Category: Neuroimaging (Non-PD)
Objective: To develop a neural network architecture based on multimodal MRI brain imaging to discriminate between multiple system atrophy (MSA) and healthy controls (HC), and validate it against a previously published method.
Background: MSA is a rare neurodegenerative disorder. The differential diagnosis from Parkinson’s disease and other parkinsonian syndromes can be challenging, especially in the early stages of the disease. Currently, there are no objective biomarkers to distinguish in vivo MSA from HC. Previous works presented a data-driven method based on machine learning and multi-modal MRI imaging to discriminate between MSA and HC. An open question is if artificial neural network could be suitable for this kind of task.
Method: This study presents a complete multimodal MRI protocol including structural, diffusion and functional MRI in 29 MSA patients and 26 HC. We used the same dataset and validation scheme as in our previously published study  to validate our deep learning approach adopting a convolutional neural network model based on the adaptation to 3D data of” VGG” models  composed of a sequence of convolution and pooling layers. We aimed to test if a neural network approach is feasible in a small sample compatible with clinical practice.
Results: The proposed three-dimension multi-modal Convolutional Neural Network (CNN) allowed to obtain a best accuracy of 89.5 ± 1.6 % (91.4 ± 1.7 % and 87.3 ± 3.0 % for sensitivity and specificity, respectively). We developed an interpretation method of our model that highlights areas that lead to the model decision, allowing the visualization of a topological signature of the pathology compatible with the known pathophysiology of the disease. The most discriminant modality was the mean diffusivity modality in the cerebellum and the putamen.
Conclusion: Our proposed neural network architecture can successfully discriminate between MSA and HC in a small sample. Moreover, information about the discriminating power of the different indexes used and the topological signature agree with our previously published study that used the same dataset (with a support vector machine approach) and more generally to the known pathophysiology of MSA. This approach is promising to help diagnosis prospectively at early stages.
References:  Nemmi F, Pavy-Le Traon A, Phillips OR, Galitzky M, Meissner WG, Rascol O, Péran P. A totally data-driven whole-brain multimodal pipeline for the discrimination of Parkinson’s disease, multiple system atrophy and healthy control. Neuroimage Clin. 2019 May 13;23:101858. doi: 10.1016/j.nicl.2019.101858.  Simonyan, K., & Zisserman, A. (2014). Very Deep Convolutional Networks for Large-Scale Image Recognition. Retrieved from http://arxiv.org/abs/1409.1556
To cite this abstract in AMA style:E. Villain, F. Nemmi, A. Pavy-Le Traon, W. Meissner, O. Rascol, M.V Le Lann, X. Franceries, P. Péran. Convolutional neural network for discriminating between Multiple System Atrophy and Healthy Control, comparing MRI modalities and highlighting the disease signature [abstract]. Mov Disord. 2020; 35 (suppl 1). https://www.mdsabstracts.org/abstract/convolutional-neural-network-for-discriminating-between-multiple-system-atrophy-and-healthy-control-comparing-mri-modalities-and-highlighting-the-disease-signature/. Accessed December 11, 2023.
« Back to MDS Virtual Congress 2020
MDS Abstracts - https://www.mdsabstracts.org/abstract/convolutional-neural-network-for-discriminating-between-multiple-system-atrophy-and-healthy-control-comparing-mri-modalities-and-highlighting-the-disease-signature/