探討在以深度學習方式生成類澱粉蛋白沉積正子斷層掃描影像的過程中加入多模態解剖性影像之成效

Abstract

背景介紹 類澱粉蛋白正子斷層掃描（PET）在診斷腦部疾病，如阿茲海默症方面是一項強大並且廣泛應用的造影技術。最近的研究亦顯示從早期（Early-phase）類澱粉蛋白放射性示蹤劑成像中能夠獲取腦部血流的相關訊息。然而，受檢者需要經過長時間的等待使示蹤劑於腦部的吸收達到穩定狀態，此漫長的等待過程可能會給病患帶來一定程度的不適和焦慮感，增加病患在掃描過程中移動的可能性，進而造成影像上假影的產生，降低其診斷品質。為了縮短整體類澱粉蛋白PET的檢查時間，以及在單次掃描中能同時取得腦血流和腦部類澱粉蛋白沉積的相關資訊，將利用早期類澱粉蛋白的PET影像透過深度學習的方法生成晚期（Late-phase）類澱粉蛋白的PET影像，並且在生成過程中加入多模態的解剖性影像（電腦斷層掃描[CT]和磁共振造影[MRI]），探討其對於影像生成之影響。 研究方法 我們在此回溯性研究中所使用的影像資料來自於52位受試者，PET檢查中所使用的類澱粉蛋白放射性示蹤劑為[11C]-Pittsburgh Compound-B (PiB)，而PET影像的收取採用動態掃描模式。研究中我們使用MATLAB (R2021b) 和SPM12進行影像的前處理。晚期類澱粉蛋白PET影像（Ground Truth）利用示蹤劑注射後40~70分鐘的PET影像產生，而早期類澱粉蛋白PET影像則透過示蹤劑注射之後前20分鐘的PET影像來生成。我們訓練U-net神經網絡來生成影像，神經網絡的輸入有五種不同情況（Model 1：只有早期類澱粉蛋白PET影像、Model 2：早期類澱粉蛋白PET和CT影像、Model 3：早期類澱粉蛋白PET和T1加權MR影像、Model 4：早期類澱粉蛋白PET、CT和T1加權MR影像以及Model 5：早期類澱粉蛋白PET和T2加權MR影像）。此外，在過程中我們加入資料增強與交叉驗證的方法幫助神經網絡訓練。最後利用峰值訊噪比（PSNR）、結構相似性（SSIM）以及標準攝取值比值（SUVr）對生成影像與晚期類澱粉蛋白PET影像進行比較，以此來評估網絡生成影像之成效。 結果 由神經網絡所生成的影像和作為網絡輸入的早期類澱粉蛋白PET影像相較之下有著明顯圖像視覺上的改善，更加接近晚期類澱粉蛋白PET影像，但卻顯得較為平滑。此外，於峰值訊噪比以及結構相似性兩項指標中皆表現出統計上顯著性的增加（p<0.05）。然而，經由不同組合的影像輸入所生成的結果在彼此之間雖然有數值上的差異，但皆未呈現出統計上的顯著性。在SUVr指標中，根據生成影像與晚期類澱粉蛋白PET影像在額葉、頂葉、顳葉、後扣帶皮層以及楔前葉這五個腦區以小腦做為參考區域所計算出的結果，顯示兩者於這五個腦區的SUVr值在統計上皆未有顯著性的差異存在。但是在加入多模態解剖性影像前後並未觀察到明顯的趨勢變化。 結論 研究結果顯示透過深度學習的神經網絡對晚期類澱粉蛋白PET影像進行預測具有可行性。此外，在訓練過程中加入CT與MR等不同模態的解剖性影像對於網絡生成PET影像亦是可行的。然而在峰值訊噪比、結構相似性指標中，加入多模態解剖性影像前後所得到的數值並未有顯著性的差異，而在SUVr指標中亦未觀察到明顯的趨勢變化。因此，在神經網絡訓練過程中加入解剖性影像並不會對影像預測及生成產生負面的影響，然而是否具有正向的助益還需要更進一步的研究來驗證。

Introduction Amyloid positron emission tomography (PET) is a powerful and widespread imaging technique in diagnosing brain disorders such as Alzheimer's disease. Recent findings have also shown that information related to blood flow can be obtained with early-phase amyloid radiotracer imaging. However, it takes a long period of time for the radiotracer uptake to reach steady-state, which may bring discomfort and anxiety to patients and increase the possibility of image artifacts caused by patient motion. For shortening the total duration of the amyloid PET exam and to obtain both flow-related and amyloid plaque deposition information in a single scan, deep learning methods have been proposed to generate late-phase amyloid PET images from dynamic early-phase amyloid PET images. Here, the contribution of multimodal anatomical images (computed tomography [CT] and magnetic resonance imaging [MRI]) in PET image synthesis is investigated. Methods The images used in this retrospective study were obtained from 52 participants. The amyloid radiotracer [11C]-Pittsburgh Compound-B (PiB) was administered and scanned with a dynamic PET protocol. MATLAB (R2021b) and SPM12 were used for image preprocessing. The late-phase PET image (ground truth) was created by the 40-70 minutes post-injection dynamic PET data, and the early-phase PET image was generated by averaging the frames from 0 to 20 minutes post-injection, respectively. U-nets were trained in this work with the inputs being combinations of various multimodal images (Model 1: early-phase PET images only, Model 2: early-phase PET and CT images, Model 3: early-phase PET and T1-weighted MR images, Model 4: early-phase PET, CT, and T1-weighted MR images, and Model 5: early-phase PET and T2-weighted MR images). Data augmentation and 10-fold cross-validation were used in the network training. For each axial slice, the image quality of the synthesized PET image was compared to the ground truth image using peak signal-to-noise ratio (PSNR), structural similarity (SSIM), and standardized uptake value ratio (SUVr). Results Qualitatively, the images generated by the networks exhibited visible improvements compared to the early-phase PET images and were visually similar to the late-phase PET images. However, the synthesized images were smoother than the late-phase PET images. In terms of PSNR and SSIM, the results generated from the five networks all demonstrated significant improvements (p-value < 0.05) when compared to the original early-phase PET images. Nevertheless, no significant differences were observed in PSNR and SSIM between all models. In terms of SUVr, there were no significant differences in results calculated from frontal lobe, parietal lobe, temporal lobe, posterior cingulate, and precuneus with cerebellum as reference region between all synthesized images and the late-phase PET images, but no specific trend was observed after incorporating CT and MR images into network training. Conclusion This study demonstrated that predicting the late-phase PET image from the early-phase PET image via deep learning neural networks is feasible. Moreover, the feasibility of incorporating multimodal anatomical images into the late-phase PET image generation by neural networks was also verified. However, there were no significant differences in PSNR as well as SSIM, and no specific pattern was observed in SUVr when incorporating various combinations of multimodal anatomical images. Therefore, incorporating multimodal anatomical images into neural network training does not have negative influences on image prediction. Nevertheless, whether it has positive effects on image synthesis requires further investigation.