A Deep Learning-Based Approach for Automated Coarse Registration (ACR) of Image-Guided Surgical Navigation

Abstract

Coarse registration is the first step in determining the accuracy of surgical navigation. The purpose of this study was to present an automated coarse registration (ACR) methodology to improve the convenience and accuracy. For this purpose, a deep learning model based on a convolutional neural network was used. The input variable used for learning was virtual patient point-cloud (VPPC) generated based on medical image. Output variables were values of coordinate transformation obtained in the process of sending the VPPC to the surrounding space of a medical image. The ACR model consisted of a step of extracting global features of point-clouds from medical image and patient space and a step of predicting the information of 3-dimensional coordinate transformation through global features. The coefficients of determination that evaluated the similarity between predicted and actual rotation values on the x, y, and z axes were 0.993, 0.989, and 0.990, respectively. The coefficients of determination of the predicted and actual translation values on x, y, and z were 0.993, 0.989, and 0.994, respectively. As a result of coarse registration of three phantoms using the ACR, the registration errors between the patient and the computed tomography point-cloud were 3.813 +/- 0.792, 3.786 +/- 0.734, and 3.653 +/- 0.668 mm, which were significantly improved over the conventional method's registration error (4.671 +/- 0.738, 4.865 +/- 0.776, and 4.670 +/- 0.455 mm). The proposed method can provide convenience in the pre-operative preparation stage by automating coarse registration. It is expected that repeatability and reproducibility can be provided by eliminating random errors that might occur by the operator.