PROJECT SUMMARY
Dementia with Lewy body (DLB), Parkinsons Disease (PD) and Alzheimers Disease (AD) are among the most
debilitating neurodegenerative disorders that afflict patients in all countries and of all nationalities. One of the
Alzheimer related dementias (ADRD) national research priorities for DLB is to develop and validate imaging
techniques to improve the differential diagnostic accuracy of DLB versus other diseases. Magnetic Resonance
Imaging (MRI) is currently one of the most widely used diagnostic imaging techniques for detection of neuro-
degenerative disorders. However, standard T1 and T2-MRI may not provide the needed sensitivity and specificity
for differential diagnosis of DLB vs. AD and PD. Recently, Diffusion MRI (dMRI), specifically Diffusion Tensor
Imaging (DTI), has exhibited better sensitivity to the detection of some of these disorders. However, DTI is known
for its inability to cope with complex fiber geometries prevalent in the brain. This limitation can however be over-
come by using sophisticated mathematical models in conjunction with high angular resolution diffusion imaging
(HARDI). Our preliminary data suggests that learned micro-structural features from HARDI lead to high
sensitivity and specificity in differentiating PD vs. control and others in literature have shown discrimina-
tion between different stages of AD using macro-structural features derived from T1-MRI. This motivates
us to combine micro- and macro-structural features via a multi-modality approach to differentiate DLB
vs. PD, AD and controls. Differentiating between DLB, PD and AD is challenging because of possible overlap
in clinical symptoms leading to misdiagnosis. Further, differentiating between them is of high significance since
treatments including counseling for each are distinct. We propose a multi-modal approach that combines
the advantages of T1- and diffusion-MRI to achieve this goal. Recently, convolutional neural nets (CNNs)
have had great success in image classification tasks in computer vision and medical imaging. CNNs however
can not cope with HARDI data in its native form, which are samples of functions defined on non-Euclidean
(curved) domains. This motivates us to develop a novel higher order CNN that is a parameter efficient, inter-
pretable geometric deep learning network possessing improved model capacity, which we call the VolterraNet.
The VolterraNet will be designed for such data with the goal of facilitating the classification of DLB, PD and AD
groups. Further, VolterraNet will automatically localize the regions in the brain that are significantly discriminatory
of these patient groups. We will test the VolterraNet on HARDI scans acquired from a cohort of 356 Controls, 355
PD, 216 DLB and 240 AD scans obtained from a medley of data sites including the PDBP, 1Florida-ADRC and
PPMI. The VolterraNet will be validated using the standard leave-k-out cross-validation method with the precision
recall measure. The gold standard used will be the specialist-assigned clinical diagnosis from contributing stud-
ies (e.g. consensus assignment from ADRCs). The VolterraNet will have significant benefits to the Neurology
community through better detection and diagnosis of several neurodegenerative disorders.
