Contributions
Abstract: EP1542
Type: ePoster
Abstract Category: Pathology and pathogenesis of MS - 21 Imaging
We present a method (DeepSCAN-MS) for automated volumetry of healthy appearing and white-matter lesion tissue from multispectral magnetic resonance imaging in multiple sclerosis. The method, which is an application of deep learning, is trained on a combination of automated tissue classification and manual lesion delineation, and provides a method for tracking the two of the NEDA-4 biomarkers for disease activity. These biomarkers cannot be adequately tracked manually, either because they are infeasible (atrophy) or time consuming and non-repeatable (lesion load)
FLAIR imaging of 123 MS patients from the Inselspital were examined by two raters, who delineated the FLAIR hyperintense lesions on each axial slice. The delineations from the first rater were superimposed on segmentations obtained from FSL FAST (2) (showing WM, GM and CSF) and Freesurfer (subcortical grey matter). The fused segmentations from 80 randomly selected cases, together with the FLAIR, T1-weighted and T2-weighted images were, after rigid registration, used to train an ensemble of deep learning classifiers (variations on the recently introduced Densenet architecture) The method was then applied to the remaining 43 cases. Segmentations of intracranial volume, white matter, grey matter and lesion tissue were signficantly correlated with ground truth volume (p< 0.001) in each case. A Bayesian method for providing confidence intervals for volumes allows to track significant volume change over time.
Disclosure:
Richard McKinley: nothing to disclose
Franca Wagner: nothing to disclose
Fabian Aschwanden: nothing to disclose
Mauricio Reyes: nothing to disclose
Anke Salman: has received speaker honoraria and/or travel compensation for activities with Almirall Hermal GmbH, Biogen, Merck, Novartis, Roche and Sanofi Genzyme, none related to this work.
Andrew Chan: has received personal compensation and research support from Almirall, Bayer, Biogen, Genzyme, Merck, Novartis, Roche, Teva and UCB in the last 3 years
Roland Wiest: nothing to disclose
Abstract: EP1542
Type: ePoster
Abstract Category: Pathology and pathogenesis of MS - 21 Imaging
We present a method (DeepSCAN-MS) for automated volumetry of healthy appearing and white-matter lesion tissue from multispectral magnetic resonance imaging in multiple sclerosis. The method, which is an application of deep learning, is trained on a combination of automated tissue classification and manual lesion delineation, and provides a method for tracking the two of the NEDA-4 biomarkers for disease activity. These biomarkers cannot be adequately tracked manually, either because they are infeasible (atrophy) or time consuming and non-repeatable (lesion load)
FLAIR imaging of 123 MS patients from the Inselspital were examined by two raters, who delineated the FLAIR hyperintense lesions on each axial slice. The delineations from the first rater were superimposed on segmentations obtained from FSL FAST (2) (showing WM, GM and CSF) and Freesurfer (subcortical grey matter). The fused segmentations from 80 randomly selected cases, together with the FLAIR, T1-weighted and T2-weighted images were, after rigid registration, used to train an ensemble of deep learning classifiers (variations on the recently introduced Densenet architecture) The method was then applied to the remaining 43 cases. Segmentations of intracranial volume, white matter, grey matter and lesion tissue were signficantly correlated with ground truth volume (p< 0.001) in each case. A Bayesian method for providing confidence intervals for volumes allows to track significant volume change over time.
Disclosure:
Richard McKinley: nothing to disclose
Franca Wagner: nothing to disclose
Fabian Aschwanden: nothing to disclose
Mauricio Reyes: nothing to disclose
Anke Salman: has received speaker honoraria and/or travel compensation for activities with Almirall Hermal GmbH, Biogen, Merck, Novartis, Roche and Sanofi Genzyme, none related to this work.
Andrew Chan: has received personal compensation and research support from Almirall, Bayer, Biogen, Genzyme, Merck, Novartis, Roche, Teva and UCB in the last 3 years
Roland Wiest: nothing to disclose