3516
Task-aware 3D-Convolutional Neural Networks for Detailed Brain Parcellation1YIWEI Medical Technology Co., Ltd, Shenzhen, China, 2College of Biomedical Engineering & Instrument Science, Zhejiang University, Hangzhou, China
Synopsis
Morphological changes in neurodegenerative diseases can be detected with structural MR images, but it requires detailed brain parcellation. Therefore, a task-aware V-Net was proposed to segment the brain into 40 regions. Task-aware features were achieved by three cascading branches, including brain and non-brain regions, 25 regions with the bilateral regions grouped, and 40 regions, respectively. The proposed model was developed on 8938 subjects and validated using additional196 subjects. The proposed method outperformed the typical 3D U-Net and V-Net, and achieved state-of-the-art results on both datasets with a mean Dice score of 0.886 ± 0.029 and 0.874 ± 0.013, respectively.
INTRODUCTION
Neurodegenerative diseases typically lead to brain morphologic changes1 that can be assessed by volumetric measurement partly2. The sensitivity of such detection is closely associated with the region selection. Therefore, it requires detailed brain parcellation with high accuracy. Conventional segmentation based on atlas can provide detailed parcellation. However, the segmentation time is excessively long. It is not feasible for clinical feedback in real-time or large-scale studies. Deep learning-based methods have demonstrated promising results in automatic segmentation3, but most previous approaches focused on a small number of brain regions that are not detailed enough for clinical use. When a large number of regions are desired, the complexity of the deep learning model increases, and the label unbalance becomes worse, both of which make the segmentation task more challenging4. In this work, we developed a task-aware V-Net for brain segmentation using 8938 MRI scans and validated it using additional 196 MRI scans. The goal of this preliminary study is to segment the brain into 40 regions accurately and efficiently.METHODS
The first dataset includes a total of 8938 MRI 3D T1-weighted brain images, collected from public datasets including ABIDE5, ADHD6, ICBM7, OASIS8, and PPMI9. Although manual segmentation is commonly used as the ground truth, it is costly to get for such a large dataset. Therefore, pseudo ground truths were provided by Freesurfer10. A second dataset was included to validate the proposed method. It has 196 samples (40 brain structures per sample, mean age: 42.2±17.7 years), and the ground truth was manually segmented by experienced physicians. The proposed task-aware V-Net is the 3D V-Net11 with dense skip connections to fuse the low-level and high-level features. Pooling was performed three times as appropriate to preserve delicate features during the feature extraction. The original V-Net showed good performance with six brain regions but a low accuracy when segmenting the brain into 40 regions. Therefore, we introduced a cascaded structure11 to perform the segmentation with task-ware features. Specifically, three cascading branches were derived from the main branch at the output, Figure 1: the first branch differentiated brain and non-brain regions; in the second branch, bilateral brain regions were grouped, and the remaining 25 regions were segmented; the third branch segmented the brain into 40 regions. This approach is supposed to use task-aware features to improve the accuracy of the model. The dice loss was used in the first branch, and a combination of the median frequency balanced dice loss12, and the cross-entropy loss was used in the second and the third branch. The above training procedure can help alleviate the imbalance effects caused by the difference in volumes of different brain regions. Lastly, the loss of the three branches was added and constituted the total loss. The proposed model was trained on the first dataset with a training set of 5720 samples, a validation set of 1430 samples, and a testing set with 1788 samples. The trained model was further validated on the second dataset by comparing it to the manual segmentation. Original 3D V-Net and popular 3D U-Net were compared as benchmarks. Pytorch version 1.7.0 was used on a cluster of i7-7820X CPUs and two Nvidia GTX 1080Ti GPUs.RESULTS
Figure 2 shows the segmentation result of the proposed method, comparing to the original V-Net. The proposed method can effectively avoid the segmentation of non-brain regions, which was mislabeled by the original V-Net, as highlighted by the arrows in Figure 2. The proposed model achieved a mean Dice score of 0.886 ± 0.029 on the first datasets with the highest dice value of 0.948 ± 0.017 in the right cerebral white matter and the lowest value of 0.706 ± 0.077 in the left choroid plexus. In the validation with the second dataset, it provided mean Dice of 0.874 ± 0.013 with the highest dice value of 0.945 ± 0.013 in the brain stem and the lowest value of 0.687±0.067 in the right choroid plexus. The mean Dice score of our method is significantly higher than the widely used 3D U-Net13 (mean Dice of 0.875 ± 0.029 and 0.860 ± 0.014 on the public and the validation dataset, respectively) and 3D V-Net (mean Dice of 0.879 ± 0.029 and 0.865 ± 0.012 on the public and the validation dataset respectively), with p<0.0001 in Wilcoxon signed-rank test. In addition, compared with the excessive computation time when using Freesurfer (6 hours), our method predicted each case in less than 1 second based on GPU and in less than 20 seconds based on CPUs.DISCUSSION and CONCLUSION
We proposed an effective method to improve the performance of brain parcellation tasks. The task-aware segmentation strategy using the cascading branches provided a superior segmentation accuracy compared to the previous deep learning methods. The validation on the second dataset confirmed the accuracy of the proposed method with manual segmentation further. Comparing with the atlas-based Freesurfer method, our method provided an efficient solution for large scale studies and real-time clinical feedback. In future studies, we will continue to optimize the partitioning strategy to extract more discriminative features that can contribute to the segmentation.Acknowledgements
No acknowledgement found.References
1. J. L. Whitwell et al., ‘Rates of cerebral atrophy differ in different degenerative pathologies’, Brain, vol. 130, no. Pt 4, pp. 1148–1158, Apr. 2007, doi: 10.1093/brain/awm021.
2. K. Nithyakalyani, R. Kalpana, and R. Vigneswaran, ‘Volumetric assessment of human brain morphology using pixel counting technique’, in 2014 International Conference on Science Engineering and Management Research (ICSEMR), Nov. 2014, pp. 1–6, doi: 10.1109/ICSEMR.2014.7043556.
3. M. H. Hesamian, W. Jia, X. He, and P. Kennedy, ‘Deep Learning Techniques for Medical Image Segmentation: Achievements and Challenges’, J Digit Imaging, vol. 32, no. 4, pp. 582–596, 2019, doi: 10.1007/s10278-019-00227-x.
4. Z. Akkus, A. Galimzianova, A. Hoogi, D. L. Rubin, and B. J. Erickson, ‘Deep Learning for Brain MRI Segmentation: State of the Art and Future Directions’, J Digit Imaging, vol. 30, no. 4, pp. 449–459, Aug. 2017, doi: 10.1007/s10278-017-9983-4.
5. ‘ABIDE’. http://fcon_1000.projects.nitrc.org/indi/abide/ (accessed Dec. 10, 2020).
6. ‘ADHD200’. http://fcon_1000.projects.nitrc.org/indi/adhd200/ (accessed Dec. 10, 2020).
7. ‘LONI | ICBM Data Use Agreement’. https://ida.loni.usc.edu/collaboration/access/appLicense.jsp;jsessionid=828B7742C9F836E364DCBFE6E99D779F (accessed Dec. 10, 2020).
8. ‘OASIS Brains - Open Access Series of Imaging Studies’. https://www.oasis-brains.org/ (accessed Dec. 10, 2020).
9. ‘Parkinson’s Progression Markers Initiative |’. https://www.ppmi-info.org/ (accessed Dec. 10, 2020).
10. Bruce, Fischl, and, et al. Whole Brain Segmentation: Automated Labeling of Neuroanatomical Structures in the Human Brain[J]. Neuron, 2002.
11. F. Milletari, N. Navab, and S.-A. Ahmadi, ‘V-Net: Fully Convolutional Neural Networks for Volumetric Medical Image Segmentation’, arXiv:1606.04797 [cs], Jun. 2016, Accessed: Dec. 11, 2020. [Online]. Available: http://arxiv.org/abs/1606.04797.
12. A. G. Roy, S. Conjeti, D. Sheet, A. Katouzian, N. Navab, and C. Wachinger, ‘Error Corrective Boosting for Learning Fully Convolutional Networks with Limited Data’, arXiv:1705.00938 [cs], Jul. 2017, Accessed: Dec. 11, 2020. [Online]. Available: http://arxiv.org/abs/1705.00938.
13. Ö. Çiçek, A. Abdulkadir, S. S. Lienkamp, T. Brox, and O. Ronneberger, ‘3D U-Net: Learning Dense Volumetric Segmentation from Sparse Annotation’, arXiv:1606.06650 [cs], Jun. 2016, Accessed: Dec. 11, 2020. [Online]. Available: http://arxiv.org/abs/1606.06650.
Figures
