Brain Disorders (Mar 2021)

AI4AD: Artificial intelligence analysis for Alzheimer's disease classification based on a multisite DTI database

  • Yida Qu,
  • Pan Wang,
  • Bing Liu,
  • Chengyuan Song,
  • Dawei Wang,
  • Hongwei Yang,
  • Zengqiang Zhang,
  • Pindong Chen,
  • Xiaopeng Kang,
  • Kai Du,
  • Hongxiang Yao,
  • Bo Zhou,
  • Tong Han,
  • Nianming Zuo,
  • Ying Han,
  • Jie Lu,
  • Chunshui Yu,
  • Xi Zhang,
  • Tianzi Jiang,
  • Yuying Zhou,
  • Yong Liu

Journal volume & issue
Vol. 1
p. 100005

Abstract

Read online

Background: Diffusion tensor imaging (DTI) has been widely used to identify structural integrity and to delineate white matter (WM) degeneration in Alzheimer's disease (AD). However, the validity and replicability of the ability to discriminate AD patients and normal controls (NCs) of WM measures are limited due to the use of small cohorts and diverse image processing methods. As yet, we still do not have a clear idea of whether WM characteristics are biomarkers for AD. Methods: We conducted a competition with diffusion measurements along 18 fiber tracts as features extracted via the automated fiber quantification (AFQ) method based on one of the largest worldwide DTI multisite biobanks (862 individuals, consisting of 279 NCs, 318 ADs, and 265 MCIs). After quality control, 825 subjects (276 NCs, 294 ADs, and 255 MCIs) were divided into a public training set (N=700) and a private testing set (N=125). Forty-eight teams submitted 130 solutions that were estimated on the private testing samples. We reported the final results of the top ten models. Results: The performance of white matter features in AD classification was stable and generalizable, which indicated the potential of WM to be a biomarker for AD. The best model achieved a prediction accuracy of 82.35% (with a sensitivity of 86.36% and a specificity of 78.05%) on the private testing set. The average accuracy of the top ten solutions was over 80%. Conclusions: The results of this competition demonstrated that DTI is a powerful tool to identify AD. A larger dataset and additional independent cohort cross-validation may improve the discriminant performance and generalization power of the classification models, thus revealing more precise disease severity factors associated with AD. For this purpose, we have released this database (https://github.com/YongLiuLab/AI4AD_AFQ) to the community, with the expectation of new solutions for the accurate diagnosis of AD.

Keywords