Displaying One Session

Parallel Session Sat, Sep 12, 2020
Session Type
Parallel Session
Date
Sat, Sep 12, 2020
Time (ET)
12:45 - 14:15
Invited Presentations Invited Abstracts

PS16.01 - Incorporating Machine Learning Approaches to Assess Putative Risk Factors for MS

Speakers
Authors
Presentation Number
PS16.01
Presentation Topic
Invited Presentations
Lecture Time
12:45 - 13:00

Abstract

Abstract

Multiple sclerosis (MS) susceptibility is multi-factorial with prominent genetic and non-genetic risk components, and there are complex interactions within and amongst these components that additively and synergistically contribute to MS risk. Efforts to characterize these risk components, and identify specific relationships underlying MS risk has significantly accelerated in the era of big-data. The challenge has been how best to analyze these rich and often-times unwieldy data, particularly when the number of predictors likely out-number the number of observations or where there are complex correlation patterns amongst predictors. Machine learning algorithms are well-suited for interrogating these complex big-data, as they rely on minimal assumptions. In general, a machine learning algorithm first parses the data, learns from it, and then assesses the prediction of what was learnt. We have successfully used machine learning to identify promising metabolomic candidates and complex genetic patterns contributing to MS risk. In both studies, Random forests (a supervised machine learning algorithm) was used to identify highly informative predictors for MS, and the relationships between these predictors and MS risk were formally tested using standard statistical models. Thus, we will present findings from two studies where machine learning was used as a means of data reduction which allowed for conservation of statistical power for association testing.

Collapse
Invited Presentations Invited Abstracts

PS16.02 - Machine Learning Techniques for Predicting MS Clinical Course

Presentation Number
PS16.02
Presentation Topic
Invited Presentations
Lecture Time
13:00 - 13:15

Abstract

Abstract

Machine learning (ML) based imaging diagnostics has recently reached or even surpassed the level of clinical experts in several clinical domains. In MS, ML techniques have been developed with two main objectives: 1) detection and segmentation of brain lesions and 2) classification of clinical forms and prediction of patient disability based on conventional or/and advanced MRI methods.

In this presentation, we review the recent developments in ML approaches including deep learning (DL) techniques based on conventional and/or advanced MRI biomarkers for clinical classification and monitoring as well as disability prediction.

Classical ML techniques such as Support Vector Machine (SVM), Random-Forest (RF), k-nearest neighbor (kNN), Linear Discriminant Analysis (LDA), initially used for lesion detection and segmentation, were compared with recent DL techniques. These methods include different types of neural networks such as Convolutional Neural Networks (CNN) and graph-based neural networks (GNN) that are designed for image and graph MRI data, respectively. In one hand, conventional MRI provides multispectral images such as pre- and post-gadolinium-contrast T1-weighted (T1w), T2-weighted (T2w), proton-density-weighted (PDw) or fluid-attenuated inversion recovery (FLAIR) for lesion detection and segmentation. On the other hand, advanced MRI techniques such tensor diffusion imaging (DTI) combined with graph models provide brain structural connectivity data allowing the extraction of global and local metrics for white matter (WM) characterization.

Recent developments of CNN based on conventional MRI images out performed ML techniques and provide excellent performances for lesion detection as well as for prediction of patient disability. Best results were obtained by measuring new and enlarging lesions (volume). Furthermore, GNN based on structural connectivity data out performed previous results for disability prediction. Finally, CNN and others ML methods are developed to improve the performances of classification or prediction, and further, to better visualize and interpret the CNN decision.

Deep learning frameworks based on 3D CNN or other advanced ML techniques provide to-day excellent performances for lesion segmentation and disability prediction on conventional MRI, and even better results, on WM fiber bundles and structural connectivity data. New CNN are improved by taking in account several parameters such data characteristics, measures uncertainty and by providing information on decision.

Collapse
Machine Learning/Network Science Oral Presentation

PS16.03 - Use of machine learning classifiers based on structural and functional visual metrics to predict diagnosis in children with acquired demyelination.

Speakers
Presentation Number
PS16.03
Presentation Topic
Machine Learning/Network Science
Lecture Time
13:15 - 13:27

Abstract

Background

Predicting diagnosis in youth at the first episode of demyelination is feasible in some but not all cases. Machine learning classifiers (MLC) can be trained to identify relationships between numerous multimodal input features and disease classifications to provide highly accurate predictions.

Objectives

To assess performance of machine learning classifiers for early disease diagnosis based on visual metrics in youth with demyelination.

Methods

Standardized clinical and visual data was prospectively collected at disease onset from 141 pediatric subjects with acquired demyelinating syndromes (ADS) and 75 healthy controls (HC). Participants were recruited through The Hospital for Sick Children (Toronto, Ontario (2010-2020)) and University of Calgary (2010-2017). Patients were classified using consensus definitions of demyelinating disorders and serum antibody testing for myelin oligodendrocyte glycoprotein (MOG) and aquaporin 4 (AQP4). Twenty-two auto-segmented Optical Coherence Tomography (OCT) features, 4 functional visual and 4 clinical features were used in a stratified manner alone or in combination to identify which combination of features provided the highest predictive accuracy. These input features were analyzed using 9 supervised MLC (Random Forest (RF), AdaBoost, XGBoost, Decision Tree (DT), Logistic Regression, Support Vector Machines (SVM), k-Nearest Neighbors, Stochastic Gradient Descent, Multinomial Naive Bayes). Data was split 80/20 between training and test sets. Backward feature selection was performed to re-run the analysis with best scoring predictor features in the MLC with highest predictive accuracy.

Results

AdaBoost, SVM, and DT were the best performing MLC with a test set accuracy between 82-88% in distinguishing between ADS and HC eyes. Multiple sclerosis (MS) was distinguished from HC with 92% accuracy. In descending order, fovea thickness, inferotemporal ganglion cell layer (GCL) thickness, low contrast visual acuity, outer inferior macular thickness, temporal peripapillary retinal nerve fiber layer and superior GCL thicknesses were the most important contributors to disease classification.

Conclusions

MLC can be used to combine visual metrics and clinical parameters to distinguish ADS from HC, and to predict MS. In addition to commonly used clinical metrics, we identified other structural and functional metrics that contribute importantly to classification. Among the machine learning algorithms tested, AdaBoost, SVM and DT performed best for this model.

Collapse
Machine Learning/Network Science Oral Presentation

PS16.04 - RimNet: A deep 3D multimodal MRI architecture for paramagnetic rim lesions assessment in multiple sclerosis

Speakers
Presentation Number
PS16.04
Presentation Topic
Machine Learning/Network Science
Lecture Time
13:27 - 13:39

Abstract

Background

In multiple sclerosis (MS), perilesional chronic inflammation appears on in vivo 3T susceptibility-based magnetic resonance imaging (MRI) as non-gadolinium-enhancing paramagnetic rim lesions (PRL). A higher PRL burden has been recently associated with a more aggressive disease course. The visual detection of PRL by experts is time-consuming and can be subjective.

Objectives

To develop a multimodal convolutional neural network (CNN) capable of automatically detecting PRL on 3D-T2*w-EPI unwrapped phase and 3D-T2w-FLAIR images.

Methods

124 MS cases (87 relapsing remitting MS, 16 primary progressive MS and 21 secondary progressive MS) underwent 3T MRI (MAGNETOM Prisma and MAGNETOM Skyra, Siemens Healthcare, Erlangen, Germany). Two neurologists visually inspected FLAIR magnitude and EPI phase images and annotated 462 PRL. 4857 lesions detected by an automatic segmentation (La Rosa et al. 2019) without overlap with PRL were considered non-PRL. The prototype RimNet was built upon two single CNNs, each fed with 3D patches centered on candidate lesions in phase and FLAIR images, respectively. A two-step feature-map fusion, initially after the first convolutional block and then before the fully connected layers, enhances the extraction of low and high-level multimodal features. For comparison, two unimodal CNNs were trained with phase and FLAIR images. The areas under the ROC curve (AUC) were used for evaluation (DeLong et al. 1988). The operating point was set at a lesion-wise specificity of 0.95. The patient-wise assessment was conducted by using a clinically relevant threshold of four rim+ lesions per patient (Absinta et al. 2019).

Results

RimNet (AUC=0.943) outperformed the phase and FLAIR image unimodal networks (AUC=0.913 and 0.855, respectively, P’s <0.0001). At the operating point, RimNet showed higher lesion-wise sensitivity (70.6%) than the unimodal phase network (62.1%), but lower than the experts (77.7%). At the patient level, RimNet performed with sensitivity of 86.8% and specificity of 90.7%. Individual expert ratings yielded averaged sensitivity and specificity values of 76.3% and 99.4%, respectively.

Conclusions

The excellent performance of RimNet supports its further development as an assessment tool to automatically detect PRL in MS. Interestingly, the unimodal FLAIR network performed reasonably well despite the absence of a paramagnetic rim, suggesting that morphometric features such as volume or shape might be a distinguishable feature of PRL.

Collapse
Neuromyelitis Optica and Anti-MOG Disease Oral Presentation

PS16.05 - Application of deep-learning to NMOSD and unclassified seronegative patients

Speakers
Presentation Number
PS16.05
Presentation Topic
Neuromyelitis Optica and Anti-MOG Disease
Lecture Time
13:39 - 13:51

Abstract

Background

Current diagnostic criteria of neuromyelitis optica spectrum disorders (NMOSD) allow the diagnosis of aquaporin-4 (AQP4) seropositive patients with limited manifestations, whereas seronegative patients with limited phenotypes remain unclassified and are usually considered as prodromal phases of multiple sclerosis (MS) or different entities themselves. Nowadays, there is great effort to perform an automatic diagnosis of different neurological diseases using deep-learning-based imaging diagnostics, which is a form of artificial intelligence, allowing predicting or making decisions without a priori human intervention.

Objectives

To provide a deep-learning classification of NMOSD patients with different serological profiles and to compare these results with their clinical evolution.

Methods

228 T2- and T1-weighted brain MRIs were acquired from patients with AQP4-seropositive NMOSD (n=85), early MS (n=95), AQP4-seronegative NMOSD (n=11, 3 with anti-myelin oligodendrocyte glycoprotein antibodies) and unclassified double-seronegative limited phenotypes (n=17 idiopathic recurrent optic neuritis [IRON], n=20 idiopathic recurrent myelitis [IRM]). The latter had a clinical re-evaluation after 4-year follow-up. The neural network architecture was based on four 3D convolutional layers. It was trained and validated on MRI scans (n=180) from AQP4-seropositive NMOSD and MS patients. Then, it was applied to AQP4-seronegative NMOSD and double-seronegative patients with limited phenotypes to evaluate their classification as NMOSD or MS in comparison with their clinical follow-up.

Results

The final algorithm discriminated between AQP-4-seropositive NMOSD and MS with an accuracy of 0.95. Forty-seven/48 (97.9%) seronegative patients were classified as NMOSD (one patient with IRON was classified as MS). Clinical follow-up was available in 27/37 (73%) double-seronegative limited phenotypes: one patient evolved to MS, three developed NMOSD and the others did not change phenotype.

Conclusions

Deep-learning may help in the diagnostic work-up of NMOSD. Our findings support the inclusion of AQP4-seronegative patients to the spectrum of NMO and suggest its enlargement to double-seronegative limited phenotypes.

Collapse