Deep Learning disconnectomes to accelerate & improve long-term predictions for post-stroke symptoms

2182 

Abstract Submission 

Anna Matsulevits¹, Pierrick Coupé², Huy-Dung Nguyen², Lia Talozzi³, Chris Foulon⁴, Parashkev Nachev⁵, Maurizio Corbetta⁶, Thomas Tourdias⁷, Michel Thiebaut de Schotten⁸

¹University Bordeaux, Institut des Maladies Neurodégénératives CNRS UMR 5293 Université de Bordeaux, Bordeaux, Gironde, ²University Bordeaux, Bordeaux, Gironde, ³Stanford Medical School, Stanford, CA, ⁴Groupe d'Imagerie Neurofonctionnelle (GIN), Institut des Maladies Neurodegeneratives-UMR 5293, CNRS, Bordeaux, FRANCE!, ⁵UCL Queen Square Institute of Neurology, London, London, ⁶Clinica Neurologica, Department of Neuroscience, University of Padova, Padova, Italy, Padova, Padova, ⁷University Bordeaux, Bordeaux , Gironde, ⁸Groupe d’Imagerie Neurofonctionnelle, Institut des Maladies Neurodégénératives- UMR 5293, CNRS, CEA, Bordeaux, France

Anna Matsulevits  
University Bordeaux, Institut des Maladies Neurodégénératives CNRS UMR 5293 Université de Bordeaux
Bordeaux, Gironde

Pierrick Coupé  
University Bordeaux
Bordeaux, Gironde

Huy-Dung Nguyen  
University Bordeaux
Bordeaux, Gironde

Lia Talozzi  
Stanford Medical School
Stanford, CA

Chris Foulon  
Groupe d'Imagerie Neurofonctionnelle (GIN), Institut des Maladies Neurodegeneratives-UMR 5293, CNRS
Bordeaux, FRANCE!

Parashkev Nachev  
UCL Queen Square Institute of Neurology
London, London

Maurizio Corbetta  
Clinica Neurologica, Department of Neuroscience, University of Padova, Padova, Italy
Padova, Padova

Thomas Tourdias  
University Bordeaux
Bordeaux , Gironde

Michel Thiebaut de Schotten  
Groupe d’Imagerie Neurofonctionnelle, Institut des Maladies Neurodégénératives- UMR 5293, CNRS, CEA
Bordeaux, France

White matter connections are recognized as fundamental building blocks of behavior and cognition, and their disconnections can be quantified to facilitate personalized prediction. This is particularly relevant in the context of stroke that is going to damage a specific brain region but also disconnect several remote areas. Being able to anticipate the risk of developing motor, cognitive, and emotional impairments following stroke could help to refer the patients to dedicated training to improve their outcomes. With the rise of Artificial Intelligence applications in healthcare, we explored and evaluated the potential of deep-learning models to accurately generate disconnectomes in a population of stroke survivors in order to speed up and accelerate the individualized prediction of neuropsychological scores one year post-stroke.

We implemented a 3D U-Net network for predicting individual deep-disconnectomes from binary masks of infarcts that was trained on N=1333 synthetic lesions and their corresponding disconnectomes, and tested on N=1333 real stroke lesions. The level of similarities between deep-disconnectome and conventional disconnectomes was assessed by the percentage of the variance of disconnection reproduced by the 3D U-Net. To explore any systematic differences in terms of disconnected voxels, we contrasted frequency maps of disconnected voxels. To predict clinical scores, we embedded the deep-learning-based disconnection pattern of each of the 1333 patients within a 2D morphospace using UMAP dimensionality reduction. With the achieved association between location within the morphospace and neuropsychological scores, we were able to predict scores for an out-of-sample population. We tested this on 139 new stroke patients using multiple regression and validated out-of-sample on 20 patients. Finally, we compared the accuracy of prediction between the two methods.

The trained 3D U-Net algorithm was able to capture most information obtained in conventional disconnectomes, i.e., statistical maps filtering normative white-matter networks, but outputed a deep-disconnectome 720 times faster – compared to disconnectome computation with the state-of-the-art software. Moreover, through the morphospace, the deep-disconnectomes predicted neuropsychological outcome at 1 year with an average accuracy of 85.2% (R²=0.208) which was significantly better (p=0.009) than prediction from the conventional disconnectome approach and confirmed out-of-sample. In order to understand why the score prediction of the deep-disconnectome outperforms the conventional disconnectome, a systematic comparison across the embedding coordinates was performed for each patient. To find explanations for this, we assessed the structure of the morphospace by means of the average Euclidean distance between each embedding set of coordinates against the rest of the embedding points. This calculation suggests a greater differentiation for deep-disconnectomes derived from similar stroke lesions. This may have improved the segregation between similar profiles of white-matter damage and, accordingly, could have led to better modeling of fine differences within the same neuropsychological assessment.

→ Our 3D U-Net algorithm is able to accurately resemble the ground truth and produce an accurate deep-disconnectome from a binary lesion mask that captures statistical maps filtering normative white-matter networks, just as the conventional disconnectome, but 720 times faster.
→ For long-term stroke outcome predictions, the deep-disconnectome's predictive power outperformed the conventional disconnectome predictions for neuropsychological scores.
→ This work demonstrates the potential of practical application of AI-driven models for clinical settings related to stroke outcome predictions and stroke management, which might enhance efficiency within healthcare systems, and ultimately contribute to an improved quality of life for stroke survivors.

Neurodegenerative/ Late Life (eg. Parkinson’s, Alzheimer’s) ²

Classification and Predictive Modeling

Methods Development

White Matter Anatomy, Fiber Pathways and Connectivity ¹

Workflows

Machine Learning

Tractography

Other - Deep-Learning, Stroke, Precision Medicine, Neuroimaging