Brain Disease Diagnosis Using Deep Learning Features from Longitudinal MR Images

Deep learning-based brain disease diagnoses utilizing magnetic resonance (MR) images has attracted increasing attention in the field of computer-aided diagnosis. However, most existing methods require computationally expensive preprocessing before feature extraction, such as 3D MR image registration and landmark detection. Additionally, these methods only employ cross-sectional MR images. Recent studies have demonstrated that longitudinal images acquired at different time points can comprehensively reflect the pathological changes of diseases. To date, effectively capturing information from variable numbers of longitudinal MR images has not been adequately investigated. In this study, we propose a deep learning method taking advantage of longitudinal MR images for disease diagnoses. In particular, we first extract features from slice images employing a Deep Convolutional Neural Network (DCNN) in an end-to-end manner. This avoids 3D image registration and landmark detection. We then generate longitudinal-level features by using the Bag-of-Words (BoW) model. Lastly, we devise a Recurrent Neural Network (RNN) to capture the pathological changes for facilitating disease diagnoses. We evaluate the proposed method on the public Alzheimer’s Disease National Initiative (ADNI) dataset. Extensive experiments show that the proposed method is superior to baseline methods and is robust to both the Alzheimer’s disease (AD) and mild cognitive impairment (MCI) diagnoses. Moreover, the proposed method can effectively learn pathological changes from the longitudinal MR images for disease diagnosis.