Towards Multiple 3D Bone Surface Identification and Reconstruction Using Few 2D X-Ray Images for Intraoperative Applications

Abstract

This article discusses a possible method to use a small number, e.g. 5, of conventional 2D X-ray images to reconstruct multiple 3D bone surfaces intraoperatively. Each bone's edge contours in X-ray images are automatically identified. Sparse 3D landmark points of each bone are automatically reconstructed by pairing the 2D X-ray images. The reconstructed landmark point distribution on a surface is approximately optimal covering main characteristics of the surface. A statistical shape model, dense point distribution model (DPDM), is then used to fit the reconstructed optimal landmarks vertices to reconstruct a full surface of each bone separately. The reconstructed surfaces can then be visualised and manipulated by surgeons or used by surgical robotic systems.

2D–3D reconstruction of distal forearm bone from actual X-ray images of the wrist using convolutional neural networks

“…Most of the conventional SSM-based 2D–3D reconstruction methods incur high computational costs and are operationally complicated, including manual operation during the initial registration. Some methods do not require manual operation, but the accuracy of reconstruction is inferior (3.04 mm) 12 , and the accuracy of reconstruction has not been evaluated 13 . There are some reports of SSM-based 2D–3D reconstruction methods using CNNs 14 , 15 .…”

“…To address this issue, 2D–3D reconstruction has been studied, in which a 3D bone shape is estimated and constructed from 2D X-ray images without CT. Methods for estimating the 3D bone shape underlying X-ray images have been reported in which the 2D projections of the 3D statistical shape model (SSM) 8 – 12 or generic model 13 calculated from the training data set were compared with X-ray images, resulting in a similarity score, and the shape parameters of SSM or generic model were optimized to maximize the similarity. Traditionally, because most of these methods require an initial registration via manual operation, they have limitations in terms of complexity and reproducibility.…”

Estimating and abstracting the 3D structure of bones using neural networks on X-ray (2D) images

“…There are even a few fully automated approaches 9 10 . However, the first fully automated approach that we found requires previous knowledge about the bone geometry in order to identify bone boundaries in the input image 9 , and the second approach requires five X-ray images taken from different angles 10 . A fully automated approach which can easily adapt to new data without previous knowledge about bone geometry or other attributes is therefore desirable.…”

Estimating and abstracting the 3D structure of feline bones using neural networks on X-ray (2D) images

“…Currently, multiple approaches for estimating the 3D structure of bones from their 2D X-ray images exist [1][2][3][4][5][6][7][8] , which help reduce the cost and the radiation-related health risks for the patient, as well as the necessity of anesthesia in most of the animals. There are even a few fully automated approaches for bone shape estimation 9,10 . However, one such approach requires previous knowledge about the bone geometry in order to identify bone boundaries in the input image 9 , and the other requires five X-ray images taken from different angles 10 .…”

“…There are even a few fully automated approaches for bone shape estimation 9,10 . However, one such approach requires previous knowledge about the bone geometry in order to identify bone boundaries in the input image 9 , and the other requires five X-ray images taken from different angles 10 .…”