Introduction
The increasing health costs and lost productivity from neurodegenerative diseases such as Alzheimer’s disease (AD), exacerbated by the ageing population, is a major National problem estimated to be $5.6b in 2002, including $3.2b in direct health costs (Access Economics 2005). The population >65 is expected to increase from 13% to 30% by 2040, with 54,000 new AD cases estimated in 2006 alone (of 210,000 new dementias and with 730,000 anticipated by 2040; CEDA). Delaying onset of the major dementia (i.e. AD) by 5 years could reduce new cases by 50% with cumulative health cost savings of up to $13.5 billion by 2020 (Access Economics 2005 estimation).
The causes of dementia are not well understood and current diagnosis is difficult because there are as yet no known biological markers. The relatively advanced loss of cognitive function necessary for current clinical diagnosis of dementia generally results in irreversible neuronal dysfunction. If objective evidence of AD pathological lesions could be found early (before there is evidence of cognitive or behavioural symptoms), appropriate treatment and care could be provided, resulting in delayed onset or prevention of AD.
The Biomedical Imaging team at the Australian e-Health Research Centre (AEHRC) is developing key technologies for in vivo quantitative assessment of Amyloid-Α (Aß) deposition suspected to be an early marker of AD. We are collaborating through CSIRO’s Preventative Health Flagship with AIBL, the Australian Imaging, Biomarker and Lifestyle cluster study. AIBL has enrolled more than 1000 volunteers for psycho-cognition and blood biomarker evaluation and includes 200 volunteers who will have brain-image scans using Pittsburgh compound B (11C-PIB), a novel Positron Emission Tomography (PET) biomarker. Other information from different imaging modalities will also be acquired from those volunteers such as fluorodeoxyglucose (18FDG) PET imaging, and Magnetic Resonance Imaging (MRI).
Specifically, the Biomedical Imaging team is developing a library of image processing algorithms that can be called from our core software MilxView, a generic medical imaging viewer. From MilxView individual or large batch of image analysis tasks can be scheduled and run automatically. This white paper lists some of those technologies.
AIBL Consortium
The Australian Imaging, Biomarker and Lifestyle (AIBL) Flagship Study of Ageing study aims to improve understanding of the causes and diagnosis of Alzheimer’s disease, examine lifestyle and diet factors that may influence its onset, and help develop preventative strategies. It involves CSIRO, through its Preventative Health National Research Flagship, working with the University of Melbourne (Austin Health PET Centre), the Edith Cowan University (WA), Neurosciences Australia and the Mental Health Research Institute of Victoria. The study was launched on November 14, 2006, at Neurosciences Australia in the National Neurosciences Facility, Melbourne.
The three-year prospective longitudinal study of ageing will involve recruiting 1,000 volunteers from a cross-section of Australia’s population. It is the largest study in the world involving PET scans using PIB, a PET amyloid-imaging agent.
PIB is a radioactive marker used for in vivo imaging of Aß deposits using PET. PIB is derived from Thioflavin T, a dye used to stain Aß plaques in histology. PIB has several properties that make it attractive for PET imaging: it bounds to insoluble species of Amyloid with high affinity, it crosses the blood-brain barrier and enters brain in sufficient amount to be imaged using PET, and it clears rapidly from normal tissues as shown by several animal and human studies.
Example of PET-PIB scans of an Alzheimer’s patient and a normal control
Multimodality Imaging
Over 200 patients from the AIBL study will be imaged using Magnetic Resonance Imaging and PET including at least 40 patients with Mild AD, 40 patients with Mild Cognitive Impairment ( MCI), and about 120 normal elderly controls (NC). In addition to PET-PIB, several scans will be performed including PET-FDG to assess brain function, and a series of Magnetic Resonance Imaging contrasts for anatomical characterization (T1W, PDW, T2W), structural integrity of the white matter (DWI), and pathological imaging (FLAIR, SWI). Patients will be scanned at the Brain Research Institute and the Austin Health PET Centre in Melbourne, as well as the Hollywood Private Hospital and the Royal Perth Hospital.
Multimodality imaging provides complimentary information. From left to right on the same patient: PET-FDG (neronal activity), MRI-T1W (anatomical delineation), PET-PIB (Aß plaques deposition).
PET Normalisation
PET scanners measure radioactivity of markers such as FDG and PIB, and the images intensity need to be normalised to allow quantitative analysis and comparison between patients. We are developing automatic methods to perform the normalisation which is usually done manually in a very tedious and long process. For example in PET-PIB, our method automatically extracts the cerebellum gray matter and uses its mean intensity to normalise signal intensity across all the patients of our database.
Automatic segmentation from PET-PIB of the cerebellum cortex (shown in white overlay over the PET-PIB) for intensity normalisation across patient population.
Inter-patient Registration
In order to relate patient information from the different imaging modalities, an essential step is to spatially align all the image volumes together and to a template where brain structures are identified. We used a robust automatic technique that maximizes normalized cross-correlation within a block matching scheme. Further used of a multi-resolution technique allows to register fully automatically a patient dataset in less than one minute.
Rigid body registration of a patient MRI T1W with its corresponding PET-PIB in less than one minute.
Population Specific Atlas Generation
We have also implemented a robust non rigid registration technique to warp different patients to a common template. We use a free form deformable method which models the deformation field with B-Spline that are constrained to enforce smoothness. For example, using this technology all the MRI scans of our patient’s population can be warped to a MRI template allowing statistical methods to be applied such as voxel based morphometry. We also use this technique to build atlases and templates that are representative of our patient population, in particular matched for age. We found that those custom-build templates help tremendously further statistical analysis avoiding the bias associated with using a standard template or a unique individual as is often the case.
The left panel shows the widely used Colin27 template, whereas the right panel shows a template computed from our normal control patient population. Note the enlarged ventricles characteristics of the advanced age of our patient cohort.
Cortical Thickness Estimation from MRI
MRI is a powerful method to image with exquisite anatomical details the different tissues of the brain. It has become an essential tool to assess Alzheimer’s because of the cortical atrophy associated with neuro-degenerescence, which can be measured by MRI. In particular, relatively fast T1W acquisitions offer very good contrast between white and gray matters allowing the precise delineation of the outer cortical mantel. However, manually measuring gray matter thickness on a high resolution 3D dataset (1 mm isotropic) is not only tedious, but prone to error. We have developed algorithms to perform fast and accurate automatic segmentation of the brain into its main tissues type and estimate the thickness of the gray matter in 3D without any human intervention. We first model the tissue intensities using normal distributions, before estimating the fractional tissue content of each pixel. The last step implements a novel technique to measure the cortical thickness using a multi-layer physical model. Because of our ability to spatially register each patient to a common template, we are able to study the statistical significance of brain atrophy in our patient population and benchmark any new individual MRI scan against its "normal" template.
Example of cortical thickness estimation. From the top left, panels show the original MRI-T1W of a patient, the fractional gray matter content estimated for each pixels, and the final thickness estimation.
The two brain renderings below show the average thickness for the AD and the NC group for different brain areas.
Towards a Complete Solution for Early Diagnostic of AD
Our vision is to provide clinicians with a software tool easy to use and fully automatic. The physicians would read the scans from a patient, and we would compute several quantitative measurements from the images otherwise hard or impossible for a human to obtain. From our study on the AIBL patients we will be able to benchmark each individual patient against the typical age-matched individual and provide to physicians relevant statistics. This kind of information should be valuable not only to help in treatment planning of individual patients, but also to help design therapies and test scientific hypothesis.
