P007 AI4MS: A Deep Learning Approach for Multimodal Prediction of Multiple Sclerosis Progression
Shailesh Appukuttan*1,2, Adrien Amberto1, Mounir Mohamed El Mendili1, Bertrand Audoin1,3, Ismail Zitouni1,Audrey Rico1,3, Hugo Dary1, Maxime Guye1,Jean-Philippe Ranjeva1,Ronan Sicre4,Jean Pelletier1,3,Wafaa Zaaraoui1, Matthieu Gilson2, Adil Maarouf1,3
1Aix Marseille Univ, CNRS, CRMBM, Marseille, France
2Aix Marseille Univ, CNRS, INT, Marseille, France
3APHM, Hôpital de la Timone, Maladie Inflammatoire du Cerveau et de la Moelle Epinière (MICeME), Marseille, France
4University of Toulouse, CNRS, IRIT, France.
*Email: shailesh.appukuttan@univ-amu.fr
Introduction:
Multiple Sclerosis (MS) is a chronic neurological disorder of the central nervous system. Disease progression in MS can be highly variable. Reliable prediction of disease progression has a huge impact on optimizing individualized treatment plans. Traditionally, MRI-based assessments rely heavily on clinical expertise. However, with the notable advancements in the field of AI in recent times, AI-based approaches offer potential for improving the accuracy and reproducibility of such predictions [1]. With the AI4MS project, we aim to develop and validate a deep-learning model that integrates multimodal MRI and clinical data to improve MS prognosis prediction. Our approach incorporates advanced deep learning architectures to enhance predictive power, with a focuson clinical applicabilityby targeting explainable models.
Methods:
Inthis project we leverage a cohort of 300+ MS patients that have been followed for over 10 years. We have access to multimodal MRI (T1w, T2w) as well as the associated clinical data (such as EDSS and MSFC scoresthat quantify disease severity) [2]. The deep-learning model employs 3D ResNetextracting spatial features from the MRI images, while a bidirectional recurrent network (GRU) with time-aware attention is used to incorporate temporal dynamics.The decision of the model is explained by the means of a saliency mapthat identifies parts of the images influencing the classification,obtained with a CAM-basedinterpretabilitymethod [3].
Results:
In our preliminary tests, we use CNN-based models to predict the Sustained Accumulation of Disability (SAD) [4] using data from a subset of the patients (n = 104) and only employing the EDSS clinical scores. Data is grouped into triplets of visits to capturehow the disease progresses over time.We systematically test different models to evaluate the prediction capability of each MRI modality, as well as data selection / augmentation on the cross-validated classification accuracy to test the generalization capability of the prediction pipeline.The study also suggests the need for incorporating additional clinical measures (e.g., MSFC scores) and MRI-based metrics to capture a more holistic representation of disease progression.
Discussion:
The AI4MS project aims to build on our preliminary findings and overcome its limitations. We adopt a more multimodal approach by integrating diverse clinical and imaging data. The model is developed in a modularized manner, with spatial and temporal components being trained separately. This promises to ensure better learning and efficiency. Visualization tools, such as heatmaps and saliency maps, are incorporated to enhance interpretability of the model predictions. The project also explores various data augmentation techniques to address any problems of data scarcity and imbalance. The AI4MS project aims to assist clinicians with reliable predictions to guide individualized treatment plans for MS patients.
Acknowledgements
All MRI acquisitions were funded by Fondation ARSEP. This project has received funding from the Excellence Initiative of Aix-Marseille Université - AMidex, a French “Investissements d’Avenir programme” AMX-21-IET-017 (via the institutes NeuroMarseille and Laënnec). We would also like to thank AMU mésocentre for access to HPC resources.
References
[1]https://doi.org/10.1038/s41591-018-0300-7
[2]http://doi.org/10.1186/1471-2377-14-58
[3]https://doi.org/10.1007/s11263-019-01228-7
[4]https://doi.org/10.1093/brain/aww173
