Quantifying Neurodegenerative Progression With DeepSymNet, an End-to-End Data-Driven Approach

Abstract

Alzheimer's disease (AD) is the most common neurodegenerative disorder worldwide and is one of the leading sources of morbidity and mortality in the aging population. There is a long preclinical period followed by mild cognitive impairment (MCI). Clinical diagnosis and the rate of decline is variable. Progression monitoring remains a challenge in AD, and it is imperative to create better tools to quantify this progression. Brain magnetic resonance imaging (MRI) is commonly used for patient assessment. However, current approaches for analysis require strong a priori assumptions about regions of interest used and complex preprocessing pipelines including computationally expensive non-linear registrations and iterative surface deformations. These preprocessing steps are composed of many stacked processing layers. Any error or bias in an upstream layer will be propagated throughout the pipeline. Failures or biases in the non-linear subject registration and the subjective choice of atlases of specific regions are common in medical neuroimaging analysis and may hinder the translation of many approaches to the clinical practice. Here we propose a data-driven method based on an extension of a deep learning architecture, DeepSymNet, that identifies longitudinal changes without relying on prior brain regions of interest, an atlas, or non-linear registration steps. Our approach is trained end-to-end and learns how a patient's brain structure dynamically changes between two-time points directly from the raw voxels. We compare our approach with Freesurfer longitudinal pipelines and voxel-based methods using the Alzheimer's Disease Neuroimaging Initiative (ADNI) database. Our model can identify AD progression with comparable results to existing Freesurfer longitudinal pipelines without the need of predefined regions of interest, non-rigid registration algorithms, or iterative surface deformation at a fraction of the processing time. When compared to other voxel-based methods which share some of the same benefits, our model showed a statistically significant performance improvement. Additionally, we show that our model can differentiate between healthy subjects and patients with MCI. The model's decision was investigated using the epsilon layer-wise propagation algorithm. We found that the predictions were driven by the pallidum, putamen, and the superior temporal gyrus. Our novel longitudinal based, deep learning approach has the potential to diagnose patients earlier and enable new computational tools to monitor neurodegeneration in clinical practice.

Toward a Multimodal Computer-Aided Diagnostic Tool for Alzheimer’s Disease Conversion

“…The imaging feature importance was completed on both a voxel-wise level and a brain regional level (subcortical vs. cortical regions), as seen in Figure 6. These saliency maps are smoother than our previous work (Pena et al, 2019), and we attributed this improvement to the use of attention-based modules and a less parameterized network. These model characteristics perform significant regularization, highlighting only the most informative regions for the given task.…”

“…As shown in Figure 1 , the T1-weighted MRI images were pre-processed according to the steps outlined in our previous work ( Pena et al, 2019 ). In summary, the two images at two time points were normalized and aligned to each other first and then registered to a common space using a linear registration algorithm.…”

“…The final common interpatient space was derived from 2 mm MNI T1 template which was cropped of the background space to reduce the computational complexity of the network for a final resolution of 64 × 80 × 64. This pipeline was shown to drastically decrease the pre-processing time compared with conventional image processing pipelines such as the wull FreeSurfer-based ones ( Pena et al, 2019 ). These steps were extended to the full MCI cohort used in this study.…”

“…The network architecture employed was inspired by a model that learned from spatial symmetry between brain hemispheres in the stroke detection task ( Barman et al, 2019 ; Sheth et al, 2019 ). Our previous work extended this model in the AD-progression and time domain ( Pena et al, 2019 ). This study has implemented a new network that combines cross-sectional and longitudinal imaging data with clinical features, which can be trained end-to-end on the MCI conversion classification task.…”

“…These layers were composed of three parallel operations: 1 × 1 × 1, 3 × 3 × 3, and 5 × 5 × 5 convolutions with two filters. Previous work has shown that this module can produce meaningful results in neuroimaging applications (Barman et al, 2019(Barman et al, , 2020Pena et al, 2019). This style of operation allows the model to view an image or input at different scales to learn different types of spatial information.…”