利用數位模擬驗證基於深度學習之正子斷層造影之部分體積效應校正方法

Abstract

正子斷層掃描（PET）能提供關鍵的生物性資訊，特別是在診斷如阿茲海默症（Alzheimer’s disease, AD）等神經退化性疾病上具有重要價值。 正子斷層掃描（PET）影像中常見的體積效應（Partial Volume Effects, PVE）會導致顯著的影像錯誤。為了解決此問題，研究人員已發展出基於解剖結構資訊的體積效應校正（Partial Volume Correction, PVC）方法。此類方法通常結合來自磁振造影（MRI）的解剖資訊，但多數仍需獲取掃描儀的點擴散函數（Point Spread Function, PSF）。MRI-styled PET 是一種基於深度學習的 PVC 方法，僅利用 PET 影像與由 MRI 影像衍生的解剖資訊進行校正，不需額外的掃描儀特定參數。初步驗證結果顯示，MRI-styled PET 相較於傳統 PVC 方法，具有更良好的校正性能。

本研究進一步透過數位模擬驗證流程，透過控制模擬參數調整不同的影像品質，以進行更完整的評估 MRI-styled PET 的校正表現。

為驗證 MRI-styled PET 的穩健性，我們設計了四項實驗，分別控制不同變因，包括腦部結構差異、病理疾病差異、掃描儀結構與影像重建方法。模擬流程如下：從 MRI 影像產生五十組數位 [18F]-FDG PET 數位模型，涵蓋不同認知階段的受試者。針對各個解剖區域進行分割並賦予對應的放射性活性濃度，接著依據具代表性的掃描儀建構成像模型產生正弦圖並加入衰減與雜訊的因子，使用 PET 影像重建方法進行影像重建。主要採用 OSEM 演算法（10 個子集、20 次迭代），少數情況下則使用其他方法，如濾波反投影法（FBP）與核期望最大化法（KEM）。模擬成像與影像重建皆透過 SIRF（Synergistic Image Reconstruction Framework）框架實現。

模擬結果顯示，本研究設計之模擬流程具備高度彈性，能夠模擬多樣且真實的影像條件，包括掃描儀特性、解剖結構與放射性分布等多樣化的特徵，有助於進行全面且可控的校正方法驗證。此外，MRI-styled PET 在多種模擬情境下皆展現出穩定可靠的校正效能，且最佳解剖輸入會依不同情況有所不同。綜合而言，研究結果凸顯出 MRI-styled PET 在 PET 影像體積效應校正方面的穩健性、適應性與高度泛用性。

Positron emission tomography (PET) provides crucial biological information, particularly valuable for detecting neurodegenerative diseases such as Alzheimer's disease (AD). Partial volume effects (PVE) in PET imaging can lead to significant quantification inaccuracies. To address these effects, anatomy-guided post-reconstruction partial volume correction (PVC) techniques have been developed. These methods incorporate anatomical information(typically from MRI)but often require accurate estimation of the point spread function (PSF) of scanners. MRI-styled PET is a deep learning-based PVC method that utilizes only PET data and anatomical inputs derived from MR images, eliminating the need for scanner-specific parameters. Initial validations have shown that MRI-styled PET delivers superior correction performance compared to conventional PVC approaches.

This study aims to further evaluate the correction performance of MRI-styled PET using a digital phantom simulation pipeline that systematically varies key parameters to emulate different imaging conditions.

To assess the robustness of MRI-styled PET, four experiments were conducted, each controlling a specific variable, such as pathological condition, scanner model, or reconstruction method. The simulation protocol is summarized as follows: Fifty digital [18F]-FDG PET phantoms were generated from MR images, with balanced representation across different cognitive stages. Anatomical regions of interest were segmented and assigned relative activity concentrations. Phantom sinograms were generated using acquisition models based on selected scanners, with additional factors such as attenuation and noise incorporated. PET images were then reconstructed which primarily using the OSEM algorithm (10 subsets and 20 iterations) with a few cases reconstructed using alternative methods such as filtered back-projection (FBP) and kernel expectation maximization (KEM). Both acquisition modeling and image reconstruction were implemented using the Synergistic Image Reconstruction Framework (SIRF).

The results indicate that the simulation pipeline is highly flexible and capable of modeling a wide range of realistic imaging conditions, including variations in scanner characteristics, anatomical structures, and tracer distributions. This flexibility allows for comprehensive and controlled validation of correction techniques. Furthermore, the results demonstrate that MRI-styled PET consistently achieves reliable correction performance across diverse scenarios. The optimal anatomical input may vary depending on the specific condition. Overall, these findings underscore the robustness, adaptability, and generalizability of MRI-styled PET as an effective PVC solution for PET imaging.