Premium
O4‐04‐02: IDENTIFICATION OF FAST CONVERTERS AMONG MCI PATIENTS BY INTEGRATING INCOMPLETE MULTI‐MODALITY IMAGE DATA USING A LONGITUDINAL TRANSFER LEARNING MODEL
Author(s) -
Liu Xiaonan,
Li Jing,
Chen Kewei,
Wu Teresa,
Lure Fleming,
Su Yi,
Weidman David,
Wang Peijun
Publication year - 2019
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 - 10.1016/j.jalz.2019.06.4759
Subject(s) - dementia , modality (human–computer interaction) , modalities , classifier (uml) , medicine , cognitive impairment , artificial intelligence , magnetic resonance imaging , nuclear medicine , computer science , radiology , cognition , social science , disease , psychiatry , sociology
structural MRI alone for A/T/N staging. A multi-class convolutional neural network (MC-CNN) model for simultaneous prediction of A/T/N stages was trained on 5000+ structural-MRIs from ADNI cases with known "A,” "T,” and "N” status measured by CSFAb42, CSF p-tau, and FDG-PET, respectively. A/T/N staging for each structural MRI casewas deemed normal or abnormal based on predefined CSF and FDG-PET biomarker cutoffs[Jack et al. Lancet Neurol. 2017]. The predictive performance of the MCCNN model was assessed on an independent validation cohort of 2000+ ADNI cases with known A/T/N staging based on the same CSF and FDG-PET biomarker cutoffs. Both training and independent validation cohorts included cases from different clinical diagnostic groups, including healthy controls, MCI, and AD. Results: MC-CNN model predicted "A” at an overall accuracy of 88%, "T” at 89%, and "N” at 95%; model performance varied across the clinical diagnostic groups with improved performance in symptomatic cases as reported in the Table. MC-CNNmodel performed relatively better for predicting "N” stage (i.e., virtually perfect "N” staging in AD cases), consistent with the widely accepted notion of structural MRI being a surrogate marker of neurodegeneration. Conclusions: Using structural-MRI alone and an artificial intelligence approach, we have developed a unified model that can predict A/T/N stages simultaneously for a spectrum of individuals ranging from healthy control toMCI and AD. The model could be used as an image-screening tool to comprehensively classify individuals into A/T/N stages and potentially facilitate cohort enrichment. In future studies, personal demographics or clinical measures could be included to boost the performance of the model further.