibilities of combining FRI with artificial intelligence, and more specifically, machine learning applications were studied, to ensure more optimized patient-specific diagnoses, treatments and follow-up. FRI endpoints were first tested for their sensitivity, by studying the effects of bronchodilators in asthma and COPD patients with both FRI and standard spirometry endpoints. From the results it showed that FRI parameters were substantially more sensitive, which makes them ideal input parameters for machine learning applications.
Two types of machine learning projects were established, i.e. prediction of patient-specific disease progression and responder phenotyping for medical treatments. In a first project, imminent exacerbations in COPD patients could be predicted with an accuracy of 80%. For a second project, responders to a new compound for IPF treatment could be identified with an accuracy of 86.5%. Finally, in the last machine learning project, lung transplant rejection could be predicted with an accuracy of 85%.
With this work it has been proven that in the respiratory space as well, artificial intelligence algorithms in combination with comprehensive imaging endpoints allow for the creation of very robust predictive models, enabling major potential for well-established, patient-specific diagnosis, follow-up and treatment.