雙頭平面PET系統的影像重建與質子治療中的深度學習應用&quot;Image Reconstruction for Dual Head Plane PET System and Deep Learning Based Application in Proton Therapy

NCU Institutional Repository > 理學院 > 物理研究所 > 博碩士論文 > Item 987654321/95041

jsp.display-item.identifier=請使用永久網址來引用或連結此文件: http://ir.lib.ncu.edu.tw/handle/987654321/95041

题名:	雙頭平面PET系統的影像重建與質子治療中的深度學習應用";Image Reconstruction for Dual Head Plane PET System and Deep Learning Based Application in Proton Therapy
作者:	爾拉;Rahman, Atiq Ur
贡献者:	物理學系
关键词:	正子發射斷層掃描;正子發射斷層掃描（PET）成像;三角度影像重建;質子劑量預測;醫學成像中的深度學習;Positron Emission Tomography;PET Imaging;Three angle image reconstruction;Proton Dose Prediction;Deep Learning in Medical Imaging
日期:	2024-04-29
上传时间:	2024-10-09 15:44:54 (UTC+8)
出版者:	國立中央大學
摘要:	本論文的第一部分探討了正子發射斷層掃描（PET）成像的前沿技術，通過開發一種新型旋轉雙頭PET系統，旨在克服緊湊型PET探測器中固有的空間分辨率限制。該方法的核心是引入一種基於最大似然估計方法（MLEM）的三角重建技術。我們使用GATE/Geant4 10.4模擬工具包進行探測器模擬，並開發了自己的圖像重建框架。估計了圖像分辨率和探測器靈敏度。使用三角重建方法，沿有限角度軸的分辨率從3.6毫米提高到1.7毫米。論文的第二部分深入研究了通過深度學習提高粒子治療劑量分佈的準確性。我們關注從探測器數據到內在劑量分佈的轉換，使用GATE/Geant4對暴露於高能質子的人類CT模型進行蒙特卡羅模擬。採用條件生成對抗網絡，我們開發了一個神經網絡模型，從PET符合分佈中推斷劑量圖。我們的模型通過平均相對誤差和布拉格峰位置偏差進行評估，顯示在單能量輻照中，劑量偏差在1％以內，範圍在2％以內，且在實際的展寬布拉格峰條件下性能保持穩定。這項工作證明了深度學習用於將低計數數據映射到劑量分佈的可行性，為粒子治療成像帶來了進展。項目的第三部分檢查了質子治療劑量沉積的精確度，重點是使用中研院PET的模塊化設計以及SiPMs和STiC asic讀出，進行範圍驗證。該研究展示了130 MeV質子輻照PMMA時正子發射體的深度分佈，並將長庚紀念醫院緊湊的32通道設置的測量結果與Geant4模擬進行比較。通過計時能力和不同同位素的多指數擬合分析的結果驗證了解剖變化估計在連續治療期間的實施。此外，我們介紹了AS-PET的模塊化設計及其模擬成像性能，突出其在質子治療範圍驗證中的潛力。這項論文研究可以幫助改善平板PET系統的PET圖像質量，可作為粒子治療中治療計劃和質量保證的有價值工具。;The first part of the thesis explores the frontier of Positron Emission Tomography (PET) imaging through the development of a novel rotating dual-head PET system, aimed at overcoming the inherent spatial resolution limitations within compact PET detectors. Central to this method is the introduction of a three-angle reconstruction technique utilizing the basic Maximum Likelihood Estimation Method (MLEM). We use GATE/Geant4 10.4 simulation toolkit to perform detector simulation and have developed our own image reconstruction framework.Image resolution and detector sensitivity are estimated. The resolution along limited-angle axis is improved from 3.6 mm to 1.7 mm using three-angle reconstruction approach. The second part of the thesis delves into enhancing dose distribution accuracy in particle therapy through deep learning. Focusing on the transition from detector data to intrinsic dose distributions, we utilize Monte Carlo simulations with GATE/Geant4 on a human CT phantom exposed to high-energy protons. Employing a conditional generative adversarial network, we develop a neural network model to infer dose maps from PET coincidence distributions. Our model, evaluated by mean relative error and deviations in Bragg peak position, demonstrates deviations within 1% for dose and 2% for range in mono-energetic irradiations, with performance sustained under realistic spread-out Bragg peak conditions. This work underscores the feasibility of deep learning for mapping low count data to dose distributions, promising advancements in particle therapy imaging. The third part of project examines the precision of proton therapy dose deposition, focusing on range verification using the Academia Sinica PET′s modular design with SiPMs and STiC asic readout. This study presents the positron emitter depth distribution in PMMA irradiated by 130 MeV protons, comparing measurements from a compact 32-channel setup at Chang Gung Memorial Hospital with Geant4 simulations. The results, validated by timing capabilities and multi-exponential fit analysis of different isotopes, confirms the implementation on anatomical change estimation during successive treatment sessions. Additionally, we introduce the AS-PET′s modular design and its simulated imaging performance, highlighting its potential in range verification for proton therapy. This thesis research can help to improve the PET image quality for flat-panel PET systems and can be a valuable tool for treatment planning and quality assurance in particle therapy.
显示于类别:	[物理研究所] 博碩士論文

文件中的档案:

档案	描述	大小	格式	浏览次数
index.html		0Kb	HTML	31	检视/开启

在NCUIR中所有的数据项都受到原著作权保护.

社群 sharing

数据加载中.....