Automatic Plane Adjustment of Orthopedic Intraoperative Flat Panel Detector CT-Volumes

Abstract

Purpose 3D acquisitions are often acquired to assess the result in orthopedic trauma surgery. With a mobile C-Arm system, these acquisitions can be performed intra-operatively. That reduces the number of required revision surgeries. However, due to the operation room setup, the acquisitions typically cannot be performed such that the acquired volumes are aligned to the anatomical regions. Thus, the multiplanar reconstructed (MPR) planes need to be adjusted manually during the review of the volume. In this paper, we present a detailed study of multi-task learning (MTL) regression networks to estimate the parameters of the MPR planes.Approach First, various mathematical descriptions for rotation, including Euler angle, quaternion, and matrix representation, are revised. Then, three different MTL network architectures based on the PoseNet are compared with a single task learning network.Results Using a matrix description rather than the Euler angle description, the accuracy of the regressed normals improves from 7.7 • to 7.3 • in the mean value for single anatomies. The multi-head approach improves the regression of the plane position from 7.4mm to 6.1mm, while the orientation does not benefit from this approach.Conclusions The results show that a multi-head approach can lead to slightly better results than the individual tasks networks. The most important benefit of the MTL approach is that it is a single network for standard plane regression for all body regions with a reduced number of stored parameters.

Computer-assisted contralateral side comparison of the ankle joint using flat panel technology

“…Applying the pipeline to images with two ankles reaches similar accuracy even though the networks were solely trained on single ankles. The seg-mentation approach outperformed the regression method by Martin Vicario et al [11] and seems to be more suitable for the pipeline. The translational discrepancy of that approach might be explained by the downsampling to 72 3 voxels.…”

“…Their experiments on 72 fetal ultrasound images yielded a positional error of 3.83 mm and 3.80 mm and an angular error of 12.7 • and 12.6 • . Most related to the proposed work, Martin Vicario et al [11] use a CNN to directly regress coordinate system parameters for each of the three viewing planes of the ankle and calcaneotalar joint. They defined the coordinate system for each plane with a 6D representation.…”

“…The accuracy of the plane estimation was assessed in a five-fold cross validation by setting the crop size to 128 and the mask size radius to 25. Martin Vicario et al [11] evaluated their plane estimation approach based on CNN regression on the same unilateral ankle data so that their results can be compared directly to the results of the proposed approach. In both approaches, the images were downsampled and normalized differently.…”

“…a Shows the segmentation of an example with a low positional deviation whereas b shows the segmentation of an outlier example with the highest positional deviation Table 1 Plane estimation: Accuracy comparison (median±σ ) of the proposed approach of plane estimation via an auxiliary segmentation task to direct CNN regression (results were taken from [11]) on 218 unilateral ankle data. The lower table shows the overall system performance (plane estimation with prior localization) on three clinical cases and four cadaver specimen with bilateral ankle data.…”

Abstract: Automatic Plane Adjustment in Surgical Cone Beam CT-volumes

“…To this end, this study investigates a fully-automatic solution to directly regress the standard plane parameters from a CBCT volume of the calcaneus and ankle regions. A PoseNet convolutional neural network (CNN) is adapted and trained, comparing a 6D-, Euler angle-and quaternionbased approach to represent the plane rotation [1]. In addition, a cost function is optimized to incorporate orientation constraints.…”