Contributions
Abstract: P1074
Type: Poster
Abstract Category: Pathology and pathogenesis of MS - 21 Imaging
Introduction: Accurate and reproducible automatic segmentation of white matter lesions in patients with multiple sclerosis (MS) remains important, but challenging. In the ISBI 2015 competition, a relatively simple multi-view convolutional neural network (MV-CNN) outperformed various state-of-the-art automatic lesion segmentation methods (Birenbaum 2016). Batch normalization is known to further improve the accuracy of CNNs (Loffe 2015). Here, we compare the performance of a batch normalized MV-CNN against the performance of human raters and the automatic k nearest-neighbors with tissue type priors (kNN-TTP) technique (Steenwijk 2013).
Methods: An MV-CNN was 'trained' using 3.0T 3D FLAIR and T1 signal intensities as features. Preprocessing involved brain extraction (FSL-BET) and rigid registration of the T1 to the FLAIR image (FSL-FLIRT, spline interpolation). Each branch of the MV-CNN (axial, sagittal and coronal) consisted of 4 stacked convolutional layers, which were merged into a fully connected layer. All layers were followed by batch normalization, ReLu activation and 25% dropout. Training time was reduced by only using the dilated lesion mask and 4% of the normal appearing tissue for training. No post-processing was applied. The method was trained and evaluated using a leave-one-out approach among previously published 20 MS patients (Steenwijk 2013). The reference segmentation was created completely manually by two raters. Volumetric and spatial agreement were evaluated using intra-class coefficient (ICC) and Dice Similarity Index (SI), respectively.
Results: As previously shown, both raters achieved good intra-observer (SI rater 1: 0.93; SI rater 2: 0.92) and inter-observer agreement (SI=0.84±0.04). The average lesion volume was 16.33±11.49mL and kNN-TTP reached reasonable performance (ICC=0.93; SI = 0.75±0.08) (Steenwijk 2013). MV-CNN showed better performance than the inter-observer agreement and kNN-TTP, reaching ICC=0.996 and average SI=0.87±0.05.
Conclusion: Our data suggests that MV-CNN with batch normalization outperforms the agreement between human raters, and provides a significant improvement in automatic MS lesion segmentation performance over a previously published method like kNN-TTP.
References: Birenbaum et al, International Workshop on Large-Scale Annotation of Biomedical Data and Expert Label Synthesis 2016, 58-67. Loffe et al, JMLR Workshop and Conference Proceedings 2015, 1-11. Steenwijk et al, NeuroImage: Clinical 2013, 3:462-69.
Disclosure:
M.D. Steenwijk has nothing to disclose.
M. Daams has nothing to disclose.
F. Barkhof serves as a consultant for Bayer-Schering Pharma, Sanofi-Aventis, Biogen, Teva, Novartis, Roche, Synthon BV, Genzyme and Jansen Research.
P.J.W. Pouwels receives research support from the Dutch MS Research Foundation, grant number 14-876.
J.J.G. Geurts is an editor of Multiple Sclerosis Journal, a member of the editorial boards of BMC Neurology, Neurology and Frontiers in Neurology, and serves as a consultant for Biogen and Genzyme.
Abstract: P1074
Type: Poster
Abstract Category: Pathology and pathogenesis of MS - 21 Imaging
Introduction: Accurate and reproducible automatic segmentation of white matter lesions in patients with multiple sclerosis (MS) remains important, but challenging. In the ISBI 2015 competition, a relatively simple multi-view convolutional neural network (MV-CNN) outperformed various state-of-the-art automatic lesion segmentation methods (Birenbaum 2016). Batch normalization is known to further improve the accuracy of CNNs (Loffe 2015). Here, we compare the performance of a batch normalized MV-CNN against the performance of human raters and the automatic k nearest-neighbors with tissue type priors (kNN-TTP) technique (Steenwijk 2013).
Methods: An MV-CNN was 'trained' using 3.0T 3D FLAIR and T1 signal intensities as features. Preprocessing involved brain extraction (FSL-BET) and rigid registration of the T1 to the FLAIR image (FSL-FLIRT, spline interpolation). Each branch of the MV-CNN (axial, sagittal and coronal) consisted of 4 stacked convolutional layers, which were merged into a fully connected layer. All layers were followed by batch normalization, ReLu activation and 25% dropout. Training time was reduced by only using the dilated lesion mask and 4% of the normal appearing tissue for training. No post-processing was applied. The method was trained and evaluated using a leave-one-out approach among previously published 20 MS patients (Steenwijk 2013). The reference segmentation was created completely manually by two raters. Volumetric and spatial agreement were evaluated using intra-class coefficient (ICC) and Dice Similarity Index (SI), respectively.
Results: As previously shown, both raters achieved good intra-observer (SI rater 1: 0.93; SI rater 2: 0.92) and inter-observer agreement (SI=0.84±0.04). The average lesion volume was 16.33±11.49mL and kNN-TTP reached reasonable performance (ICC=0.93; SI = 0.75±0.08) (Steenwijk 2013). MV-CNN showed better performance than the inter-observer agreement and kNN-TTP, reaching ICC=0.996 and average SI=0.87±0.05.
Conclusion: Our data suggests that MV-CNN with batch normalization outperforms the agreement between human raters, and provides a significant improvement in automatic MS lesion segmentation performance over a previously published method like kNN-TTP.
References: Birenbaum et al, International Workshop on Large-Scale Annotation of Biomedical Data and Expert Label Synthesis 2016, 58-67. Loffe et al, JMLR Workshop and Conference Proceedings 2015, 1-11. Steenwijk et al, NeuroImage: Clinical 2013, 3:462-69.
Disclosure:
M.D. Steenwijk has nothing to disclose.
M. Daams has nothing to disclose.
F. Barkhof serves as a consultant for Bayer-Schering Pharma, Sanofi-Aventis, Biogen, Teva, Novartis, Roche, Synthon BV, Genzyme and Jansen Research.
P.J.W. Pouwels receives research support from the Dutch MS Research Foundation, grant number 14-876.
J.J.G. Geurts is an editor of Multiple Sclerosis Journal, a member of the editorial boards of BMC Neurology, Neurology and Frontiers in Neurology, and serves as a consultant for Biogen and Genzyme.