Impact of Variability in Brain Volume Measurements from MRI scans in Multi-Centre Neurodegeneration Assessments
TL;DR
This work illustrates that when neurodegeneration in a patient is evaluated in different centres based on an MRI scan, there can be significant differences in the hippocampal volume measurement that can hinder accurate and timely clinical intervention; highlighting an urgent need for standardisation of imaging pipelines to ensure consistency in diagnosis, treatment and assuring patient safety.
Abstract
Dementia is characterised by the cumulative loss of cognitive and emotional abilities to extents that it disrupts everyday life. Dementia-causing diseases can induce structural and chemical changes in the brain leading to neuronal loss and brain volume shrinkage, and have a prolonged onset period which can go unnoticed for years before significant symptoms manifest or a diagnosis can be made. Thus, an early and accurate diagnosis for dementia and its subtypes is crucial to deliver appropriate patient treatment and care, but that is an enormous challenge. Current gold standards for diagnosis involve neuroimaging techniques such as Magnetic Resonance Imaging (MRI) to measure volumes in regions of the brain repeatedly over time, where hippocampal shrinkage is indicative of Alzheimer’s dementia. Brain volumetry typically involves segmentation algorithms to isolate the relevant areas of the brain. Rapid advancements in artificial intelligence (AI) have led to a plethora of segmentation methods, with many vendors providing AI tools in their medical imaging instrument software. AI is a powerful tool for clinicians to analyse data, though it lacks transparency and compounds the ‘black box’ issues. Critical information regarding the data pre- and post- processing, in addition to details of the AI model implementation, hyper-parameters, training data, et cetera, are not typically available. Moreover, these aspects can change over time with no warning to the users through software updates. There is a lack of standardisation between the segmentation algorithms resulting in variation between AI outputs and, highlighted in the work reported here, the measured brain volumes from MRI data. We investigate the differences between brain volumes extracted from the same patient scan at different centres to examine variability in the measured volumes. We use the Alzheimer’s Disease Neuroimaging Initiative data consisting of imaging data acquired under a standardised protocol curated at one site, with data processing and analysis occurring at several independent sites. We find a low degree of concordance between measurements of a patient’s hippocampal volumes between centres, with differences between centres exceeding thresholds for mild, moderate and severe dementia from Vijayakumar et al, ISRN Radiology [2003]. Comparing hippocampal volumes between three centres, we find that there are many cases where patients were diagnosed as cognitively normal or with mild cognitive impairment but had more than a 5 % or 12 % difference in volume between centres. This indicates that a patient evaluated at different centres risks receiving different diagnoses between the centres. This difference risks patients not being diagnosed and treated in a timely manner due to poor measurement of hippocampal volumes. Results indicate that hippocampal volume measurement between centres suffers inconsistencies due to a lack of standardisation in image post-processing, segmentation, and extraction pipelines at each centre. Our work illustrates that when neurodegeneration in a patient is evaluated in different centres based on an MRI scan, there can be significant differences in the hippocampal volume measurement that can hinder accurate and timely clinical intervention; highlighting an urgent need for standardisation of imaging pipelines to ensure consistency in diagnosis, treatment and assuring patient safety.