Contributions
Abstract: P467
Type: Poster Sessions
Abstract Category: Pathology and pathogenesis of MS - MRI and PET
Introduction: Atrophy is a hallmark of neurodegeneration in Multiple Sclerosis (MS) that can be quantified by MRI. Brainstem (BS) atrophy is under-investigated in MS.
Objective and aims: To assess accuracy and reproducibility of a fully-automated deep learning-based segmentation method for BS volumetry in 3D high-resolution T1w MRI data of healthy controls (HC) and MS patients.
Methods: Segmentation was done using multi-dimensional gated recurrent units (MD-GRU; Andermatt et al., 2016 (DOI 10.1007/978-3-319-46976-8_15), Andermatt et al., 2017 (DOI 10.1007/978-3-319-75238-9_3)) a deep learning-based, fully-automated semantic segmentation approach employing a convolutional adaptation of gated recurrent units (GRU; Cho et al., 2014 (http://arxiv.org/abs/1409.1259)). In brief, MD-GRU traverses an image forward and backward along each of its spatial dimensions to infer the current segmentation class label from the local appearance and its surrounding context. The respective neural network was trained for 100'000 iterations on 67 scans (17 HC, 50 patients). Mean Dice score wrt. an expert-labeled manual ground truth was used to select the final training state for evaluation:the state producing the highest score on the 3 labeled sub-regions of the BS (midbrain (M), pons (P) and medulla oblongata (MO)) in a separate set of 20 patients' scans was chosen for further analyses.
Expert-labeled manual BS segmentations were then used to validate the accuracy of the automated segmentation in another independent set of 20 patients' scans using Dice scores. The reproducibility of the segmentations was assessed in 11 HC that underwent a MR test-retest experiment with repositioning in-between. The mean %-change betw. test and retest and the respective intra-class correlation coefficients (ICC) were calculated.
Results: Accuracy: In the validation set, the mean Dice scores comparing automated to the manual segmentations were (mean/SD):0.97/0.006 (total BS); 0.95/0.015 (M);0.97/0.008 (P); 0.96/0.014 (MO). Reproducibility: The mean %-change/SD between test-retest scans was 0.47%/0.004 for the automated and 0.82%/0.005 for the manual segmentation of the total BS. The ICC of the automated test-retest segmentations of the total BS, M and P were all >0.99, of MO 0.97.
Conclusions: This fully-automated BS segmentation provides accurate, reproducible segmentations in HC and MS patients in 200sec/scan on a Nvidia GeForce GTX 1080 GPU and has potential for use in longitudinal studies.
Disclosure: Sander L, Andermatt S, Pezold S, Amann M, Meier D, Wendebourg MJ: nothing to disclose. Sinnecker T has received travel support from Actelion and Roche, and speaker fees from Biogen. Naegelin Y, Granziera C, Kappos L: nothing to disclose. Wuerfel J: CEO of MIAC AG Basel, Switzerland. He served on scientific advisory boards of Actelion, Biogen, Genzyme-Sanofi, Novartis, and Roche. He is or was supported by grants of the EU (Horizon2020), German Federal Ministeries of Education and Research (BMBF) and of Economic Affairs and Energy (BMWI). Cattin P, Schlaeger R: nothing to disclose.
Abstract: P467
Type: Poster Sessions
Abstract Category: Pathology and pathogenesis of MS - MRI and PET
Introduction: Atrophy is a hallmark of neurodegeneration in Multiple Sclerosis (MS) that can be quantified by MRI. Brainstem (BS) atrophy is under-investigated in MS.
Objective and aims: To assess accuracy and reproducibility of a fully-automated deep learning-based segmentation method for BS volumetry in 3D high-resolution T1w MRI data of healthy controls (HC) and MS patients.
Methods: Segmentation was done using multi-dimensional gated recurrent units (MD-GRU; Andermatt et al., 2016 (DOI 10.1007/978-3-319-46976-8_15), Andermatt et al., 2017 (DOI 10.1007/978-3-319-75238-9_3)) a deep learning-based, fully-automated semantic segmentation approach employing a convolutional adaptation of gated recurrent units (GRU; Cho et al., 2014 (http://arxiv.org/abs/1409.1259)). In brief, MD-GRU traverses an image forward and backward along each of its spatial dimensions to infer the current segmentation class label from the local appearance and its surrounding context. The respective neural network was trained for 100'000 iterations on 67 scans (17 HC, 50 patients). Mean Dice score wrt. an expert-labeled manual ground truth was used to select the final training state for evaluation:the state producing the highest score on the 3 labeled sub-regions of the BS (midbrain (M), pons (P) and medulla oblongata (MO)) in a separate set of 20 patients' scans was chosen for further analyses.
Expert-labeled manual BS segmentations were then used to validate the accuracy of the automated segmentation in another independent set of 20 patients' scans using Dice scores. The reproducibility of the segmentations was assessed in 11 HC that underwent a MR test-retest experiment with repositioning in-between. The mean %-change betw. test and retest and the respective intra-class correlation coefficients (ICC) were calculated.
Results: Accuracy: In the validation set, the mean Dice scores comparing automated to the manual segmentations were (mean/SD):0.97/0.006 (total BS); 0.95/0.015 (M);0.97/0.008 (P); 0.96/0.014 (MO). Reproducibility: The mean %-change/SD between test-retest scans was 0.47%/0.004 for the automated and 0.82%/0.005 for the manual segmentation of the total BS. The ICC of the automated test-retest segmentations of the total BS, M and P were all >0.99, of MO 0.97.
Conclusions: This fully-automated BS segmentation provides accurate, reproducible segmentations in HC and MS patients in 200sec/scan on a Nvidia GeForce GTX 1080 GPU and has potential for use in longitudinal studies.
Disclosure: Sander L, Andermatt S, Pezold S, Amann M, Meier D, Wendebourg MJ: nothing to disclose. Sinnecker T has received travel support from Actelion and Roche, and speaker fees from Biogen. Naegelin Y, Granziera C, Kappos L: nothing to disclose. Wuerfel J: CEO of MIAC AG Basel, Switzerland. He served on scientific advisory boards of Actelion, Biogen, Genzyme-Sanofi, Novartis, and Roche. He is or was supported by grants of the EU (Horizon2020), German Federal Ministeries of Education and Research (BMBF) and of Economic Affairs and Energy (BMWI). Cattin P, Schlaeger R: nothing to disclose.