深度學習於自動分割腦膜瘤放射手術後之水腫區域;Deep Learning for Automated Segmentation of Brain Edema in Meningioma after Radiosurgery

NCUIR > College of Electrical Engineering & Computer Science > Graduate Institute of Electrical Engineering > Electronic Thesis & Dissertation > Item 987654321/89945

Please use this identifier to cite or link to this item: http://ir.lib.ncu.edu.tw/handle/987654321/89945

Title:	深度學習於自動分割腦膜瘤放射手術後之水腫區域;Deep Learning for Automated Segmentation of Brain Edema in Meningioma after Radiosurgery
Authors:	陳至宏;Chen, Zhi-Hong
Contributors:	電機工程學系
Keywords:	腦膜瘤;腦水腫;加馬刀放射治療;深度學習;腦部分割;數位影像處理;Meningioma;Brain edema;Gamma knife Radiosurgery;Deep Learning;Brain Segmentation;Digital Image Processing
Date:	2022-08-03
Issue Date:	2022-10-04 12:05:27 (UTC+8)
Publisher:	國立中央大學
Abstract:	腦膜瘤(meningioma)是最常見的良性腦部腫瘤之一，在接受加馬刀立體定位放射手術(Gamma knife radiosurgery, GKRS)治療後，可能會在病灶周邊產生一定程度的腦水腫(perifocal edema)。臨床治療上對於腦水腫體積的測定扮演著重要的角色。放射治療後病灶周邊水腫變化在腦部磁振T2權重影像(T2‐weighted)中可以清楚顯示，但是目前沒有客觀且準確的工具可以分割與量化T2權重影像中水腫的範圍與體積，所以不利於水腫消長過程、量化嚴重程度和病例差異性的研究。本研究藉由DeepMedic網路架構，和使用Mask R‐CNN模型取代手動式前處理的步驟，以遷移式學習(transfer learning)的概念訓練模型達到自動分割及量化T2權重影像中腦膜瘤GKRS後腦水腫區域。此量化工具將用以研究GKRS治療後所造成周邊組織的影響。本研究收集21位腦膜瘤患者齊接受過放射手術治療且定期追蹤的T2權重影像共154筆。經過篩選的130筆影像被隨機區分為訓練集94筆、驗證集22筆，以及測試集14筆。T2權重影像中的實際水腫範圍由臨床醫師手動標記作為監督式學習的依據。14筆測試資料經由訓練模型自動分割後，與醫師手動標記相比，達到平均相似係數(Dice similarity coefficient, DSC) 84.7%。本研究所建立的系統對於自動分割量化腦膜瘤患者接受放射手術治療後產生的腦水腫達到了相當優秀的效能，並且展露了預測模型建立的可能性。;Meningioma is one of the most common benign brain tumors, the radiation to the brain tissue surround the lesions may cause different degree of perifocal edema few months after Gamma knife radiosurgery (GKRS). Volumetric assessment of perifocal edema is highly relevant for therapy planning and monitoring. The post radiosurgery brain perifocal edematous change can be clearly identified in brain magnetic resonance T2‐weighted (T2‐w) images in which it appears as more hyper‐dense area compared with normal brain tissue. However, there is a lack of an objective tool to segment and quantify the volume of these T2‐w hyper‐dense area. So, it is not conducive to the research of the process of edema growth, quantification of severity and case differences. This study trained Mask R‐CNN to replace manual pre-processing steps to generate region interest, and trained DeepMedic architecture with the concept of transfer learning to automatically segment and quantify brain edema regions in images. This quantitative result will be used to explore the research of GKRS treatment on brain edema caused by meningioma in the future. Twenty‐one patients with meningiomas who had undergone GKRS treatment, and a total of 154 regularly tracked T2‐w scans were collected. The selected 130 scans are randomly distributed to a training set of 80 scans, a validation set of 80 scans, and a test set of 20 scans. The actual range of edema in the T2‐w images is labeled manually by the clinical radiologist as a gold standard for supervised learning. The 14 test scans were automatically segmented by the trained model and compared with the manual segmentation. The average Dice similarity coefficient is 84.7%. The automatic segmentation and quantification method this research proposed for brain edema after radiosurgery demonstrates excellent results. Possibilities for building predictive models are revealed.
Appears in Collections:	[Graduate Institute of Electrical Engineering] Electronic Thesis & Dissertation

Files in This Item:

File	Description	Size	Format
index.html		0Kb	HTML	54	View/Open

社群 sharing

Loading...