Deep learning of baseline brain lesion masks better predicts SPMS progression than baseline clinical, demographic and MRI measures
ECTRIMS Online Library. Law M. 09/13/19; 278656; P1616
Marco Law
Marco Law
Contributions
Abstract

Abstract: P1616

Type: Poster Sessions

Abstract Category: Poster Session 3

M. Law1, A. Traboulsee2, D. Li3, R. Carruthers2, M. Freedman4, S. Kolind3, R. Tam3

1School of Biomedical Engineering, 2Department of Neurology, 3Department of Radiology, University of British Columbia, Vancouver, BC, 4Neurology, University of Ottawa and The Ottawa Hospital Research Institute, Ottawa, ON, Canada

Introduction: Predicting disability progression in multiple sclerosis (MS) has largely been limited using traditional approaches. Deep learning (DL) is a class of artificial intelligence capable of extracting features from multi-dimensional data and has been shown to learn features from binary brain lesion masks that correlate better with MS disability scores than total lesion volume.
Objective: To compare deep-learned baseline brain lesion mask features against user-defined baseline features for predicting disability progression in secondary progressive MS (SPMS).
Aim: To develop deep learning models for the accurate prognostication of MS.
Methods: Participants from a 2-year placebo-controlled (negative) SPMS trial assessing MBP8298 were categorized as progressors if there was 6-month sustained increase in EDSS (≥1.0 and ≥0.5 for baseline ≤5.5 and ≥6.0 respectively). A DL network was used to automatically extract features from Euclidean-distance transformed brain lesion masks. This was compared to baseline Expanded Disability Status Scale score, MS functional composite (timed 25-foot walk, 9-hole peg test, paced auditory serial addition), disease duration, age, brain volume and T2 lesion volume. Logistic regression (LR) was used to predict disability progression. Random under-sampling was used to correct for class-imbalance during model fitting. Training and testing were done using 10-fold cross validation. Area under the receiver-operator characteristic curve (AUC), sensitivity, precision, sensitivity and negative predictive value (NPV) were used to compare the two sets of features using paired t-tests on each fold.
Results: There were 115 progressors and 370 non-progressors. The deep-learned features were better than user-defined features in AUC (55.0% vs. 45.0%), precision (27.0% vs 22.0%) and NPV (79.1% vs. 74.9%) at p < 0.05. Combining both sets of features did not alter performance. There were no differences in sensitivity and specificity between deep learned features and user-defined features.
Conclusion: Using only lesion masks, DL was able to learn features more predictive of disability progression in SPMS than user-defined features. Observed improvements in precision and NPV of deep-learned features compared to user-defined features may be due to the DLN learning use of spatial information that is lost in global BPF and T2LV measures.
Disclosure: Marco T. K. Law has nothing to disclose. Anthony Traboulsee has the following competing financial interests: research funding from Biogen, Chugai, Novartis, Roche, Sanofi Genzyme, and consultancy honoraria from Biogen, Roche, Sanofi Genzyme, Teva Neuroscience. David K. B. Li has received research funding from the Canadian Institute of Health Research and Multiple Sclerosis Society of Canada. He is the Emeritus Director of the UBC MS/MRI Research Group which has been contracted to perform central analysis of MRI scans for therapeutic trials with Novartis, Perceptives, Roche and Sanofi-Aventis. The UBC MS/MRI Research Group has also received grant support for investigator-initiated independent studies from Genzyme, Merck-Serono, Novartis and Roche. He has acted as a consultant to Vertex Pharmaceuticals and served on the Data and Safety Advisory Board for Opexa Therapeutics and Scientific Advisory Boards for Adelphi Group, Celgene, Novartis and Roche. He has also given lectures which have been supported by non- restricted education grants from Biogen-Idec, Novartis, Sanofi-Genzyme and Teva. Robert Carruthers is Site Investigator for studies funded by Roche, Novartis, MedImmune, EMD Serono and receives research support from Teva Innovation Canada, Roche Canada and Vancouver Coastal Health Research Institute. Robert Carruthers has done consulting work and has received honoraria from Roche, EMD Serono, Sanofi, Biogen, Novartis, and Teva. Mark S. Freedman has received a research/educational grant from Genzyme; received honoraria or consultation fees from Actelion, BayerHealthcare, BiogenIdec, Chugai, Clene Nanomedicine, EMD Canada, Genzyme, Merck Serono, Novartis, Hoffman La-Roche, Sanofi-Aventis, Teva Canada Innovation; is member of a company advisory board, board of directors or other similar group of Actelion, BayerHealthcare, BiogenIdec, Hoffman La-Roche, Merck Serono, MedDay, Novartis, Sanofi- Aventis and is on speaker ́s bureau for Genzyme. Shannon Kolind has received a research/educational grant funding from Genzyme and Roche. Roger Tam has received research support as part of sponsored clinical studies from Novartis, Roche, and Sanofi Genzyme.

Abstract: P1616

Type: Poster Sessions

Abstract Category: Poster Session 3

M. Law1, A. Traboulsee2, D. Li3, R. Carruthers2, M. Freedman4, S. Kolind3, R. Tam3

1School of Biomedical Engineering, 2Department of Neurology, 3Department of Radiology, University of British Columbia, Vancouver, BC, 4Neurology, University of Ottawa and The Ottawa Hospital Research Institute, Ottawa, ON, Canada

Introduction: Predicting disability progression in multiple sclerosis (MS) has largely been limited using traditional approaches. Deep learning (DL) is a class of artificial intelligence capable of extracting features from multi-dimensional data and has been shown to learn features from binary brain lesion masks that correlate better with MS disability scores than total lesion volume.
Objective: To compare deep-learned baseline brain lesion mask features against user-defined baseline features for predicting disability progression in secondary progressive MS (SPMS).
Aim: To develop deep learning models for the accurate prognostication of MS.
Methods: Participants from a 2-year placebo-controlled (negative) SPMS trial assessing MBP8298 were categorized as progressors if there was 6-month sustained increase in EDSS (≥1.0 and ≥0.5 for baseline ≤5.5 and ≥6.0 respectively). A DL network was used to automatically extract features from Euclidean-distance transformed brain lesion masks. This was compared to baseline Expanded Disability Status Scale score, MS functional composite (timed 25-foot walk, 9-hole peg test, paced auditory serial addition), disease duration, age, brain volume and T2 lesion volume. Logistic regression (LR) was used to predict disability progression. Random under-sampling was used to correct for class-imbalance during model fitting. Training and testing were done using 10-fold cross validation. Area under the receiver-operator characteristic curve (AUC), sensitivity, precision, sensitivity and negative predictive value (NPV) were used to compare the two sets of features using paired t-tests on each fold.
Results: There were 115 progressors and 370 non-progressors. The deep-learned features were better than user-defined features in AUC (55.0% vs. 45.0%), precision (27.0% vs 22.0%) and NPV (79.1% vs. 74.9%) at p < 0.05. Combining both sets of features did not alter performance. There were no differences in sensitivity and specificity between deep learned features and user-defined features.
Conclusion: Using only lesion masks, DL was able to learn features more predictive of disability progression in SPMS than user-defined features. Observed improvements in precision and NPV of deep-learned features compared to user-defined features may be due to the DLN learning use of spatial information that is lost in global BPF and T2LV measures.
Disclosure: Marco T. K. Law has nothing to disclose. Anthony Traboulsee has the following competing financial interests: research funding from Biogen, Chugai, Novartis, Roche, Sanofi Genzyme, and consultancy honoraria from Biogen, Roche, Sanofi Genzyme, Teva Neuroscience. David K. B. Li has received research funding from the Canadian Institute of Health Research and Multiple Sclerosis Society of Canada. He is the Emeritus Director of the UBC MS/MRI Research Group which has been contracted to perform central analysis of MRI scans for therapeutic trials with Novartis, Perceptives, Roche and Sanofi-Aventis. The UBC MS/MRI Research Group has also received grant support for investigator-initiated independent studies from Genzyme, Merck-Serono, Novartis and Roche. He has acted as a consultant to Vertex Pharmaceuticals and served on the Data and Safety Advisory Board for Opexa Therapeutics and Scientific Advisory Boards for Adelphi Group, Celgene, Novartis and Roche. He has also given lectures which have been supported by non- restricted education grants from Biogen-Idec, Novartis, Sanofi-Genzyme and Teva. Robert Carruthers is Site Investigator for studies funded by Roche, Novartis, MedImmune, EMD Serono and receives research support from Teva Innovation Canada, Roche Canada and Vancouver Coastal Health Research Institute. Robert Carruthers has done consulting work and has received honoraria from Roche, EMD Serono, Sanofi, Biogen, Novartis, and Teva. Mark S. Freedman has received a research/educational grant from Genzyme; received honoraria or consultation fees from Actelion, BayerHealthcare, BiogenIdec, Chugai, Clene Nanomedicine, EMD Canada, Genzyme, Merck Serono, Novartis, Hoffman La-Roche, Sanofi-Aventis, Teva Canada Innovation; is member of a company advisory board, board of directors or other similar group of Actelion, BayerHealthcare, BiogenIdec, Hoffman La-Roche, Merck Serono, MedDay, Novartis, Sanofi- Aventis and is on speaker ́s bureau for Genzyme. Shannon Kolind has received a research/educational grant funding from Genzyme and Roche. Roger Tam has received research support as part of sponsored clinical studies from Novartis, Roche, and Sanofi Genzyme.

By clicking “Accept Terms & all Cookies” or by continuing to browse, you agree to the storing of third-party cookies on your device to enhance your user experience and agree to the user terms and conditions of this learning management system (LMS).

Cookie Settings
Accept Terms & all Cookies