適用於超音波系統之低複雜度波束成像引擎演算法及架構設計

Abstract

超音波影像（ultrasonic imaging）系統可以提供生理組織，血流等診療資訊。相比於其他醫用影像系統如Ｘ光、電腦斷層掃描和核磁共振成像等，超音波影像系統具有低成本、非侵入性（non-invasiveness）、無放射性、高顯像速率（Frame Rate）以及可攜式等特性，所以是一種被廣泛使用的造影方式。隨著VLSI及製程技術的演進，可攜式超音波影像系統在近十幾年來逐漸成為重要的趨勢，提升病患就診方便性也降低病患的痛苦；此外，可攜式系統在緊急救護上更扮演著不可或缺的重要角色，能大幅度的提升存活率。 目前，超音波系統中最主要的應用為波束成型(beamforming)技術，用以產生B模式(B mode)成像以表達接受訊號的功率強度而形成生理組織影像。由於超音波的自然特性，波束成像對於接收訊號而言，相當重要。其中，最重要且也最古老的波束成像方法即是利用延遲和加總(delay-and-sum)將收集到的訊號作整理以及凝聚。由於傳統的波束成像方法產生的影像其影像解析度並不理想，因此可適性最小變異無失真響應(minimum variance distortionless response, MVDR/MV) 演算法近年來被廣泛的應用在醫療領域上來提升波束成像的影像品質。但是可適性權重演算法的高計算複雜度使得其演算法很難用在超音波系統上，所以如何設計出低複雜度的可適性權重演算法變成一個重要的研究趨勢。 波束成像方法可概括分類為真實孔徑（Real Aperture）以及合成孔徑（Synthetic Aperture）。相較起真實孔徑，合成孔徑擁有較低的複雜度和成本，因此適用於可攜式高速超音波成像系統。雖然合成孔徑擁有上述特性，但由於合成孔徑的輸出影像是疊加處理多次探頭激發的影像資料而成，所以若是在合成孔徑成像的過程中目標物體有位移現象，會在影像資料間產生非同調現象（Inhomogeneous）。在正常診療情況下，受測者可藉由短時間閉氣而減少位移現象，或是經由現存的離線（Off-line）系統位移補償演算法來進行修正。然而當應用在救護車，戰場或是孩童身上時，位移現象(motion artifact)將會難以避免而影響影像品質甚鉅。 本論文的研究主題主要有三個部分，第一部分首先針對用於可適性權重的演算法提出改進，利用超音波信號在計算MV時所需的樣本共變異數矩陣(sample covariance matrix, SCM)具有空間穩定(spatial stationarity)的特性，提出一個近似樣本共變異數矩陣(Approximate SCM)。接著再使用矩陣求逆引理(matrix inverse lemma)來推導出一個不用直接使用反矩陣的MV公式來降低傳統計算MV的計算量。相較於目前文獻可維持相近的影像品質卻可以大幅降低MV計算量從O(L3)到O(L)。在本論文的第二部分中，對於合成孔徑成像有非同調現象的產生提出一個即時二維位移補償演算法。此演算法運用合成發射孔徑的空間特性來估計位移量，大幅降低其運算複雜度，並讓系統產生高品質的影像。本論文的第三部分以硬體實現二維位移補償演算法的演算法架構，並提出一個適用於線性陣列(linear array)的低複雜度延遲加總 (delay-and-sum)架構。所提出的波束成像引擎可達到即時成像，畫面更新率為42 fps (frames per second)。

Ultrasonic imaging system provide diagnostic information like tissue images and blood velocity. Compared to other medical imaging systems such as X-ray, computed tomography and magnetic resonance imaging, ultrasonic imaging system has features like non-invasive, non-radioactive, low cost, high frame rate and portable. With the progress of VLSI technique, portable ultrasound imaging systems have become a trend for tens of years. They are easily carried to where patients are, which largely decrease the inconvenience and pain for patients. Moreover, portable systems become more indispensable in supporting immediate diagnosis for emergency rescue to increases the survival rate. Currently, beamforming is the main applications in ultrasound system to generate the B mode imaging which expresses the power of the received echo. Since the natural characteristics of the propagation wave, beamforming plays a vital role to focus the received echoes. The oldest but the most important beamforming method is to delay and sum for received echoes alignment. Since DAS beamformer has a wide main lobe and higher sidelobe levels, adaptive minimum variance directionless response (MVDR)-based beamformers, or also known as MV-based beamformers, are proposed to enhance the image quality of ultrasound imaging. However, it is not suitable for MV beamformer to implement in ultrasound imaging system due to its high computational complexity. Therefore, how to design a low-complexity MV beamformer become an important research issue. Beamforming could be roughly classified into real aperture and synthetic aperture. Compared to real aperture, synthetic aperture is more suitable in high frame rate ultrasound imaging system due to lower complexity and cost. However, the output image of synthetic aperture is formed with accumulating series of low resolution images (LRIs) which is obtained from multiple probes, therefore it is susceptible to motion, which will cause the inhomogeneous LRIs. In the normal clinics, motion can be reduced by holding breath or compensated by existing off-line algorithm. However, when the system is used on the ambulance, battlefield, or children, motion is difficult to avoid and it will degrade the image quality severely. There are three main topics in this work. In the first part, the low-complexity MV beamformer is proposed to reduce the computational complexity of traditional MV beamformer. We applied Approximate Sample Covariance Matrix (ASCM) and the matrix inversion lemma to derive a new formula to perform MV beamforming without computing matrix inversion. Compared with traditional MV beamformer, the proposed method reduce the computational complexity from O(L3) to O(L) with similar image quality. In the second part of this work, a low-complexity two-dimensional motion compensation algorithm is proposed. The proposed method can reduce computational complexity significantly by geometry characteristics of synthetic transmit aperture, and generate high quality images. In the third part, a low-complexity linear array delay-and-sum architecture is proposed. The hardware of the proposed algorithms is also implemented in 90 nm technology. The implementation results of beamforming engine have 42 fps (frames per second).