PosterPDF Available

Collaborative study in deep learning for predicting disease activity in multiple sclerosis (deepMS)

Authors:

Abstract

Objective and Aims: Collecting and procuring an MRI and clinical data repository from centres around the world for collaborative, expert scientiic contributions in predicting disease activity in MS patients. With this data repository we aim to develop new MRI-derived clinically applicable biomarkers using a combination of convolutional neural networks (CNNs) and visualisation techniques. Methods: We will be collecting the following data: a) 3D T1-, T2-weighted brain MRIs and optionally: T2*-EPI, T2-weighted spine; b) associated clinical data including: diagnosis, age, sex, ethnicity, initial attack type & date, EDSS, MSFC and optionally: lesion load, NEDA-3 Results (from pilot study, see reference (1)): A classification accuracy of 87% was reached, when CNN was fine-tuned from Alzheimer's Disease and transferred to MS data. The algorithm independently learned patterns in MRIs to separate MS patients from healthy participants. Visualisation analysis revealed that CNNs focused not only on lesion voxels, but also incorporates information from lesion location, as well as non-lesional white and grey matter areas, such as the thalamus. Conclusions: Initial results show immense promise in using CNNs to automatically evaluate brain MRIs in MS patients. In this collaborative study, we aim to extend our methods now for the more complex task of predicting individualized disease activity and prognosis (Figure 1). For robust, reliable, and clinically relevant future applications, further data contribution is required. Together with neurologists, neuroradiologists, and researchers, the data will be analysed in a collaborative, expert-led, and world-wide manner. References: 1. Eitel F, Soehler E, Bellmann-Strobl J, Brandt AU, Ruprecht K, Giess RM, et al. Uncovering convolutional neural network decisions for diagnosing multiple sclerosis on conventional MRI using layer-wise relevance propagation. arXiv:190408771 [cs] [Internet]. 2019 Sep 5, accepted in NeuroImage: Clinical Funding: This project is supported by the Deutsche Multiple Sklerose Gesellschaft Bundesverband e.V., Deutsche Forschungsgemeinschaft (389563835) and Manfred and Ursula Müller-Stiftung.
ECTRIMS 2019
P1665
Collaborative study in deep learning for predicting
disease activity in multiple sclerosis (deepMS)
Claudia Chien, Fabian Eitel, Alexander Brandt, Friedemann Paul,
Kerstin Ritter
Experimental and Clinical Research Center, Max Delbrück Center for Molecular Medicine & Charité-Universitätsmedizin Berlin, Corporate Member of Freie Universität Berlin,
Humboldt-Universität zu Berlin, and Berlin Institute of Health
NeuroCure Clinical Research Center, Charité-Universitätsmedizin Berlin, Corporate Member of Freie Universität Berlin, Humboldt-Universität zu Berlin, and Berlin Institute of Health
Department of Psychiatry and Psychotherapy, Berstein Center for Computational Neuroscience, Berlin Center for Advanced Neuroimaging, Charité-Universitätsmedizin Berlin,
Corporate Member of Freie Universität Berlin, Humboldt-Universität zu Berlin, and Berlin Institute of Health
Department of Neurology, University of California Irvine, Irvine, CA, USA
Department of Neurology, Charité-Universitätsmedizin Berlin, Corporate Member of Freie Universität Berlin, Humboldt-Universität zu Berlin, and Berlin Institute of Health
Objective and Aims:
Collecting and procuring an MRI and clinical data repository from centres around the world for collaborative, expert
scientic contributions in predicting disease activity in MS patients. With this data repository we aim to develop new
MRI-derived clinically applicable biomarkers using a combination of convolutional neural networks (CNNs) and
visualisation techniques.
Methods:
Results (from pilot study, see reference (1)):
- a classication accuracy of 87%, when ne-tuned from
Alzheimer's Disease and transferred to MS data
- the algorithm independently learned patterns in MRIs
to separate MS patients from healthy participants
- visualisation analysis revealed that CNNs focused not
only on lesion voxels, but also incorporates information
from lesion location, as well as non-lesional white and
grey matter areas, such as the thalamus.
Conclusions:
Initial results show immense promise in using CNNs to
automatically evaluate brain MRIs in MS patients. In this
collaborative study, we aim to extend our methods now
for the more complex task of predicting individualized
disease activity and prognosis (Figure 1). For robust,
reliable, and clinically relevant future applications,
further data contribution is required. Together with
neurologists, neuroradiologists, and researchers, the
data will be analysed in a collaborative, expert-led, and
world-wide manner.
Learned
Model
Visualisation
References:
1. Eitel F, Soehler E, Bellmann-Strobl J, Brandt AU, Ruprecht K,
Giess RM, et al. Uncovering convolutional neural network
decisions for diagnosing multiple sclerosis on conventional
MRI using layer-wise relevance propagation. arXiv:190408771
[cs] [Internet]. 2019 Sep 5, accepted in NeuroImage: Clinical
Funding:
This project is supported by the Deutsche Multiple Sklerose
Gesellschaft
Bundesverband e.V., Deutsche Forschungsgemeinschaft
(389563835) and Manfred and Ursula Müller-Stiftung.
Prediction Disease
Activity
Data Required:
- 3D T1-, T2-
weighted
brain MRIs
Optional: T2*-
EPI,
T2-weighted
spine
- associated
clinical data
including:
diagnosis, age,
sex, ethnicity,
initial attack
type & date,
EDSS, MSFC
Optional: lesion
load, NEDA-3
Single Patient
Test &
Training Data
Personalised Prognosis
Clinical & Radiological
Input
Figure 1. Overview of the methods and aims of deepMS.
Conicts of Interest:
CC has nothing to disclose.
FE has nothing to disclose.
AUB is cofounder and holds shares of Motognosis GmbH and Nocturne GmbH. He is
named as inventor on several patent applications describing multiple sclerosis serum
biomarkers, perceptive visual computing-based motion analysis and retinal image ana-
lysis; none of this is related to the present abstract.
FP has received research support from Bayer, Novartis, Biogen Idec, Teva, Sano-
Aventis/Genzyme, Merck Serono, Alexion, Chugai, Arthur Arnstein Foundation Berlin,
Guthy Jackson Charitable Foundation and the US National Multiple Sclerosis Society;
has received travel funding and/or speaker honoraria from Bayer, Novartis, Biogen Idec,
Teva, Sano- Aventis/Genzyme and Merck Serono; and has consulted for Sano Genzy-
me, Biogen Idec and MedImmune; none of this is related to the present abstract.
KR has nothing to disclose.
{
Visit us here!
ResearchGate has not been able to resolve any citations for this publication.
ResearchGate has not been able to resolve any references for this publication.