RNN-Based Dementia Disease Prediction from Prodromal Stage using Diffusion
Tensor Imaging 
   Lead Investigator:    Matthew Velazquez
   Institution      :    UMKC
   E-Mail           :    mv3md@umkc.edu
   Proposal ID      :    1371



   Proposal Description:
   The goal of our proposed research is to build a Recurrent Neural Network that can predict Dementia prognosis based on diffusion tensor imaging from the mild cognitive impairment stage. This research is an extension on our recently presented paper at BIBM 2019: Prodromal Stage using Diffusion Tensor Imaging.    In that work we were able to determine at 96.5 accuracy whether someone diagnosed with Mild Cognitive Impairment would go on to develop Alzheimer's Disease. This extension is concerned with broadening the classes to include the other dementia types rather than only being a binary problem around Alzheimer's.    Our hypothesis is that there are detectable differences in white molecule diffusion at the MCI stage between dementia sub-types. As this stage typically occurs years before a more specific dementia diagnosis, this research could lead to more accurate treatment plans with a better estimated prognosis.    Currently most of the deep-learning applications on MRI images are concerned with the classification problem (i.e. Is this scan normal, MCI, or AD). In terms of the few that focus on prediction, most of their work has centered around sMRI or PET images. Their method is to train their neural network with enough scanned images that their network can determine on a fresh image the likelihood of that individual eventually being diagnosed with Alzheimer's Disease.  Our model is also built on computer vision and trained accordingly, however, we focus on Diffusion Tensor Images as we've noticed significantly higher accuracy when allowing our neural network to observer the water molecule diffusion in the brain.  We originally used the ADNI dataset which was a longitudinal study where we could see whether patients would eventually develop AD or not. With additional classes we would be able to expand our binary problem so that our solution does more than only rule out Alzheimer's. Our initial publication on this work saw our accuracy at 96.5 for predicting Alzheim