深度學習輔助低劑量半導體單光子斷層心肌灌注掃描血流定量

Abstract

心血管疾病是全球的頭號死因，其中冠狀動脈心臟病貢獻了很大的比例。冠狀動脈心臟病的診療過程中，心肌灌注掃描是一項重要的功能性評估工具。傳統靜態掃描只能評估相對血流與相對血流儲備，對於多血管疾病造成的平衡性缺血、瀰漫性或微小血管疾病難以有效偵測。動態心肌灌注掃描可以用來定量心肌血流的絕對值，進而克服這些問題。但過去礙於時間解析度的問題，動態心肌灌注掃描需要以正子斷層掃描進行。雖然近年半導體攝影機引進之後有機會可以用較低成本及更為普及的單光子斷層進行動態心肌灌注掃描達成血流定量，常用的鎝-99m標記的心肌灌流藥物受限於較差的生理特性；而鉈-201雖有較佳的生理特性，但有較高的輻射劑量及影像雜訊。本論文旨在開發和驗證一種使用深度學習方法的低劑量鉈-201動態單光子斷層心肌灌注掃描造影方式，其目的是在減少輻射暴露的同時，提升影像品質和心肌血流定量的準確度。 首先，本文對標準劑量動態心肌灌注掃描造影方式進行了改良與驗證，利用可變時間解析度先驗知識降低影像雜訊，及深度學習驅動的自動化處理減少人為誤差，並與黃金標準的動態正子斷層心肌灌注掃描測量結果進行比較。接著，本文設計了一種模擬低劑量動態研究的方法，生成配對的低劑量和全劑量動態單光子斷層掃描資料集。深度學習模型被訓練來對這些低劑量影像降噪，並從降噪後的影像定量出心肌血流，並將結果與正子掃描及全劑量測量結果進行比較，以評估其準確性。本研究還將心肌血流定量結果與冠狀動脈心臟病的存在及患者的預後進行了相關性分析。 結果顯示，對全劑量鉈-201動態單光子斷層掃描影像降噪進而定量出的心肌血流與正子掃描的定量結果有窄的一致性界線。定量出的壓力心肌血流能夠有效預測冠狀動脈狹窄，其結果與傳統的總灌注缺損相當，甚至對於缺損較輕微的案例能夠更為準確的預測冠狀動脈心臟病的存在。較高的壓力心肌血流與較低的主要不良心血管事件及全因死亡的概率相關，更表明此方法具有預後價值。對於模擬出僅有20%光子密度的低劑量動態影像所定量出的血流數值與正子掃描有寬的一致性界線。而經由深度學習降噪模型處理過後可以有效的縮窄一致性界線。由降噪後的低劑量影像定量出的壓力心肌血流也能預測冠狀動脈狹窄，對於缺損輕微的案例也能更為準確地預測冠狀動脈心臟病的存在。低劑量降噪影像的定量結果也能預測主要不良心血管事件及全因死亡。 本論文探討了深度學習技術在低劑量動態鉈-201單光子斷層心肌灌注掃描中輔助心肌血流量量化的應用，旨在減少患者輻射暴露的同時保持診斷準確性。研究顯示，深度學習顯著提升了低劑量掃描的影像品質，使其可與標準劑量圖像相當。經與正子掃描定量結果驗證，降噪後的低劑量掃描能有效預測冠狀動脈心臟病，並在早期疾病檢測和預後評估方面優於傳統方法。本論文還強調了鉈-201在高流量條件下具有更優異的線性萃取分率，以實現精確的血流量化。儘管樣本量較小等限制存在，研究結果表明，深度學習輔助的低劑量鉈-201單光子斷層心肌灌注掃描是動態正子斷層心肌灌注掃描一個有前景的替代方案，提供了一種在核醫心臟學中更易於獲取的診斷方法。

Cardiovascular diseases are the global leading cause of death, with coronary artery disease (CAD) contributing significantly to this statistic. Myocardial perfusion imaging (MPI) is an essential functional assessment tool for the diagnosis and treatment of CAD. Traditional static MPI can only evaluate relative blood flow and relative flow reserve, making it difficult to effectively detect balanced ischemia caused by multivessel, diffuse, or microvascular diseases. Dynamic MPI can quantify the absolute value of myocardial blood flow (MBF), thereby overcoming these issues. However, owing to limitations in time resolution, dynamic MPI traditionally requires positron emission tomography (PET). Although the introduction of semiconductor cameras in recent years has made it possible to achieve blood flow quantification through dynamic MPI using more cost-effective and widely available single-photon emission computed tomography (SPECT), the commonly used Tc-99m-labeled perfusion agents are limited by their poor physiological characteristics. Although Tl-201 has better physiological characteristics, it has higher radiation dosimetry and image noise. This thesis focuses on developing and validating a low-dose dynamic Tl-201 SPECT MPI protocol with the assistance of deep learning methods. The primary aim is to create a protocol that enhances the image quality and quantification accuracy of MBF while minimizing radiation exposure. A standard-dose dynamic SPECT MPI protocol was initially validated against gold-standard dynamic PET measurements. We developed a variable temporal resolution prior to reduce image noise during the reconstruction process and a deep learning-driven automatic process to reduce human processing error. Subsequently, a method to simulate low-dose dynamic studies was devised, generating a paired low-dose and full-dose dynamic SPECT dataset. Deep learning models were trained to denoise these low-dose images, and the MBF was quantified from the denoised images. The results were compared with those of PET to evaluate accuracy. Furthermore, this study correlated MBF measurements with the presence of CAD and patient prognosis. These findings indicate that MBF measurements from denoised full-dose SPECT scans have narrow limits of agreement with PET measurements. The quantified stress MBF can effectively predict CAD. Stress MBF outperformed conventional stress total perfusion deficit (TPD) in patients with minor perfusion defects. Moreover, a higher stress MBF was associated with lower probabilities of major adverse cardiovascular events (MACE) and all-cause mortality, demonstrating the prognostic value of this approach. In simulated low-dose scans containing only 20% count density, the quantified MBF measurements have wider limits of agreement with PET measurements. The deep learning denoise model can effectively narrow these limits. Stress MBF from low-dose scans can also predict CAD and outperform TPD in patients with minor defects. Low-dose measurements also predict MACE and all-cause mortality. This thesis explores deep learning techniques to assist in the quantification of MBF using low-dose dynamic Tl-201 SPECT MPI, aiming to reduce patient radiation exposure while maintaining diagnostic accuracy. Research has demonstrated that deep learning significantly improves the quality of low-dose SPECT images, making them comparable to standard-dose images. Validated against PET measurements, denoised low-dose scans effectively predicted CAD and outperformed traditional methods in early disease detection and prognosis assessment. The study also highlighted Tl-201's superior linear extraction fraction for accurate flow quantification under high-flow conditions. Despite limitations such as small sample size, the findings suggest that deep learning-assisted low-dose SPECT MPI is a promising alternative to dynamic PET MPI, offering a more accessible approach in nuclear cardiology.