在三維超音波影像中,行列式二維陣列相較於全採樣二維陣列大幅地降低了所需要的硬體以及計算複雜度。然而,當可控制的傳感器元件從N^2減少至2N以及較長孔徑造成的邊界效應,將導致成像品質受到影響。因此本論文提出從行列式二維陣列的三維影像重建出全採樣二維陣列的全資料集的方法,藉由重建的全資料集不僅可以改善行列式二維陣列的成像品質,也可能適用於其他被全採樣陣列發展的成像方法,預期能達到更廣泛的應用。本論文所提出方法之核心為透過空間濾波來重建全採樣空間資料,為達到較佳之效果,本方法首先使用端對端的深度學習框架,將行列式二維陣列的三維影像增強至全採樣二維陣列的雙向動態聚焦影像,接著應用N^2組K空間濾波器至增強後的影像,以估計出全採樣二維陣列各個發射事件所獲得的低解析度影像,最後藉由波束和分解方法,計算各張低解析度影像對於原始通道資料中每一個取樣點的貢獻量,再將該貢獻量加總以重建出全採樣二維陣列的全資料集。在Field II模擬實驗中,使用128行與128列、11MHz、通道間隔為一倍波長的行列式二維陣列,將線仿體於-6dB與-20dB的側向波束寬度從0.42mm及0.64mm改善至0.35mm及0.57mm,旁瓣等級也從-15.92dB降低至-24.78dB。在囊腫仿體中,將對比與對比雜訊比從5.09dB及2.59dB改善至19.10dB與4.85dB,廣義對比雜訊比也從0.71提高至0.99。提出的方法預期也能拓展至其他不能存取全資料集的超音波系統上,且無關於陣列的發射方式。
In 3-D ultrasound imaging, the row-column addressed (RCA) 2-D array greatly reduces the hardware requirements and computational complexity compared to a fully sampled (FS) 2-D array. However, with a reduced channel count and longer array elements, the overall image quality is degraded, and the edge effect associated with the more extended aperture becomes inevitable. To improve the RCA image quality, we propose reconstructing the channel data of an FS array using the data from an RCA array for 3-D imaging. The core component of the proposed method is spatial filtering. To achieve optimal performance, we first use the end-to-end deep learning framework to enhance the 3-D image of the RCA 2-D array. Second, we apply N^2 sets of k-space filters to the enhanced image to estimate the low-resolution images obtained by the FS 2-D array when only a single element is used on transmit. Finally, using the proposed beamsum decomposition method, we decompose the beamsum data into channel data of all individual elements of the FS 2-D array. With the Field II simulations, an RCA 2-D array of 128 rows, 128 columns, 11MHz center frequency, one wavelength pitch is used. The lateral beamwidths of the wire target at -6dB and -20dB are improved from 0.42mm and 0.64mm to 0.35mm and 0.57mm, respectively, and the sidelobe level is also reduced from -15.92dB to -24.78dB. With the cyst phantom, contrast ratio (CR) and contrast-to-noise ratio (CNR) are improved from 5.09dB and 2.59dB to 19.10dB and 4.85dB, and generalized contrast-to-noise ratio (GCNR) is also improved from 0.71 to 0.99. The proposed method can be extended to other ultrasound imaging schemes when the full dataset is not available (e.g., plane wave imaging).