基於非監督式學習之動態電腦斷層影像去雜訊方法

Abstract

電腦斷層灌注影像（CTP）是一種透過動態 CT 掃描進行非侵入式、定量腦功能評估的影像技術，在急性缺血性中風的診斷與臨床管理中扮演重要角色。然而，動態 CT 掃描的輻射劑量較高。雖然降低管電流–時間乘積可減少輻射劑量，但同時會導致影像雜訊增加，進而影響腦灌注分析的可靠性。為了降低影像雜訊，我們提出一種基於非監督式學習的方法，整合一維卷積神經網路、雙向長短期記憶、雙向閘門循環單元以及多頭注意力機制。由於其非監督式學習的特性，該方法無需大量訓練資料與真實有標註的資料。此外，我們將CTP 資料轉換為一系列一維的時間–衰減曲線（TACs），以降低記憶體使用量，並允許同時處理所有 TACs。本方法於模擬與病患 CTP 資料上進行評估，並與高度限制反投影局部重建法（HYPR）、非區域均值法（NLM）及其組合(HYPR-NLM)進行比較。在模擬 CTP 資料中，所提出的方法相較於其他去雜訊方法具有較低的均方根誤差，並產生組織邊界更為銳利的灌注圖。在病患 CTP 資料中，本方法所得到的灌注數值與先前研究結果一致，且灌注圖呈現出較高的對比度與更清晰的組織輪廓。上述結果顯示，所提出的方法在動態 CT 影像去雜訊方面具有潛力，並能有效提升灌注圖的品質。

Computed tomography perfusion (CTP) is an imaging technique that enables non-invasive, quantitative assessment of brain function through dynamic CT acquisitions, serving an important role in the diagnosis and management of acute ischemic stroke. However, the radiation dose for a dynamic CT scan is high. While using a low tube current–time product setting can reduce the dose, it leads to increased image noise, thereby compromising the reliability of perfusion analysis. To reduce image noise, we propose an unsupervised learning based method integrating one-dimensional (1D) convolutional neural networks, recurrent neural networks, and an attention mechanism. Due to its unsupervised learning property, the proposed method eliminates the need for extensive training and ground-truth datasets. Moreover, we transform the spatiotemporal CTP data into a series of 1D time–attenuation curves (TACs) that reduce memory consumption and allow simultaneous processing of all TACs. The proposed method was evaluated using both simulated and patient CTP data, and its performance was compared with highly constrained back-projection local reconstruction, non-local means, and a combination of the two. For the simulated CTP data, the proposed method achieved lower root mean square errors than the other image denoising methods and produced perfusion maps with sharper tissue boundaries. For the patient CTP data, the perfusion values obtained with the proposed method were consistent with those reported in the previous studies, and the perfusion maps exhibited higher contrast and clearer tissue delineation. These results demonstrate the potential of the proposed method to effectively denoise dynamic CT images and improve the quality of perfusion maps.