立體視覺導引之游離皮瓣機械手臂追蹤系統

Abstract

對於因為癌症手術而造成大面積周圍組織損傷的病患，游離皮瓣重建手術能提供針對患部功能性及外觀性良好的修復，也能減少皮瓣供給部位的術後副作用。雖然手術成功率已能達到95%，但皮瓣區域仍有機會在術後出現血液循環障礙，嚴重的話可能導致皮瓣區域壞死，若及早發現皮瓣血液灌流異常並實行再探查手術，皮瓣挽救率將能達到90%以上。若能提供醫護人員及病患低成本、即時、非侵入式及非接觸式等特性之監測方式，將能有效減輕人力負擔，且提高整體醫療品質。 為了提供皮瓣病患術後更好的監測及照護方式，本研究團隊開發一套自動化即時監控皮瓣區域血流狀況之系統，結合皮瓣區域追蹤、對位以及皮瓣熱影像及可見光影像之資訊分析，達到術後皮瓣即時監測目的。本研究論文將著重在於系統中皮瓣區域自動化追蹤以及對位演算法部分。為了測試系統臨床使用情形以及適用性，本研究與台北榮民總醫院整形外科彭成康醫師合作，並於研究期間同時進行動物實驗、臨床皮瓣手術病患收案、以及開發追蹤及對位系統演算法。本研究中動物實驗、人體試驗及其相關資料應用皆通過台北榮總審查委員會同意。 為了自動追蹤病患皮瓣區域的位置，持續擷取影像並進行對位，基於移動自由度、安全性與簡易操作性，研究中硬體設備採用達明機器人公司所開發的TM5-900 機械手臂及其內嵌之可見光相機，進行追蹤演算法開發。追蹤演算法將以光流法持續追蹤皮瓣移動情形，在偵測到皮瓣區域大幅移動之後，利用機械手臂控制結合立體視覺演算法，偵測皮瓣特徵點移動方向及其於空間中之座標點位置，並移動至適當拍攝位置。且為了達成多時間點間的影像對位並轉換皮瓣區域資訊以進行後續血管阻塞偵測，本研究發展一套多時間點皮瓣區域對位演算法，於初始影像半自動圈選皮瓣區域及邊緣之後，利用SIFT演算法擷取皮瓣區域之特徵點，作為影像形變以及對位之參考資訊，再以CPD計算不同時間點影像特徵點群之間的形變模型，並以TPS轉換皮瓣區域至新的影像位置，最後再進行對位演算法之最佳化。 結果顯示以立體視覺導引機械手臂追蹤之演算法，能夠即時偵測皮瓣區域移動，並藉由立體視覺所計算出之位置，保持相同觀察距離並移動至新的觀測位置。同時藉由多時間點皮瓣區域對位演算法的計算，能讓皮瓣區域的資訊完整的轉換到後續時間點的影像上。將本論文所開發之對位演算法應用在拍攝過程中有大幅移動的實驗豬隻連續影像中，並利用Dice Similarity Coefficient、Hausdorff Distance以及平均邊界距離，來衡量連續150張影像之對位結果與相對應半自動圈選區域之間的表現差異，能得到150張影像的Dice Similarity Coefficient平均為96.1% ± 1%，Hausdorff Distance平均為64.92 ± 13.04像素，平均邊界距離之平均值則為20.56 ± 5.36像素，且於對位過程中並沒有因為豬隻掙扎或移動而大幅影響對位結果。將演算法應用於實驗病患案例一對位結果之Dice Similarity Coefficient平均則為95.9% ± 1%，Hausdorff Distance平均為32.87 ± 8.80像素，平均邊界距離之平均值則為9.58 ± 2.65像素，表示所開發之對位演算法能用於不同情況下之皮瓣區域對位，且能夠克服因為病患移動或環境改變所造成的困難，達到多時間點對位之目的。

Microvascular free flap surgery has been a reliable and important reconstruction method for patients who are suffered from complicated large area tissue injure from cancer surgery. Although the success rate of free flap surgery has reached to 95%, flap pedicle thrombosis still has the chance to occur after the surgery, which causes irreversible damage on free flap. Therefore, to minimize flap losses, careful real-time postoperative monitoring and analyzing are generally recommended. Nevertheless, postoperative monitoring is a demanding task for nursing staffs. To aid the monitoring of free flap thrombosis, less the demand of medical crews, and minimize the flap losses, an effective free flap situation monitoring method with low-cost, non-invasive and contactless characteristics is highly desirable. In order to provide better care and assist the real-time monitoring of postoperative free flap surgery patients, we developed an automatic free flap region monitoring system, which consolidated the free flap tracking and registration algorithms, and the analysis algorithms for free flap temperature and color performance. This study focused on the automatic free flap tracking and registration system. To examine the applicability in clinical environments, this research cooperated with Dr. Cherng-Kang Perng and Taipei Veterans General Hospital, and proceeded the animal experimentation, clinical trials, and the system development at the same time. All images used and experiments in this study were approved by the institutional review board of Taipei Veterans General Hospital. Based on the characteristics of moving freedom, safety, and simple operation, the tracking system adopted TM5-900 robotic arm developed by Techman Robot Inc. and a camera embedded on the arm. To accomplish the tracking objective, we incorporated the stereovision, optical flow method and the robotic arm. Once detecting the moving of free flap region, the algorithm measured three-dimensional coordinates of points on the flap edges, and calculated the distance from the plane they formed to the camera. And then the position the robotic arm should reach could be retrieved. To register longitudinal images and transform the free flap region information between them, the free flap registration algorithm extracted the feature points around the delineated free flap edges by SIFT algorithm, which provided the feature information for tracking and registration algorithm. After capturing the next image, the free flap region edges from the formal image would be transferred by Coherent Point Drift and Thin Plate Spline algorithm to the image captured at the new position, and followed by the optimization algorithm. The results of the research have proved that the system can keep in sight of the moving path of the free flap region. Meanwhile, registered by the proposed longitudinal registration algorithm by this study, the average Dice Similarity Coefficient was 96.1% ± 1%, average Hausdorff Distance was 64.92 ± 13.04 pixels, and average edge distance between the results and delineated contours was 20.56 ± 5.36 pixels in the images captured from animal experimentation. On the other hand, applied the registration algorithm in the images acquired from the clinical trial, the average Dice Similarity Coefficient was 95.9% ± 1%, average Hausdorff Distance was 32.87 ± 8.80 pixels, and average edge distance between the results and delineated contours was 9.58 ± 2.65 pixels. Based on the performance of registration, the free flap region registration algorithm demonstrated its effectiveness under various situation, and the capability to overcome the difficulty caused by the moving of patients and environmental variation.