Thanks for your feedback, this will help us improve our content for you!
ARUK 2022 | Optimizing image segmentation tools for use in a clinical setting
Grace Gillis, MSc, University of Oxford, Oxford, UK, discusses her research aiming to integrate research quality measures into the diagnostic pathway for patients with dementia. The research aims to optimize automated segmentation tools so that they can better detect brain lesions and be used as a clinical tool to assess patients with dementia or possible dementia. After optimizing the automated segmentation tools, validation studies were carried out to test accuracy. Data from the UK Biobank was used to compare patient results to healthy people from the general population. The next steps will be to recruit more healthy, older volunteers to act as controls. This interview took place at the Alzheimer’s Research UK Conference 2022 in Brighton, UK.
Transcript (edited for clarity)
My work is based at the Oxford Brain Health Center, and what we try to do there is to integrate research quality measures into the diagnostic pathway for patients with dementia or with memory problems and they’re being investigated for dementia. So what I was trying to do in my project was to optimize the automated segmentation tools for these patients, because patients with memory problems tend to have greater levels of atrophy than healthy people, and also more extensive white matter lesions...
My work is based at the Oxford Brain Health Center, and what we try to do there is to integrate research quality measures into the diagnostic pathway for patients with dementia or with memory problems and they’re being investigated for dementia. So what I was trying to do in my project was to optimize the automated segmentation tools for these patients, because patients with memory problems tend to have greater levels of atrophy than healthy people, and also more extensive white matter lesions. These types of changes can cause some problems for the typical segmentation tools that we use. So what I wanted to do was to improve the tools for use in this setting. Then those quantitative measures, in turn, have the potential to provide a really useful information for the clinicians that then go on to diagnose whatever the cause of these memory problems truly is.
Essentially, there were three main aims for my project: to optimize the tools, then to validate their use in this memory clinic setting, and then to characterize the patients at the brain health center using these tools. So that’s essentially where we’ve gotten to now, although it is still a work in progress. In terms of tool optimization, I was able to improve the tool for the gray matter segmentation, for example, which is really one of the most important ones for people with memory problems, because it holds potential as a biomarker.
So the patients with memory problems tend to have greater lesions, as I said, and these lesions are misclassified as gray matter often. So what I did was to incorporate information from two different MRI modalities, the T1-weighted images and the T2-FLAIR images. And by incorporating both of those modalities, I could essentially improve the segmentation to exclude those lesions, which is what was being wrongly done in the first place. Those optimized segmentation volumes then provide a better kind of biomarker. Although we ultimately have to kind of evaluate now the efficacy of this biomarker, they at least visually appear much more accurate. So yeah, that’s kind of the first step all about optimizing.
I then wanted to validate the segmentations. So what I did was to correlate the quantitative volumes that I extracted with measures like the visual rating scale scores. So that’s what’s typically done in the memory clinic setting. And by correlating the quantitative measures that I derived with these visual rating scales, it shows kind of the clinical relevance of these measures and that they’re still quite accurate with respect to what’s normally done in the memory clinic. I also correlated with things like age and cognitive scores using the ACE, and this again showed the relatively high correlations, which support the validity of these segmentations.
Then just finally to characterize the brain health center patients, I used those quantitative measures that I got to then compare these patients to the general population, which I was able to kind of approximate using data from the UK Biobank. So as you could see on my poster, I have the reference distribution based on the UK Biobank, and then I could plot the brain health center patients on top of that, essentially getting a percentile. So you could say things like this patient falls in the fifth percentile, and that would be really clinician-friendly and valuable information to then incorporate hopefully into the radiology report that we send to the memory clinics to aid diagnosis.
So one of the biggest things that we want to do is to recruit more healthy, older individuals. So all of this work has been happening in the patients with memory problems, but to really optimize the potential of these percentiles, like I was talking about, we need to have the older, healthy controls, so we can kind of extend that distribution. The UK Biobank population is capped a little bit over 70, but our patients go up to 100. So it’s, right now, really difficult to compare those patients to the reference distribution from the UK Biobank.
So what we’re trying to do now is to kind of recruit more healthy individuals to extend that distribution, so we can compare more of our patients to the general population. That’s kind of one future direction. Another would be, as I said, to really integrate these percentile measures into the radiology report and then evaluate long term, if these quantitative measures have value when it comes to diagnosis and predicting long-term outcomes too.