hide
Free keywords:
-
Abstract:
Purpose or Learning Objective: Alzheimer's disease (AD) is the most common type of an irreversible neurodegenerative disorder,
affecting millions of people. Especially early stratification of patients with mild cognitive impairment (MCI) into patients who will
convert to AD remains a challenging task. We aimed to predict automatically whether MCI patients will develop the disease (MCIc) by
following subjects over time and quantifying spatial atrophy rates (AR) in magnetic resonance imaging (MRI).
Methods or Background: 3D T1w MRIs at 3T from 276 MCI patients participating in the first period of Alzheimer’s Disease National
Initiative (ADNI-1) with at least two MRIs more than 60 days apart without evident artifacts were segmented by a deep-learning-based
3D-UNet into 30 anatomical regions. Z-scores of TIV-adjusted volumes were calculated compared to a normal reference population,
and AR of these z-scores were calculated longitudinally per subject (AR=0 normal aging). Rolling AR were calculated as the mean AR
over a half-year time window (mRAR). A 80:20 train-test-partition was used to train a logistic regression to discriminate MCIc vs
MCInc.
Results or Findings: We found accelerated regional mRAR in MCIc. The temporal cortex and hippocampal regions showed the most
striking mRAR. On the test set, of 34 MCIc, the classifier predicted 27 as true positive with a median of 1.7 Y (Q1/3=2.0/0.6Y) before
conversion (sensitivity=0.79), with 5/22 false positives MCInc (stable specificity=0.77, AUC ROC=0.81).
Conclusion: Our method provides reliable results due to a stable specificity that can be obtained well before previous clinical
diagnoses for conversions to disease. Therefore, it is suitable for use in subsequent studies.
Limitations: Validation in an independent sample is missing.