研發以GPU加速之三邊濾波器從事自動化腦部磁振影像之雜訊去除

Abstract

雜訊去除在影像處理中一直是一個很重要的技術，特別在醫學影像的分析上，去除磁振影像的雜訊對於之後的處理、分析與應用是相當重要的。本研究使用三邊濾波器，其為一延伸雙邊濾波器且更能有效地去除腦部磁振影像中的隨機雜訊。然而，三邊濾波處理過程相當耗費時間且需要做數個參數的調整以得到最佳去除雜訊之影像。因此，本研究使用GPU平行運算來加速整個濾波處理時間，利用GPU強大的平行處理能力，搭配NVIDIA所開發的CUDA架構來加速原本的三邊濾波演算法。接著藉由人工智慧技術來自動最佳化濾波處理，藉由擷取磁振影像的特徵資料，結合類神經網路與支持向量機，訓練出一個可預測最佳三邊濾波模型，進一步建立自動化三邊濾波用於去除腦部磁振影像雜訊的系統。本研究使用t檢定方法加上循序前進浮動選取演算法選取最佳的特徵組合，再將此特徵組合用於訓練自動化去雜訊系統。本實驗使用Brain Web腦部模擬資料庫的磁振影像來分析三邊濾波器之效能與驗證本研究提出的自動化去除雜訊架構。實驗結果顯示本研究所提的方法不僅能大幅地加速傳統三邊濾波演算法，一張磁振影像可以達到34倍之加速。比較其他除雜訊方法，也能有效地自動化去除腦部磁振影像中的雜訊，最後獲得的重建影像有相當不錯的品質。

Noise removal is one of the fundamental and essential tasks within image processing. In medical imaging, finding an effective algorithm that can remove random noise in magnetic resonance (MR) images is important. This thesis proposes an effective noise reduction method for brain MR images. The proposed is based on the trilateral filter, which is a more powerful method than the bilateral filter in many cases. However, the computation of the trilateral filter is quite time-consuming and the choice of the filter parameters is also laborious. To address these problems, the trilateral filter algorithm is implemented using parallel computing with GPU. The CUDA, an application programming interface for GPU by NVIDIA is adopted, to accelerate the computation. Subsequently, the optimal filter parameters are selected by artificial intelligence techniques. Artificial neural networks and support vector machines associated with image feature analysis are proposed to establish the automatic mechanism. The best feature combination is selected by the t-test and the sequential forward floating selection (SFFS) methods. Experimental results indicated that not only did the proposed GPU-based version run dramatically faster than the traditional trilateral filter, but this automatic system also effectively removed the noise in various brain MR images. We believe that the proposed framework has established a general blueprint for achieving fast and automatic filtering in a wide variety of MR image denoising applications.