Sparse auto-encoded hypo-metabolism patterns in Alzheimer’s disease and mild cognitive impairment

Sidong Liu, Weidong Cai, Yang Song, Sonia Pujol, Ron Kikinis, Lingfeng Wen, Dagan Feng

Research output: Contribution to journalMeeting abstractpeer-review


Objectives: Analysis of hypo-metabolism patterns is important in the diagnosis of Alzheimer’s disease (AD) and mild cognitive impairment (MCI). Current studies of AD and MCI focus on individual brain regions, as opposed to correlation between multiple brain regions. It was our goal to identify dominant brain hypo-metabolism patterns based on region correlations in patients with AD and MCI.

Methods: A sparse auto-encoder with 9-hidden-neurons was designed for the non-linear analysis of hypo-metabolism patterns based on a 3-layer feed-forward neural network [1]. For each hidden neuron, a maximally activated 3D sparse auto-encoded brain template (SAEBT) encoding the non-linear region correlations was derived through L-BFGS optimization [2] when fed with a group of input data, i.e., the 81-dimensional region-wise feature vectors based on ICMB_152 atlas [3]. Two groups of subjects, AD and its pre-symptomatic state, MCI, were fed into our sparse auto-encoder, and 9 SAEBTs were derived from each group for analysis of the intra-group hypo-metabolism patterns and the inter-group differences between AD and MCI. The proposed method was tested on 3D image datasets (FDG-PET and T1-1.5T MRI) obtained from the ADNI baseline cohort [4] including 85 AD and 181 MCI cases.

Results: The dominant hypo-metabolism patterns among 9 AD-SAEBTs have strong positive correlations with hippocampal and parahippocampal regions, and strong negative correlations with precentral and orbital gyri. MCI-SAEBTs show similar patterns as AD, but with fewer positive correlated regions (AD: 33 vs. MCI: 24).

Conclusions: Our method automatically established the most representative SAEBTs based on the non-linear correlations between brain regions. Unlike other studies ignoring less important regions, SAEBTs take consideration of all regions and have the potential in discriminating AD patients from MCI patients.

Research Support ARC and AADRF grants.
Original languageEnglish
Article number1807
Number of pages1
JournalJournal of Nuclear Medicine
Issue numberSupplement 2
Publication statusPublished - May 2013
Externally publishedYes

Fingerprint Dive into the research topics of 'Sparse auto-encoded hypo-metabolism patterns in Alzheimer’s disease and mild cognitive impairment'. Together they form a unique fingerprint.

Cite this