使用數位微波束成形實現超音波陣列訊號的資料壓縮

Abstract

超音波影像系統由於成像所需資料運算量過於龐大，以往普遍建構在硬體電路上利用高度平行化架構達成即時影像處理，然而硬體電路的功能固定，不利於可程式化設計使演算法更新的彈性低、系統開發成本高且週期緩慢。然而現在軟體運算能力進步，以圖形介面處理器(GPU)為基礎的軟體成像系統能有效地降低硬體資源和系統開發成本。但實現軟體成像系統的主要瓶頸之一是大量超音波射頻訊號需要從硬體的記憶體中，藉由硬體傳輸介面傳送到軟體端，若要達到即時成像，傳輸率需要達到每秒數個Giga Bytes，以目標使用的USB 3.0硬體傳輸介面，其最大傳輸速率只支援每秒0.5GB，無法達到即時成像系統的需求。權衡軟體系統的優勢和硬體資源使用量後的妥協方法是，在系統的硬體前端使用資源消耗低的演算法壓縮原始資料量，以最少的硬體成本換取較低的傳輸介面需求，來達成即時軟體基礎成像系統。之前的研究中提出將射頻訊號解調，再利用沃爾什轉換基礎(Walsh transform-based)壓縮方法對振幅訊號做資料壓縮後，可達到整體4~5.6倍的壓縮率。其中也提出輸出影像的品質主要受是相位訊號影響，因此壓縮振幅訊號較為適當。本研究提出以數位微波束成形的方法，進一步壓縮振幅資料，達成更高的壓縮率，以應付前述壓縮方法有時壓縮率不足的情況。數位微波束成形的步驟為，首先將超音波的通道資料的前N個通道資料為一組，再利用前波束轉向方法，即對其內的訊號做適當的時間延遲以增加通道資料的同調性後，將延遲過的訊號加總成單一通道做輸出，其餘通道也使用相同處理，即可降低N倍通道數量，同時能維持一定的影像品質。另外為了改善加總後的振幅訊號誤差影響的影像品質，使用了補償方法來改善振幅誤差。結果顯示以四個振幅通道資料為一群組而加總的微波束成形壓縮演算法使用下，壓縮後的B-mode影像在空間及對比解析度上接近原始影像，而峰值訊噪比在有補償方法時，約保持在50dB以上；另外，對壓縮後的資料使用孔徑信號處理時，相位偏移矯正方法仍能有效矯正相位誤差，而同調性因子旁瓣抑制法也能有效增加輸出的影像對比度；在彩色都卜勒流速估計中可以正確評估血流區域訊號較大的流速，在訊號強度很小的區域則有較多流速誤差。另外，本研究將數位微波束成形結合了沃爾什轉換基礎的壓縮方法，將原本的壓縮率提升27~59%，得到約6.3~7.1倍的總體壓縮率，以USB 3.0每秒0.5GB的傳輸速率為例，可以支援傳輸原始每秒3.15GB以上的資料量。此架構下的數位微波束成形及補償方法在實作中的Virtex-6系列的FPGA上只需要低於5%的資源使用率，並且微波束成形減少了資料通道數，可以稍稍降低GPU的解壓縮時間，因此在系統架構上能有效率地整合，沒有顯著負擔。

Due to the high computational requirements, conventionally real-time ultrasound imaging systems utilize highly parallel hardware architectures, thus resulting in high hardware cost, lack of flexibility on image optimization and algorithm implementation and relatively long system development cycle. On the other hand, a software-based system utilizing highly parallel graphic processing units (GPUs) can alleviate such limitations. However, massive raw data transmission from hardware end to software end, which is up to gigabytes per second, becomes one of the bottlenecks for performing real-time software-based imaging. For example, the popular USB 3.0 can only support data rate up to 0.5GB/s, which cannot support real-time raw data transfer. A feasible solution is to compress raw channel data with low hardware resource requirement on the front end. As previous studies demonstrated, we can get overall compression ratio of 4~5.6 by demodulating the radio frequency data to baseband and applying Walsh transform-based compression methods. However, more data compression is still desired. In this study, we propose the use of micro-beamforming to further compress the amplitude data with following steps: take the first N channels as a group, and delay the channel data based on pre-steering, then sum up the N channels into one single output. The rest of the channels follow the same procedures and the number of output channels can be suppressed by N times. In addition to data compression, we also propose a compensation method to decrease the errors resulting from the micro-beamformed amplitude data. Results show that when a group of 4 channels are used, B-mode images formed by the compressed data have almost the same spatial and contrast resolution as the original ones. Furthermore, the peak signal-to-noise ratio is higher than 50 dB with the application of the compensation method. Moreover, several aperture domain processing algorithms, including phase aberration correction, coherence-based adaptive weighting and color Doppler velocity estimation, were tested with micro-beamforming and reasonable performance is achieved. The proposed method integrates micro-beamforming and the compensation method into the Walsh transform-based architecture, and overall compression ratio was improved by about 27~59%, reaching an overall compression ratio up to 6.3~7.1, which enables real-time data transfer via an USB 3.0 interface. The increased resource utilization is no more than 5% on a Virtex-6 FPGA.