AI driven reconstruction, calibration and data analysis within a dual-head PET detector framework

NCU Institutional Repository > 理學院 > 物理研究所 > 博碩士論文 > Item 987654321/99305

請使用永久網址來引用或連結此文件: https://ir.lib.ncu.edu.tw/handle/987654321/99305

題名:	AI driven reconstruction, calibration and data analysis within a dual-head PET detector framework
作者:	陳攻善;Thien, Tran Cong
貢獻者:	物理學系
關鍵詞:	套緊湊型雙面板PET 系統;Dual-head PET detector
日期:	2025-12-29
上傳時間:	2026-03-06 18:35:14 (UTC+8)
出版者:	國立中央大學
摘要:	正子放射斷層造影（Positron Emission Tomography, PET）可提供放射性示蹤劑分佈的定量影像資訊，然而傳統全環式 PET 系統成本高昂，且對基礎設施需求甚鉅，限制其應用彈性。本論文針對一套緊湊型雙面板 PET 系統，探討影像重建之計算方法與資料驅動式技術，並以深度學習為核心，發展直接影像重建方法，同時建立支援該重建流程所需之系統校正與特性量測程序。本研究所使用之 TOF-PET 系統由兩組相對配置的探測模組構成，每組包含 512 個 LYSO 閃爍晶體通道。為確保影像重建品質，首先完成系統基礎校正程序，包括能量校正（511 keV 下能量解析度為 12.9% FWHM）、時間校正（同時計時解析度達 255 ps），以及時間游移修正，使符合事件的時間視窗由 25 ns 縮減至 2.5 ns。此外，本研究亦建立一套非破壞性方法，利用 LYSO 晶體本身的內在放射性量測其鎂酸鋰成分比例，在適當樣本條件下可達低於 1% 的組成不確定度。本論文的核心貢獻為一套以深度學習為基礎的 PET 直接影像重建框架。研究中設計了一種條件式生成對抗網路（conditional GAN, cGAN），結合 2.5D sinogram 表示方式與適用於雙面板幾何結構的多角度量測策略，可直接由量測資料重建影像，並達到秒級重建速度。定量評估結果顯示，在測試資料集上可穩定達到 37 dB 以上之 PSNR，顯示良好的影像品質與穩定性。進一步地，本研究將該系統應用於質子治療射程驗證之正子湮滅伽瑪（Positron Annihilation Gamma, PAG）成像。於多組體模實驗中，在 60 秒量測時間內可達 1.0–1.7 mm 的射程不確定度。此外，亦發展一套時間分析方法以區分 O-15 與 C-11 放射性核種的貢獻，進一步評估其對 PAG 訊號之影響。總結而言，本研究完整建立了一套以重建為導向之雙面板 PET 系統研究流程，涵蓋系統校正、材料特性分析，以及深度學習式直接影像重建。研究結果顯示，AI 驅動之重建方法能在有限角度與低計數條件下產生快速且穩定的影像，顯示緊湊型 PET 系統於特定臨床與治療應用上的可行性，並為未來進一步提升影像準確度與系統效能提供明確的研究方向。;Positron Emission Tomography (PET) enables quantitative imaging of radiotracer distributions but conventional full-ring systems remain costly and infrastructure-intensive. This thesis investigates computational and data-driven methods for image reconstruction in a compact dual-head PET system, with an emphasis on direct reconstruction using deep learning. The work also provides the essential calibration and characterization steps required to support these reconstruction developments. The TOF-PET system utilized in this study consists of two opposing detector modules, each with 512 LYSO crystal channels. Baseline calibration procedures - energy calibration (12.9% FWHM at 511 keV), timing calibration (255 ps coincidence timing resolution), and time-walk correction (coincidence window reduced from 25 ns to 2.5 ns) - were implemented to establish consistent system performance for reconstruction studies. A non-destructive method for determining lutetium concentration in LYSO using intrinsic radiation measurements was developed, obtaining composition uncertainties below 1% for appropriately sized samples. The central contribution of this thesis is a deep-learning-based framework for direct image reconstruction from PET simulations and measurements. A conditional Generative Adversarial Network (cGAN) architecture was designed using a 2.5D sinogram representation and multi-angle acquisition scheme adapted to the dual-head geometry. This original approach produces PET images at second-level reconstruction speeds. Quantitative evaluation demonstrated stable performance on test datasets, with reconstructed images achieving PSNR values above 37 dB for the targeted imaging scenarios. The system was further assessed in the context of proton therapy range verification using positron annihilation gamma (PAG) imaging. Range verification uncertainties of 1.0–1.7 mm were achieved within 60-second measurements across multiple phantom configurations. A temporal-analysis method for separating O-15 and C-11 activity distributions was developed, enabling assessment of isotope-specific contributions to the PAG signal. Overall, this work provides a complete reconstruction-oriented study of a compact dual-head PET system, combining calibration, intrinsic characterization, and deep-learning-based direct reconstruction. The results demonstrate that AI-driven reconstruction methods can produce rapid and stable images from limited-angle PET acquisitions, supporting the feasibility of compact PET systems for specialized applications. A clear roadmap to expand on this work to improve the accuracy of image generation has been suggested.
顯示於類別:	[物理研究所] 博碩士論文

文件中的檔案:

檔案	描述	大小	格式	瀏覽次數
index.html		0Kb	HTML	24	檢視/開啟

在NCUIR中所有的資料項目都受到原著作權保護.

社群 sharing

資料載入中.....