請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/18288
標題: | Mirau全域式光學同調斷層掃描技術應用於活體大鼠角膜分析 Analysis of In-vivo Rat Corneal Using Mirau-based Full-field Optical Coherence Tomography |
作者: | Cheng-Tsung Tsai 蔡政錝 |
指導教授: | 黃升龍(Sheng-Lung Huang) |
關鍵字: | Mirau全域式光學同調斷層掃描術,角膜結構,活體大鼠角膜測量,機器學習,深度學習,支援向量機,U-Net,角膜內皮細胞,細胞分割, Mirau-bsaed full-field optical coherence tomography,corneal structure,In vivo rat corneal measure,machine learning,deep learning,SVM,U-Net,corneal endothelial cells,semantic segmentation, |
出版年 : | 2020 |
學位: | 碩士 |
摘要: | 由於3C產品的普及,用眼過度成為現代人常遇見的問題,臨床醫學中常被用來診斷的儀器為光學同調斷層掃描(Optical coherence tomography; OCT),協助醫師對疾病的確診及手術後復原追蹤,並提供有助於判斷病理狀況、結構、厚度、深度以及細胞形貌等數據,最後達成數據化的醫療。
本論文中 OCT系統為Mirau 全域式光學同調斷層掃描系統,光源為實驗室自行生長的摻鈰釔鋁石榴石(Ce3+:YAG)晶體光纖的自發輻射(Spontaneous emission),能產生中心波長560 nm,頻寬95 nm的光源。系統具備縱向0.91μm,橫向0.84 μm 的高解析度,用於掃描活體大鼠樣本時,能夠清楚解析角膜各個重要的分層與細胞結構,整體大鼠角膜厚度大約為175.1μm,分層的狀況為上皮細胞層39.1μm、前彈力層至基質層13.3 ±0.8μm、後彈力層8.2 ±0.4μm、內皮細胞層5.1 ±0.3μm。此外,本論文利用U-Net深度學習來訓練大鼠角膜的內皮細胞,促使未來輸入的影像能夠直接得到分割後的結果,降低人為手劃或圈選所增加的時間與錯誤,將內皮細胞標籤化後,用U-Net的模組來訓練約莫20張角膜內皮細胞數據,並且得到不會過度擬合的學習速率,以及影像對遮罩準確度為93.4%的正相關權重值,最後預測出重建後的細胞分割影像,計算量化的大鼠內皮細胞之細胞密度,大約2316 ±294 mm-2。 本論文展現OCT系統應用於活體眼睛量測之潛力,提供影像及數據可協助學術研究與臨床實驗,並撰寫U-Net深度學習來做影像分割及應用,未來希望進一步將量化後的成果應用於人體眼睛疾病診斷。 Owing to the 3C products are popular, excessive eye strain in today's modern society has become a common problem. Optical coherence tomography (OCT) is used to diagnostic instrument in clinical medicine. It is important to support doctor by confirmed disease and recovering from surgery. We can get more data to prove digital medicine, i.e. pathological condition, structure, thickness, depth, and cell morphology. In this thesis, a homemade Ce3+:YAG single-cladding crystal fiber, which generate amplified spontaneous emission centered at 560 nm with bandwidth of 95 nm. And we use it as the light source to demonstrate a full field OCT with axial resolution of 0.91μm and lateral resolution of 0.84μm. We use full-field OCT to scan In vivo rat cornea. We can get clearly analyze the important layers and cell structure of the cornea, i.e. rat corneal thickness has 175.1μm, epithelial layer 39.1μm, Bowman’s layer 13.3 ±0.8μm, Descemet’s layer 8.2 ±0.4μm, endothelial layer5.1 ±0.3μm. Thus, We training rat corneal endothelial cell by U-Net deep learning. In the future, We should want to get segmentation result. To reduce human do much more mistake. We try to train rough twenty slides endothelial data and label it. Finally We can predict image segmentation by seemly fitting and get accuracy 93.4% weight. We calculated the endothelial cell density of rats, which was approximately 2316 ±294 mm-2 . In this thesis, We demonstrate OCT image by measure In vivo cornea. Supporting research and clinical trials get more image data. And coding U-Net image segmentation. Finally, We want to get more data by human eye disease. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/18288 |
DOI: | 10.6342/NTU202003183 |
全文授權: | 未授權 |
顯示於系所單位: | 光電工程學研究所 |
文件中的檔案:
檔案 | 大小 | 格式 | |
---|---|---|---|
U0001-1308202002204400.pdf 目前未授權公開取用 | 5.28 MB | Adobe PDF |
系統中的文件,除了特別指名其著作權條款之外,均受到著作權保護,並且保留所有的權利。