A Deep-Learning Approach for Locating the Intramedullary Nail’s Holes Based on 2d Calibrated Fluoroscopic Images

Abstract

Background: Minimally invasive surgery is widely used for managing fractures. When using the intramedullary nailing for bone fixation, surgeons must insert interlocking screws to prevent rotation of the bone fragment; however, it is difficult to determine the exact screwing position of intramedullary nails inserted into the bone. Conventionally, the distal interlocking nail surgery is performed under intermittent X-ray shooting. Nevertheless, this freehand fluoroscopic technique is technically demanding and time-consuming. Currently, the failure rate of this surgery is more than [Formula: see text], and the location error requires to be controlled within 2[Formula: see text]mm. Purpose: To develop a deep-learning approach for locating the intramedullary nail’s holes based on 2D calibrated fluoroscopic images. Methods: The projection of the hole’s axis is deeply regressed in the first step. Then, the hole’s 3D axis is derived by computing the intersection line of two planes determined by the projection of the axis and the X-ray source, respectively. The benefit of the data-driven manner is that our method can be applied to the arbitrary shape of the hole’s contour. Besides, we extract hole’s contour as the distinctive feature, so as to reduce the space of the training data in a large scale. Results: Our approach is proved to be efficient and easy to be implemented, and it has been compared with traditional location method in phantom experiments. The location accuracy error of the traditional method is [Formula: see text][Formula: see text]mm, [Formula: see text], and the location error of this method is [Formula: see text][Formula: see text]mm, [Formula: see text]. Furthermore, the traditional method takes an average of 10[Formula: see text]min to complete the location, while our method takes only 4[Formula: see text]min. In addition, to further verify the robustness of our method, we carried out a preclinical study involving different neural networks for locating the hole’s axis. Conclusion: Whether in terms of time consumption or accuracy error, our method is significantly better than traditional method, and the efficiency has been significantly improved. Therefore, our method has great clinical value. In addition, our approach has potential advantages over the X-ray guided freehand solution in terms of radiation exposure, and it has tremendous application prospects.