卷積類神經網路應用於脊椎之影像校準

Abstract

影像校準的技術對於電腦輔助手術為不可或缺的一環，如何將術前到術中的影像進行校準，成為了醫生在手術定位中不可或缺的資訊。影像校準的困難度在於將不同空間場域、不同解析度甚至是不同類型的影像完成對位。 本研究將影像校準應用於脊柱手術，透過術前電腦斷層影像(Computed Tomography : CT)與術中C型臂X光透視機(C-arm)影像。2D-3D影像校準的流程會涉及C-arm初始姿態定位與影像校準。而影像校準包括了數位式影像重建X光攝影(Digital Reconstructed Radiography : DRR)、相似度比對、最佳化收斂步驟。 本研究第一部分會使用校正版做姿態估測，再對CT脊椎做影像校準。在姿態估測AP面與CRA45面投影誤差與3D點回推誤差都可以達到1mm之內，而lateral面則誤差較大。因此使用AP面與CRA45面接續進行影像校準，誤差為0.5896 mm。最後使用Cone-beam CT做驗證，誤差為1.2976 mm，可落在脊椎手術的安全範圍內(2mm內)。 本研究第二部分為使用卷積類神經網路(Convolutional Neural Network : CNN)對CT做影像校準並建立出C-arm的姿態估測模型。本研究使用DRR影像驗證該想法是否可行，並探討如何透過修改CNN模型使估測結果能更精準。在脊椎影像校準的3D點回推誤差為0.3703mm，並於後續對校正版做影像校準，誤差為0.4988 mm。而最後模擬CT做驗證，整體誤差為0.446 mm，可推論該想法架構是可行的。

Image registration is an important part of computer-assisted surgery. Register the images from preoperative and intraoperative has become an important information for doctors, which effect the total accuracy of the surgery. The difficulty of image registration lies in the alignment of different spatial fields, different resolutions and even different types of images. In this study, image registration was applied to spinal surgery, register the preoperative CT image and intraoperative C-arm image. The process will involve C-arm initial pose estimation, DRR projection, similarity comparison, optimization. In the first part of this study, use calibrator marker to estimate the C-arm pose first, and then use CT spine image to do the image registration. In the pose estimation part, the projection error of AP, CRA45 and the 3D estimation error can reach within 1mm.While the error of lateral view is larger, so we use the AP and CRA45 view to do the image registration, the error could reach to 0.5896 mm. Finally, we use the cone-beam CT to verify, the error is 1.2976 mm, which can fall within the safe range of spine surgery (2mm). Second part of the research, use the Convolutional Neural Network (CNN) to register the CT spine image and establish the C-arm pose estimation model. This study uses DRR image to verify whether the concept is feasible, and discuss about how to modify the CNN model to make the estimation results more accurate. In the CNN spine image registration process, 3D estimation error is 0.3703mm. In the following step, use the calibrator marker to do the image registration, the error is 0.4988 mm. In the final simulation verification,the error could reach to 0.4459 mm. It can be concluded that the idea of this architecture is feasible.