AI for hydrocephalus research

NAISS 2023/5-234


NAISS Medium Compute

Principal Investigator:

Anders Eklund


Umeå universitet

Start Date:


End Date:


Primary Classification:

30208: Radiology, Nuclear Medicine and Medical Imaging




Objective The project is aimed to investigate brain structures' relation to gait disorders, and cognitive level. In a previous project we investigated the relationship between T1 MRI brain images and higher level gait disorder (HLGD) with Convolutional Neural Networks (CNNs) and found a significant relationship. HLGD is one of the cardinal symptoms of idiopathic normal pressure hydrocephalus (INPH) and it is important to better understand the link between symptoms and brain structure to be able to improve the treatment for this patient group. We need to deepen our understanding about which features in the images that are strongly linked to the impaired gait, both through explainable AI (XAI) and by investigating a clinical cohort, who have undergone shunt surgery and gait assessment. The first objective of this study is to investigate if there are sufficient differences in brain morphology between shunt responders and non-responders to be able to discriminate these two groups by applying a CNN classifier to their brain MR images. INPH also often affect the patients' cognition, therefore the second objective is to investigate if there are sufficient differences in brain structure between study participants with different cognitive level so that they can be discriminated by a CNN classifier. Methods and materials Approx 900 brain MRI (T1, T2 flair & 4D flow) from the VESPR cohort, older people with and without gait disturbance, with different cognition levels. Approx 200 brain MRI (T1) from a retrospective cohort of patients with INPH, who have undergone shunt surgery at Umeå University hospital between 2007-2019. The pre-processing of the images includes co-registration to the T1 ICBM152 template brain and skull stripping. Several CNN:s will be designed to classify the images. At this point we do not strive to develop a new type of network, but rather to modify already existing promising architectures that could suit our problem well. The CNN:s will be used for classification of shunt responders versus non-responders and classification based on cognitive status as indicated by the MoCA test scores in the VESPR cohort. The classification accuracy will be verified by cross validation. Preliminary results Master thesis works at our department, together with a soon published study, shows significant results regarding classification of brain MR images of HLGD subjects and controls. This relationship needs to be investigated further, on a clinical cohort. Significance The results should further clarify the links between brain structure and symptoms, regarding both gait and cognitive level.