Premium
Comparative analysis of cerebrospinal fluid markers and multimodal imaging in predicting Alzheimer’s disease progression
Author(s) -
Romano Michael F.,
Balachandra Akshara,
Zhou Xiao,
Jadick Michalina,
Qiu Shangran,
Nijhawan Diya,
Chin Sang P.,
Au Rhoda,
Kolachalama Vijaya B.
Publication year - 2021
Publication title -
alzheimer's and dementia
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 6.713
H-Index - 118
eISSN - 1552-5279
pISSN - 1552-5260
DOI - 10.1002/alz.054457
Subject(s) - neuroimaging , medicine , cerebrospinal fluid , modality (human–computer interaction) , lumbar puncture , magnetic resonance imaging , positron emission tomography , disease , neurodegeneration , radiology , pathology , neuroscience , artificial intelligence , psychology , computer science , psychiatry
Abstract Background Cerebrospinal fluid (CSF) levels of p‐tau, t‐tau, and Aβ42 are widely accepted in vivo biomarkers for Alzheimer’s Disease (AD). However, lumbar puncture comes with limitations, including severe infection, headache, bleeding at the site, a lack of spatial information about regions of neurodegeneration, and substantial inter‐lab variability. A less invasive method based on imaging for assessing AD progression can lead to insights on AD subtypes, decrease lumbar puncture‐related morbidity, and lower the psychological burden on misdiagnosed patients. Method We developed deep learning models composed of a fully convolutional network (FCN) linked with a multi‐layer perceptron (MLP) to predict the 2‐year risk of AD progression in individuals with mild cognitive impairment (MCI) using T1‐weighted MRI, fluorodeoxyglucose (FDG) PET, and florbetapir (amyloid) PET images, respectively, from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) (n=328). Images were all co‐registered within‐subject, normalized to MNI space, and skull‐stripped utilizing SPM‐12 prior to incorporating them in the model. A fusion model incorporating all three imaging modalities was evaluated to utilize the additive benefits of each imaging modality. An MLP utilizing only the above three CSF biomarkers was trained for comparison. Result Our FCN‐MLP model trained on FDG data alone achieved the highest accuracy of all three single imaging modality models. The model based on the amyloid scans alone was the most sensitive of all three single modalities, though it was the least specific. The MLP model constructed from CSF biomarkers was more specific than all imaging modalities. The fusion model had the highest accuracy out of all models, comparable sensitivity to the amyloid model, and similar specificity to the CSF model. It also had a higher F1 score than all other models. Conclusion Our fusion model predicts AD progression risk with superior accuracy and sensitivity, and a comparable specificity to a model constructed from CSF biomarkers alone. The FCN‐MLP framework could be extended to incorporate non‐invasive features that are easily obtainable in a memory clinic to develop more accurate risk models. Future studies can focus on augmenting the deep learning model performance by including tau‐based PET images as development of more disease specific tau tracers continues to evolve.