電腦斷層攝影三維影像重建演算法評估次肺葉切除術後手術切緣

Abstract

本研究深入探討了肺部切除手術後的影像分析與處理，特別聚焦於次肺葉切除術術後的影像重建及切除邊緣距離量測問題。肺癌在全球及台灣均為主要的癌症類型，與持續吸菸及環境因素有密切關聯。其中，非小細胞肺癌是較常見的類型。針對小型周邊肺腺癌，次肺葉切除術常被選用，但此手術方式比肺葉切除術有更高的局部復發率。 美國國家癌症資訊網（NCCN）建議，次肺葉切除術的切除邊緣距離應大於或等於兩公分或腫瘤大小，以降低手術後的局部復發率。然而，對於這一建議的有效性，學界意見分歧，部分原因在於目前的測量方法依賴於臨床人員的肉眼觀察，存在多種潛在的誤差風險。 為了解決這一問題，本研究透過醫學影像學方法來重新審視切除邊緣距離的問題，並開發了新的流程和技術，以提供更精確的量測方法。我們專注於肺部影像的對位，特別是考慮到肺部的高度形變性。對於肺部影像對位演算法，雖然在放射治療和其他臨床情境中已有應用，但對於肺部切除手術前後影像的對位仍然是一個新領域。本研究的挑戰在於手術後肺部可能出現的諸多變化，如肺部塌陷、手術切口造成的局部形變和整體肺部容積的變化。 本研究主要發展了兩項核心技術：（1）為了獲取手術前後影像中的對應關係，我們開發了自動化子血管樹匹配技術，以確保匹配過程的準確性和可重複性。這項技術分為兩個匹配階段：首先是進行定位，接著計算相似度，並對子血管樹周圍區域的相似度進行分析，以提高匹配的穩定性。在完成初步匹配之後，我們會利用已建立的配對關係來進一步定位和匹配剩餘的子血管樹；（2）針對因手術影像中出現的非連續形變區域，我們開發了一種基於薄板樣條函數的創新方法。這個方法能夠在大部分影像區域維持連續的形變場，同時在手術切除區域附近產生非連續的形變場。它通過選擇具有特定位移趨勢的控制點來進行有效的插值，並使用DBSCAN演算法對這些控制點進行自動分組。 實驗結果顯示，我們的方法在無病灶和手術肺部的子血管樹匹配實驗中表現優異，平均匹配完成率分別為65%和50%。我們開發的基於薄板樣條函數的插值方法能夠更準確地重建手術切除部分，並在血管樹結構上更接近手術前的原始影像。在所有個案中，我們的方法達到了低於 1.50 mm 和 2.50 mm 的平均對位誤差，特別是在手術肺區，與比較方法相比，展現了顯著的差異。最終，我們還實踐了切除邊緣距離的量測及視覺化的呈現。

This study focuses on post-operative lung image analysis, particularly in reconstructing images and measuring resection margins after sublobar resections. Lung cancer, predominantly non-small cell lung cancer (NSCLC) and often linked to persistent smoking, is significant globally and in Taiwan. Sublobar resection, preferred for small peripheral lung adenocarcinomas, tends to have higher local recurrence rates than lobectomy. The National Comprehensive Cancer Network (NCCN) suggests a resection margin in sublobar resections of at least two centimeters or tumor size to reduce recurrence. However, current measurement methods, based on clinical observation, may lead to errors. Our study introduces new techniques for more precise measurements, focusing on lung image registration and dealing with the challenges posed by the complex deformability of the post-operative lung. Our study primarily developed two core technologies: (1) Automated sub-vascular tree matching technology to obtain the corresponding relationships in pre- and post-operative images, ensuring accuracy and repeatability in the matching process. This technology consists of two matching stages: first positioning, then calculating similarity, and analyzing the similarity of areas around the sub-vascular tree to enhance stability. After completing the initial match, we use established pairing relationships to further locate and match the remaining sub-vascular trees; (2) For the discontinuous deformation areas appearing in post-operative images, we developed an innovative method based on thin plate spline functions. This method maintains a continuous deformation field in most image areas while creating discontinuous deformation fields near the surgical resection area. It effectively interpolates by selecting control points with specific displacement trends and uses the DBSCAN algorithm for the automatic clustering of these points. Experimental results were promising, with our methods showing high accuracy in sub-vascular tree matching, and our interpolation method closely reconstructing the original pre-operative vascular structure. We achieved an average target registration error of under 1.50 mm in non-lesion lung areas and under 2.50 mm in post-operative lung areas. Finally, we effectively implemented the measurement and visual presentation of the resection margin distance.