開源PET動力學建模軟體的實作與驗證

Abstract

研究背景與目的：正子斷層造影(PET)是一種重要的醫學成像技術，已被廣泛應用於疾病診斷、藥物開發等研究領域，而在PET分析中，動力學建模具有其定量分析的獨特優勢，但需要合適且具公信力的軟體來進行建模分析。如常見的PMOD、SAAM II等建模軟體。本文旨在探討一個有別於一般商用軟體的開源建模工具箱QModeling的驗證，以及應用於分析大腦類澱粉血管病變造成的腦內出血 (CAA-ICH)、高血壓造成的腦內出血(HTN-ICH)和阿茲海默症 (AD)的初步結果。 材料與方法：本研究使用動力學建模分析最常應用的商用軟體PMOD來進行QModeling內建四個模型的數值驗證及分析。QModeling現有的模型包含簡易參考組織模型(SRTM)和簡易參考組織模型2 (SRTM2)、Logan參考組織法(Logan Reference Plot)、Patlak參考組織模型(Patlak Reference Model)，以及2-組織室模型(2-Tissue Compartment Model)。為驗證此軟體內建模型的可信度，因此準備了一組公開動態PET資料集，以及一組CAA-ICH、HTN-ICH及AD的資料集。公開資料集為NRM2018 PET Grand Challenge Dataset，此資料集為評估不同的PET分析工具之效能，其中包含5位模擬受試者，並使用模擬的神經遞質受體示蹤劑[11C]LondonPride，分別進行2次90分鐘的動態擷取，共10筆的動態PET的資料集。第二組為台大醫院核醫部所建立的資料集，包含14位 CAA-ICH、15位AD、以及10位HTN-ICH，分別進行70分鐘的動態擷取，共23幀。透過動力學模型的建立及選擇，將擬合完成的動力學參數如BP_nd進行分析比對，取得CAA-ICH、HTN-ICH、AD三者之間的差異性。 實驗結果：在使用PMOD作為標準的QModeling性能驗證中，SRTM的三項主要擬合參數(R_1、k_2、BP_nd)的絕對誤差和相對誤差分別為：R_1絕對誤差為(2.62±2.02)10-3 (l/l)，相對誤差為(5.01±2.38)10-3 (l/l)，k_2的絕對誤差為(7.24±2.29)10-4 (min-1)，相對誤差為(8.38±0.89)10-3 (min-1)，BP_nd的絕對誤差為(1.6±9.6)10-3 (l/l)，相對誤差為(4.53±3.78)10-2 (l/l)。SRTM2的三項擬合(R_1、k_2^'、BP_nd)，R_1絕對誤差為(9.22±4.87)10-4 (l/l)，相對誤差(2.4±0.58)10-3 (l/l)，k_2^'絕對誤差為(1.29±2.56)10-3 (min-1)，相對誤差為(2.09±0.69)10-2 (min-1)，BP_nd絕對誤差為(3.3±7.1)10-4 (l/l)，相對誤差為(7.11±2.4)10-3 (l/l)，Logan參考組織法的BP_nd絕對誤差為(1.1±1.08)10-4 (l/l)，相對誤差為(1.11±1.87)10-4 (l/l)。而在CAA-ICH、HTN-ICH、AD的初步實驗中，可以發現在枕葉及後扣帶皮質有較顯著的傾向能區分CAA-ICH與AD，在頂葉、灰質、白質則較無明顯差別。 結論：此研究利用PMOD和公開的動態PET數據集來驗證QModeling的效能，測試了SRTM、SRTM2和Logan參考組織法的七個動力學參數。研究結果顯示，QModeling在絕對誤差和相對誤差上與PMOD之間的差異均小於10-3，證明了其在無需動脈輸入函數情況下的可靠性和實用性。而在初步的CAA-ICH、AD和HTN-ICH的案例分析中，枕葉和後扣帶皮質的分析顯示出明顯的差異，進一步強調了QModeling作為開源工具的潛力和實用性。

Background and Objectives: Positron emission tomography (PET) is an important medical imaging technology widely used in disease diagnosis, drug development, and other research areas. In PET analysis, kinetic modeling offers unique advantages as a quantitative analysis method, requiring reliable and credible software for modeling and analysis, such as commonly used software like PMOD and SAAM II. This study aims to explore the validation of an open-source modeling toolkit, QModeling, which differs from general commercial software, and its preliminary application in analyzing images from populations with brain hemorrhages caused by cerebral amyloid angiopathy (CAA-ICH) or hypertension (HTN-ICH) and those with Alzheimer's disease (AD). Materials and Methods: This study used the commonly applied commercial software PMOD for kinetic modeling analysis to verify and analyze the numerical accuracy of four models built into QModeling. The existing models in QModeling include the Simplified Reference Tissue Model (SRTM), Simplified Reference Tissue Model 2 (SRTM2), Logan Reference Plot, Patlak Reference Model, and the 2-Tissue Compartment Model. To validate the reliability of these built-in models, a publicly available dynamic PET dataset and a dataset of CAA-ICH, HTN-ICH, and AD were prepared. The public dataset is the NRM2018 PET Grand Challenge Dataset, established to evaluate the performance of different PET analysis tools and confirm the receptor binding changes in PET radioligand neurotransmission studies. It included five simulated subjects using the simulated neurotransmitter receptor tracer [11C]LondonPride, each undergoing two 90-minute dynamic acquisitions, totaling ten dynamic PET datasets. The second dataset was established by the Nuclear Medicine Department of National Taiwan University Hospital and included 14 CAA-ICH, 15 AD, and 10 HTN-ICH patients, each undergoing 70-minute dynamic acquisitions, totaling 23 frames per acquisition. By establishing and selecting kinetic models, fitted kinetic parameters such as BP_nd were analyzed and compared to obtain differences among CAA-ICH, HTN-ICH, and AD. Results: Using PMOD as the standard for the validation of QModeling's model performance, the absolute and relative errors of the three main fitting parameters (R_1, k_2, BP_nd) for SRTM were: R_1 absolute error (2.62±2.02)10-3 (l/l), relative error (5.01±2.38)10-3 (l/l); k_2 absolute error (7.24±2.29)10-4 (min-1), relative error (8.38±0.89)10-3 (min-1); BP_nd absolute error (1.6±9.6)10-3 (l/l), relative error (4.53±3.78)10-2 (l/l). For SRTM2, the absolute and relative errors for R_1 were (9.22±4.87)10-4 (l/l) and (2.4±0.58)10-3 (l/l), respectively; for k_2^*, the absolute error was (1.29±2.56)10-3 (min-1), relative error (2.09±0.69)10-2 (min-1); for BP_nd, the absolute error was (3.3±7.1)10-4(l/l), relative error (7.11±2.4)10-3 (l/l). For the Logan Reference method, the BP_nd absolute error was (1.1±1.08)10-4 (l/l), relative error (1.11±1.87)10-4 (l/l). In the preliminary experiments on CAA-ICH, HTN-ICH, and AD, significant differentiation between CAA-ICH and AD was observed in the occipital lobe and posterior cingulate cortex, whereas no obvious differences were noted in the parietal lobe, gray matter, and white matter. Conclusion: This study utilized PMOD and public dynamic PET datasets to validate the efficacy of QModeling, testing seven kinetic parameters across SRTM, SRTM2, and Logan reference tissue methods. Results indicate that QModeling's absolute and relative errors, in comparison to PMOD, were less than 10-3, demonstrating its reliability and practicality without arterial input function. Preliminary analyses of CAA-ICH, AD, and HTN-ICH cases, particularly in the occipital lobes and posterior cingulate cortex, highlighted significant differences, underscoring QModeling's potential as an open-source tool.