Examenspresentation av Malin Jönsson

Examiner: Ida Häggström

Opponent: Viktoria de Carvalho

Abstract:
Idiopathic Normal pressure hydrocephalus (iNPH) is a progressive neurological condition, characterized by abnormal accumulation of cerebrospinal fluid in the ventricles of the brain. Disease symptoms can be significantly mitigated by shunt surgery but are, however, shared with other neurodegenerative diseases with higher prevalence, causing iNPH to be commonly misdiagnosed and rarely suspected. Neuroimaging plays a key role in diagnostic screening and the callosal angle (CA), referring to the angle between the medial walls of the lateral ventricles, is one of the most important markers. Although, it is currently measured manually in clinics which is time consuming, requires an experienced radiologist and involves significant interrater differences. This thesis aimed to adress this by developing an automated solution to compute CA with an image analysis approach. The objective was further to evaluate the model performance in comparison to manual CA measurements and to translate this performance into a clinical context in terms of prognostic value.

For this purpose, ventricular segmentations of MRI & CT scans from healthy and iNPH diagnosed cohorts were included. The model was built using Python ver 3.10.9 and involved: alignment of the input segmentations to standard orientation, key point detection, computation of the optimal coronal plane pitch and position, edge- and key point detection in the coronal plane, linear regression and angle computation. Based on this approach, a full pipeline was successfully developed that computes CA with an intraclass correlation score of 0.96, 0.97 and 0.69 (compared to manual measurements) for segmentations from MRI, CT 3.5 mm and CT 5mm respectively. Furthermore, the model successfully classified iNPH with a sensitivity of 0.947 based on the CA cutoff of below 90 degrees for possible iNPH, used as the current ground truth in clinics today. The results are promising and could serve as a basis for continued model development and motivate further validation on larger datasets.

Welcome!
Malin and Ida