順行掃描與自動追蹤之超音波系掃描在三維乳房檢查上的應用

Abstract

超音波掃描是一個常用於臨床檢查的醫學影像技術，而手持式掃描是超音波檢查的主要執行方式，但檢查結果卻可能因人而異。為了提供更完整的檢查，紀錄且提供實際掃描之影像及位置，並透過該結果實現掃描後病況分析是一個很好的解決方案，而透過機械輔助超音波掃描或是斑點訊號分析均可提供上述資訊。然而，受限於機構設計，現行的機械輔助超波掃描系統難以依照受試者身形進行順行掃描，又或是無法針對複雜表面進行掃描。再者，實際的超音波斑點訊號容易會因身體內的水分、氣體、散射子在組織中的分布不均勻、及影像接觸的程度而被影響。在本研究針對上述的兩項技術提出改良，並嘗試提出更有效的方法來追蹤超音波檢查中的路徑資訊。我們在研究的第一與第二個部分中首先一個機械輔助掃瞄系統，其目標不僅為達到掃描路徑之記錄追蹤、實現的乳房線性與放射性之順形掃瞄，我們更嘗試實現腋下區域之掃描，而該掃描模式目前難以被類似的系統達到。針對臨床需求與建議，我們設計了線性與放射狀等兩種順形掃描模組，同時掃描的影像間距約為0.76 mm，而兩種掃描均透過實際掃描一乳房訪體來驗證，其中兩種掃描分別可以在約75秒與210秒的時間內針對一面積為100 mm  50 mm的矩形與一半徑為50 mm 的圓形區域完成掃描，而基於所紀錄之掃描資訊，我們可以重建一個3D的超音波影像。本研究的最後一個部分為針對斑點訊號的追蹤方式進行修改，藉此達到改良手持式超音波檢查在路徑追中上的準確度。我們的改進包括：1) 製作一個具有均勻散射子分布的仿體墊片作為可靠的斑點訊號來源，2) 發展一套基於奇異值分析 (singular value decomposition, SVD) 的特徵擷取法，以及 3) 導入雙層類神經網路針對位移種類與位移量進行預測。相較於使用傳統進行預測，其結果誤差遠大於0.02 mm 的結果，實驗顯示使用我們提出的方法，可以再在訓練階段達到高於達 96%的預測精確度，而實際分析位移量在1)沿著Y軸方向位移，以及沿著 2) X/Z 軸進行旋轉移動的估測誤差分別約為0.00047 mm 與0.0038°/0.0018°。同時，該估測結果同樣適用於重建3D的超音波影像。

Ultrasound imaging is widely used in breast screening and often performed by the radiologist in free-hand operations. However, the examining results highly depend on the skills and experiences of operators. For full examinations, the ultrasound B mode images and the corresponding scanning path should be recorded. Utilizing a robotic system to assist scan or displacement tracking based on speckle pattern analysis are two possible solutions to providing the aforementioned information. However, it is difficult to achieve body-shape fitting and complex surface screening with the existing robotic scan system since only fixed path scanning is provided. On the other hand, the unreliable quality of the speckle pattern from soft tissues, which can be affected by tissue inhomogeneities and non-uniformly distributed scatterers in humane body, increases the difficulty of analyzing. To this end, the objectives of this study focused on providing better tracking approaches by improving the existing solutions. For the first and second part of this study, we proposed a robotic ultrasound scanning system which is able to adapt the linear and anti-radial scan path to perform conformal whole breast scan. In addition, the axillary area can also be scanned by the proposed system. A frame space of 0.76 mm can be achieved by the proposed system with both linear and radial conformal scanning. The proposed approaches were investigated using a breast phantom. Results show that 75 and 210 seconds were used to scan a rectangular region of 100 mm  50 mm and a circular region with a radius of 50 mm, respectively. Besides, 3D volumetric images can be reconstructed from the acquired B-mode images and the recorded position information. The third part of this study focused on improving the tracking accuracy in hand-held scan through speckle pattern analysis. The improvements include three parts 1) manufacturing a pad phantom with homogeneous distributed scatterers, 2) introducing an SVD-based method for feature extraction, and 3) using 2-layer neural network predictors for motion prediction. With the conventional approach using the speckle decorrelation curve for estimation, the estimated errors are too large to distinguish the types of motion and amount of displacement. By comparison with the proposed approach, the accuracy of motion prediction can reach 96% with training, and the estimated errors of elevational motion and rotational scan along X/Z axes are 0.00047 mm and 0.0038°/0.0018° respectively.