雙模態影像融合用於正子斷層造影之影像增強

Abstract

正子斷層掃描（positron emission tomography，PET）结合〔18F〕—氟代去氧葡萄糖（fludeoxyglucose，FDG）可以視覺化與神經退化性疾病相關的葡萄糖低代謝之腦區空間分布。起因於PET掃描儀導致的部分體積效應（partial volume effect，PVE）造成PET影像之空間解析度和影像品質下降，本研究旨在利用高空間解析度的解剖性影像，如磁振造影（magnetic resonance imaging，MRI）所提供之組織邊界資訊，開發出完全基於深度學習（deep learning，DL）的部分體積效應校正方法（partial volume correction，PVC），並省去在過去研究中被視為必要的針對PET掃描儀之點擴散函數的假設。 受到現有的深度學習框架用於影像融合（image fusion）的啟發，本研究提出以此技術將解剖資訊融入功能性影像並使其結構細節增強的MRI-styled PET模態。所提出模型的組成包含：一個有雙解碼器（dual encoder）的影像融合之主要網路，用來融合MRI和PET的影像資訊並生成MRI-styled PET；另外，還有兩個用來促進資訊保存的模態轉換之輔助網路，各以MRI-styled PET當作輸入並試圖還原MRI和PET。此外，本研究也提出了數個模組，用於改良基準模型（Baseline）使其更適用於校正部分體積效應，其中包含：為了有效地校正灰質活性低估和白質活性高的誤差，針對解剖資訊來源的調控模組；為了同時增強局部和全面組織對比度的原創損失函數（custom loss function）；能夠進行跨模態資訊交換且利用可訓練參數進行特徵提取和融合的交叉注意力（cross-attention）融合策略。 而因為理論上不存在校正後的影像作為校正基準，本研究也提出了一系列客觀評估的流程，將MRI-styled PET與使用其他校正方法處理後的影像（PVC-PET）或是其他模態如MRI和PET進行不同尺度的比較。分析結果顯示MRI-styled PET的結構相似性（SSIM）和峰值信噪比（PSNR）都有相對Baseline的顯著提升（p < .001），展現了所提出模組對校正的正向增益。更重要的，在阿茲海默症（Alzheimer’s disease，AD）病變相關的腦區中，MRI-styled PET展現出無論疾病階段如何，其部分體積校正程度都相較Baseline和PVC-PET更加穩定的特性。 總結，本研究的價值在於展現了基於深度學習進行部分體積校正的可能性，本方法無須針對PET掃描儀之系統參數或是受試者之疾病狀態進行任何假設，因此可以在只能取得影像資料的使用情境下進行校正。不只如此，如能取得不同腦區對於示蹤劑吸收量的比例，本研究也是一個理論上能應用於其他放射性示蹤劑影像的校正方法。為了更嚴謹的評估MRI-styled PET的校正準確度，本研究的下一步即是透過數位模擬進行體積像素（voxel）精度的驗證。

Positron emission tomography (PET) with [18F]-fludeoxyglucose (FDG) can visualize the spatial pattern of glucose hypometabolism related to neurodegeneration in brain imaging. In PET imaging, the partial volume effect often degrades image quality due to the limited spatial resolution of PET scanners. With access only to image space data, this study aims to develop a fully deep-learning (DL)-based partial volume correction (PVC) method, eliminating the need to assume the system point spread function during correction. Inspired by the existing framework of the DL-based image fusion model, we propose a novel imaging modality, MRI-styled PET, leveraging anatomical information from T1-weighted magnetic resonance imaging (T1-MRI) to enhance structural details in FDG-PET. The proposed framework consists of a baseline encoder-decoder image fusion model with dual encoder branches and two auxiliary modality conversion networks; additionally, it includes a series of task-specific modules designed to facilitate the transition from image fusion to anatomy-guided DL-based PVC. First, substituting the anatomical input source, T1-MRI, with ideal tracer maps resulted in effective correction in the underestimation of gray matter and the underestimation of white matter. Subsequently, two novel loss functions were introduced to enhance local and global tissue contrast: the adaptive multi-scale structural similarity loss and tissue-aware boundary loss. In addition, the proposed cross-attention fusion strategy outperformed other fusion strategies by incorporating cross-modality information exchange and utilizing trainable parameters for feature extraction and fusion. Compared to a reference anatomy-guided post-reconstruction PVC-PET method, MRI-styled PET demonstrated significantly higher structural similarity (SSIM) and peak signal-to-noise ratio (PSNR) than the baseline image fusion model (Baseline), showcasing the effectiveness of the proposed task-specific modules. In several Alzheimer's Disease-related brain regions, MRI-styled PET exhibited consistent increases and reduced noise-amplification in corrective effects regardless of disease stage, compared to Baseline and PVC-PET. This study represents an initial exploration of a fully DL-based PVC method without prior knowledge regarding the PVC-PET method or the underlying radiotracer uptake and without assumptions about the system point-spread function. Despite being a universal PVC pipeline, MRI-styled PET is currently applicable only to FDG-PET imaging due to the need for general agreement on the ratio of regional tracer uptake. Further investigations into the pixel-wise correction accuracy are planned to understand the MRI-styled PET's effectiveness fully.