深度學習技術深度影像先驗在雙能電腦斷層影像上的應用

Abstract

雙能電腦斷層成像技術的問世，近年來在臨床上衍生了多樣的應用。例如:物質組成影像和虛擬單能電腦斷層影像。然而，重建影像雜訊過高一直是傳統影像後 處理方法產生的最大問題。為了提高影像品質，本文提出以非監督式深度學習「深度影像先驗」的概念建構模型，實現由雙能電腦斷層影像產生物質組成影像和虛擬 單能電腦斷層影像，同時改善影像品質。我們搜集了十位接受一般腦部電腦斷層掃描的影像用以測試深度學習模型的可行性。在物質組成影像方面，我們著重於以微調架構的深度學習模型實現雙物質分離(軟組織、骨頭)和三物質分離(軟組織、骨頭和脂肪)。在虛擬單能影像方面，我們使用調整後的深度學習模型評估了重建後虛擬低能量單能影像的影像品質。我們還提出一個簡單的技術用以一次性產生多張不同能量的虛擬單能影像。最後以影像品質指標訊雜比(Signal-to-Noise Ratio, SNR)、解析度指標調製傳遞函數(Modulated Transfer Function, MTF)、結構性相似指標(structural similarity index, SSIM index)為依據，對深度學習模型產生之重建影像和傳統方法之重建影像做評估和分析。初步分析結果顯示，與傳統生成方法所產生的影像相比，深度學習模型產生之物質組成影像和虛擬單能電腦斷層影像均可減少影像雜訊的產生，並維持相似的空間分辨率，可增加臨床價值。

Since its commercialization, dual energy computed tomography (DECT) has shown many important clinical applications. Material decomposition (MD) and virtual monochromatic imaging (VMI) are two common applications of DECT. However, synthesizing basis material images and VMI by traditional methods gave rise to severe noise amplification in the reconstructed images. We proposed an image domain unsupervised deep learning-based method to enhance the quality of basis material images and VMI, based on the concept of deep image prior (DIP). We retrospectively recruited ten patients who received non-contrast DECT brain scan. In the part of MD, we utilized a convolutional neural network (CNN) model with specific modifications to investigate the feasibility of generating two (soft tissue and bone) and three basis materials (soft tissue, bone, and fat) from DECT imaging. In the part of VMI, we used a modified CNN model to generate high-quality VMI from DECT imaging. We also proposed a simple technique to generate multi-energy VMIs from one trained CNN model. In the data analysis, we evaluated the DIP-based MD images and VMI in terms of SNR, MTF, and SSIM index, and compared with conventional DECT-based MD images and VMI. Our preliminary results showed that both the DIP-based MD images and VMI showed superior noise suppression. Moreover, the DIP-based MD images maintained the same spatial resolution compared with the DECT-based MD images. Similarly, the DIP-based VMIs had similar spatial resolution compared with DECT-based VMIs.