A hybrid Convolutional and Recurrent Neural Network for Hippocampus Analysis in Alzheimer's Disease

Background: Hippocampus is one of the first structures affected by neurodegenerative diseases such as Alzheimer's disease (AD) and mild cognitive impairment (MCI). Hippocampal atrophy can be evaluated in terms of hippocampal volumes and shapes using structural MR images. However, the shape and volume features from hippocampus mask have limited discriminative information for AD diagnosis. In addition, extraction of these features is independent of classification model, resulting to sub-optimal performance for disease diagnosis.

New method: This paper proposes a hybrid convolutional and recurrent neural network for more detailed hippocampus analysis using structural MR images in AD. The DenseNets are constructed on the decomposed image patches of internal and external hippocampus to learn the intensity and shape features. Recurrent neural network (RNN) is cascaded to combine the features from the left and right hippocampus and learn the high-level features for disease classification.

Results: Our proposed method is evaluated with the baseline MR images of 807 subjects including 194 AD, 397 MCI and 216 normal controls (NC) from Alzheimer's Disease Neuroimaging Initiative (ADNI) database. Experiments show the proposed method achieves AUC (area under ROC curve) of 91.0%, 75.8% and 74.6% for classifications of AD vs. NC, MCI vs. NC and pMCI vs. sMCI, respectively.

Comparison with existing methods: The proposed method achieves better performance than the volume and shape analysis methods.

Conclusions: A hybrid convolutional and recurrent neural network was proposed by combining DenseNets and bidirectional gated recurrent unit (BGRU) for hippocampus analysis and AD diagnosis. Results show its promising performance.