以卷積神經網路為基礎的老鼠腦部磁振影像中風區域分割之研究

Infarct Region Segmentation in Rat Brain MR Images after Stroke Based on Convolutional Neural Networks

郭曉芙(Hsiao-Fu Kuo)

指導教授 ： 張恆華

國立臺灣大學/工學院/工程科學及海洋工程學研究所/碩士(2018年)

https://doi.org/10.6342/NTU201803479

摘要

腦中風是世界致死率第二高的疾病，主要可以分為缺血型與出血型兩大類，在臨床上有許多研究發展空間。而臨床前實驗模型中，多半使用齧齒動物輔以磁振影像作為實驗研究依據。研究者在分析作業前需要經過許多的處理步驟，例如將大腦區域與缺血型中風區域提取出來，然而這些區域不僅需要專家耗時且費力的手動分割外，也亦產生標準不一的問題。因此，本篇論文的主要目的是希望能藉由全卷積神經網路來達成全自動化的分割預測。為了使神經網路能夠正確分割出中風區域，我們將分割步驟分為兩個階段：大腦分割及缺血型中風區域分割，兩者皆使用相同的卷積神經網路分割系統。透過編譯-反編譯的網路架構與結合不同階層特徵的方式，我們能夠較準確地找到需要的特徵並以逐像素(pixel-wise)的方式判別。最後，再利用一些簡單的形態學處理優化分割出的影像。本篇研究中使用了35筆老鼠中風資料，實驗結果顯示本研究方法可以相當準確地截取老鼠大腦(98.12%)，對於中風區域亦有優良的分割結果(80.47%)。

關鍵字

磁振影像 ； 缺血型中風 ； 影像分割 ； 卷積神經網路 ； 編譯-反編譯

並列摘要

Stroke has the second highest fatality rate around the world. It can be divided into two major categories: ischemic and hemorrhagic. In the clinically, there is lots of development prospect. Rodents associated with magnetic resonance (MR) images are often preclinical experimental models. Researchers need to go through many processing steps before analyzing the operation, such as extracting the brain regions and infarct regions. However, these regions need lots of time for manual segmentation by experts, who may have the problems of different standards. Therefore, the main purpose of thesis is to achieve fully automated segmentation by using fully convolutional neural networks. In order to segment the right region, we divide the segmentation step into two phases: brain segmentation and infarct segmentation, both of which using the same convolutional neural network. With an encoder-decoder network and concatenating different levels of features, we are able to find the accurate features that we need and classify them pixel by pixel. Finally, some simple morphological methods are applied to optimize the segmentation results. In this study, 35 rat brain MR image data were used. The experimental results show that the proposed method accurately extracted the rat brain (98.12%) and provided good segmentation results for the infarct region (80.47%).

並列關鍵字

magnetic resonance image(MRI) ； ischemic stroke ； image segmentation ； convolutional neural network ； encoder-decoder

參考文獻

[1] W. H. Organization, "The top 10 causes of death," 2018.

[2] R. J. Mural, M. D. Adams, E. W. Myers, H. O. Smith, G. L. G. Miklos, R. Wides, et al., "A Comparison of Whole-Genome Shotgun-Derived Mouse Chromosome 16 and the Human Genome," Science, vol. 296, p. 1661, 2002.

[3] B. A. Moffat, C. J. Galbán, and A. Rehemtulla, "Advanced MRI: Translation from Animal to Human in Brain Tumor Research," Neuroimaging Clinics of North America, vol. 19, pp. 517-526, 2009/11/01/ 2009.

[4] N. D. E. Greene, M. F. Lythgoe, D. L. Thomas, R. L. Nussbaum, D. J. Bernard, and H. M. Mitchison, "High resolution MRI reveals global changes in brains of Cln3 mutant mice," European Journal of Paediatric Neurology, vol. 5, pp. 103-107, 2001/01/01/ 2001.

[5] O. Natt, T. Watanabe, S. Boretius, J. Radulovic, J. Frahm, and T. Michaelis, "High-resolution 3D MRI of mouse brain reveals small cerebral structures in vivo," Journal of Neuroscience Methods, vol. 120, pp. 203-209, 2002/10/30/ 2002.