Bi-Directional LSTM Recurrent Neural Network for Lumbar Vertebrae Identification in X-Ray Images

Abstract

Duo to the capability of providing online patient pose, mobile C-arm X-ray images play a key role in image-guided minimally invasive spine surgery. However, automatic lumbar vertebrae identification is still a challenge task because of the inherent limitation of mobile C-arm. In order to solve these problems, a novel automatic lumbar vertebrae identification method is proposed, which based on bidirectional long short-term memory (LSTM) recurrent neural network (RNN). First, in order to solve the problem of lumbar vertebrae texture overlapping in X-ray images, the curvature features of 3D lumbar vertebrae model, which are common to the 2D X-ray images, are taken as the input of the model. Second, in order to simulate the multi-view imaging of intraoperative C-arm, the bi-directional recurrent neural network is exploited to learn the correlation of lumbar curvature features at different imaging angles. Finally, in order to avoid of gradient vanishing and error blowing up, the LSTM neuron is applied to replace the notes of bi-directional RNN. Experiment results show that our method identified lumbar vertebrae more accurately than another two methods.