Contributions
Abstract: P1047
Type: Poster
Abstract Category: Pathology and pathogenesis of MS - 21 Imaging
Background: Quantitative MRI, like diffusion tensor, magnetization transfer and quantitative T1, can measure subtle and dynamic processes in MS including remyelination in lesions and pathological changes in normal appearing white matter (NAWM). This requires non-conventional imaging that is not generally implemented on clinical scanners. In conventional T1-weighted (T1w) scans, voxel intensities are arbitrary and vary from scan-to-scan. Normalization can ameliorate this, but techniques based on the whole-brain do not work well in the presence of pathology. Previously we suggested using orbital fat as a reference tissue unaffected by MS pathology, but manual orbital fat segmentation is challenging and time-consuming due to anatomical complexity (≈20 m/scan).
Deep learning (DL) is the machine learning technique that underlies most modern artificial intelligence. We trained our publicly available DL neuroimaging tool to perform high quality segmentation of orbital fat in individuals, and assessed the performance of this fully automatic technique for normalizing T1w scans.
Objectives: To train a DL system for automatic orbital fat segmentation and T1w normalization.
Methods: An expert MRI rater designed a detailed protocol for orbital fat segmentation and trained three additional raters. Each rater segmented T1w scans from (109 total) children with MS or monophasic demyelination. Scans were MPRAGE or FLASH at 1.5 T or 3 T, to provide a diverse training set. A DL network was trained on these data. All raters and DL segmented each of 15 validation scans and DICE overlap scores were calculated (1.0 is perfect agreement). The DL network also segmented 1239 scans from 262 children and the median intensity in the orbital fat was used to normalize each scan (nT1).
Results: DL agreed with the expert rater (DICE=0.84) as well as did the other raters (DICE=0.67-0.72) and required short processing time (0.5 s/scan). Standardized intra-subject NAWM variance was reduced after normalization in nT1 (0.0023) compared to T1w (0.050) by 20x. Sample size might be reduced by this much in some longitudinal studies.
Conclusions: We trained a DL algorithm to segment ocular fat from diverse MRI scans and used this for automatic T1w normalization. This technique performed as well as human raters and might allow use of conventional clinical T1w images for some quantitative and longitudinal studies, and for routine assessment of NAWM changes in MS.
Disclosure: R. A. Brown has received personal compensation for consulting services from Biogen Idec and NeuroRx Research.
R. Fratila has nothing to disclose.
D. Fetco has nothing to disclose.
S. Jiang has nothing to disclose.
G. Fadda has nothing to disclose.
N. Alkhawajah has nothing to disclose.
B. Banwell has served as a consultant for Novartis. Dr. Banwell serves as a non-remunerated advisor on clinical trial design for Biogen-IDEC, Sanofi, and Teva Neuroscience. Has participated as a speaker in meetings sponsored by and received consulting fees and/or grant support from Biogen Idec, Diogenix, Roche/Genentech, Sanofi-Genzyme, GlaxoSmithKline, Medimmune, Novartis, Ono Pharma, Teva Neuroscience, Celgene/Receptos Inc, and Merck/EMD Serono.
A. Bar-Or has participated as a speaker in meetings sponsored by and received consulting fees and/or grant support from Biogen Idec, Diogenix, Roche/Genentech, Sanofi-Genzyme, GlaxoSmithKline, Medimmune, Novartis, Ono Pharma, Teva Neuroscience, Celgene/Receptos Inc, and Merck/EMD Serono.
D. Arnold has served on advisory boards, received speaker honoraria, served as a consultant, or received research support from Adelphi, Biogen, Celgene, Genentech, Genzyme, Medday, NeuroRx Research, Novartis, Pfizer, Receptos, Roche, Sanofi, the Canadian Institutes of Health Research, and the Multiple Sclerosis Society of Canada; and holds stock in NeuroRx Research.
Abstract: P1047
Type: Poster
Abstract Category: Pathology and pathogenesis of MS - 21 Imaging
Background: Quantitative MRI, like diffusion tensor, magnetization transfer and quantitative T1, can measure subtle and dynamic processes in MS including remyelination in lesions and pathological changes in normal appearing white matter (NAWM). This requires non-conventional imaging that is not generally implemented on clinical scanners. In conventional T1-weighted (T1w) scans, voxel intensities are arbitrary and vary from scan-to-scan. Normalization can ameliorate this, but techniques based on the whole-brain do not work well in the presence of pathology. Previously we suggested using orbital fat as a reference tissue unaffected by MS pathology, but manual orbital fat segmentation is challenging and time-consuming due to anatomical complexity (≈20 m/scan).
Deep learning (DL) is the machine learning technique that underlies most modern artificial intelligence. We trained our publicly available DL neuroimaging tool to perform high quality segmentation of orbital fat in individuals, and assessed the performance of this fully automatic technique for normalizing T1w scans.
Objectives: To train a DL system for automatic orbital fat segmentation and T1w normalization.
Methods: An expert MRI rater designed a detailed protocol for orbital fat segmentation and trained three additional raters. Each rater segmented T1w scans from (109 total) children with MS or monophasic demyelination. Scans were MPRAGE or FLASH at 1.5 T or 3 T, to provide a diverse training set. A DL network was trained on these data. All raters and DL segmented each of 15 validation scans and DICE overlap scores were calculated (1.0 is perfect agreement). The DL network also segmented 1239 scans from 262 children and the median intensity in the orbital fat was used to normalize each scan (nT1).
Results: DL agreed with the expert rater (DICE=0.84) as well as did the other raters (DICE=0.67-0.72) and required short processing time (0.5 s/scan). Standardized intra-subject NAWM variance was reduced after normalization in nT1 (0.0023) compared to T1w (0.050) by 20x. Sample size might be reduced by this much in some longitudinal studies.
Conclusions: We trained a DL algorithm to segment ocular fat from diverse MRI scans and used this for automatic T1w normalization. This technique performed as well as human raters and might allow use of conventional clinical T1w images for some quantitative and longitudinal studies, and for routine assessment of NAWM changes in MS.
Disclosure: R. A. Brown has received personal compensation for consulting services from Biogen Idec and NeuroRx Research.
R. Fratila has nothing to disclose.
D. Fetco has nothing to disclose.
S. Jiang has nothing to disclose.
G. Fadda has nothing to disclose.
N. Alkhawajah has nothing to disclose.
B. Banwell has served as a consultant for Novartis. Dr. Banwell serves as a non-remunerated advisor on clinical trial design for Biogen-IDEC, Sanofi, and Teva Neuroscience. Has participated as a speaker in meetings sponsored by and received consulting fees and/or grant support from Biogen Idec, Diogenix, Roche/Genentech, Sanofi-Genzyme, GlaxoSmithKline, Medimmune, Novartis, Ono Pharma, Teva Neuroscience, Celgene/Receptos Inc, and Merck/EMD Serono.
A. Bar-Or has participated as a speaker in meetings sponsored by and received consulting fees and/or grant support from Biogen Idec, Diogenix, Roche/Genentech, Sanofi-Genzyme, GlaxoSmithKline, Medimmune, Novartis, Ono Pharma, Teva Neuroscience, Celgene/Receptos Inc, and Merck/EMD Serono.
D. Arnold has served on advisory boards, received speaker honoraria, served as a consultant, or received research support from Adelphi, Biogen, Celgene, Genentech, Genzyme, Medday, NeuroRx Research, Novartis, Pfizer, Receptos, Roche, Sanofi, the Canadian Institutes of Health Research, and the Multiple Sclerosis Society of Canada; and holds stock in NeuroRx Research.