Detection And Segmentation Of Vertebrae In Medical CT Images Based On Deep Learning

Scoliosis,spinal deformity,and spondylolisthesis are diseases that have a high incidence today.If the symptoms of spinal disease can be found in early time,it will play a vital role in treatment.However,the first for the doctor to diagnose the vertebrae is localization and identification of the vertebrae via medical imaging,such as Computed Tomography and Magnetic Resonance Imaging.Medical CT examination has become an essential auxiliary tool in disease screening,diagnosis and treatment.One of the commonly used methods for disease screening,diagnosis,and treatment planning is medical image examination.However,lots of medical images are produced in daily operations,due to the low image clarity and inconspicuous contrast,it has caused certain difficulties for accurate data through medical images.Thus,it’s essential to develop a vertebrae’s detection and segmentation technology.These years,the development of big data and computing power has promoted the rapid advancement of intelligent medical treatment.Location and detection of vertebrae in medical CT images can be completed by constructing a deep learning-based detection algorithm,it can assist radiologists to increase the sensitivity and work efficiency.Deep learning algorithms can train large amounts of doctor-labeled data,generate effective vertebra detection and segmentation models thant can be applied to areas with underdeveloped medical care.This paper mainly focuses on vertebra detectopm and segmentation via deep learning methods.Our contribution are as follows: Fristly,a method is proposed to convert the traditional convolutional neural network classifier into a fully convolutional network classifier.Secondly,the Faster R-CNN network structure is migrated from natural images to medical image datasets and optimized to improve positioning accuracy and recognition accuracy.Thirdly,by comparing different feature extraction networks and studying the effect of different feature maps on positioning accuracy and recognition accuracy,the detection performance of the entire network is improved.Fourthly,research on multi-scale problems,find the scale and aspect ratio that are most suitable for vertebra detection and positioning.The contributions of this paper are as follows: Firstly,for different feature extraction networks,Res Net work as a feature extraction network can help to improve the detection accuracy of vertebrae.Secondly,for multi-scale problems,since the shape of the vertebrae is regular,a single-scale setting can be used.it helps to reduce the compute cost and reduce the complexity of the model.Thirdly,Non-maximum suppression does not harm the accuracy of detection.Fourthly,Increasing the dataset is good for improving the detection accuracy and generalization.Finally,Res Net50 is a little bit better than Res Net101 in Mask RCNN for vertebrae segmention.