Human Vertebra Segmentation Based On Deep Learning

Computer tomography(CT)is a commonly used clinical medical imaging technique.CT can well detect vertebral fractures and unstable injuries,so it is widely used in clinical diagnosis.Because the lumbar spine bears most of the body's strength,it is easy to cause compression fractures when it encounters strong external forces.In clinical diagnosis,the initial and mid-term compression fractures are deformed less in vertebral morphology.In most cases,doctors will misdiagnose so as to aggravate the degree of fractures.Therefore,very accurate segmentation results are needed to better assist the diagnosis.However,manual segmentation in the clinic is time-consuming and labor-intensive,and different doctors have different results,and there is no unified segmentation standard.In addition to the above reasons,the repeatability between vertebrae morphology,anatomical variability between normal and pathology,and the difference in image acquisition parameters(including resolution and visible field of view and so on),make the automatic segmentation of a single vertebrae extremely difficult.Big difficulty.Although the existing traditional segmentation algorithms and machine learning algorithms can segment human vertebrae,their accuracy is too low to be used in clinics.Therefore,there is an urgent need for an algorithm that can automatically complete the segmentation of vertebral segments,thereby greatly reducing the workload of clinicians.In response to this problem,this paper proposes a human vertebra multi-class segmentation algorithm,which greatly improves the accuracy of vertebra segmentation,including the following three parts:1.CT image preprocessing.It includes operations such as limiting contrast adaptive histogram equalization,threshold segmentation based on HU value,and image dilation,which improves the quality of the input image and enhances the feature information of the image.2.Multi-scale L1 Loss Generative Adversarial Networks.Taking the GANs as the basic structure,a multi-scale L1 loss function is proposed in its discriminator to supervise the segmentation results;The U-Nnet network widely used in medical image segmentation is used for the generator After verifying its convergence and stability,the multi-scale L1 loss function supervises the segmentation results and contains more feature information of the segmentation results,so the network can effectively segment the vertebrae,And the result is better than the existing image segmentation algorithms.3.Multi-scale adaptive attention network.The network performs more detailed multi-class segmentation on the results of the preliminary segmentation to complete the segmentation of the human lumbar vertebrae L1-L5.In view of the problem that traditional CNNs networks will lose image information as the network deepens,a multi-scale strategy and attention mechanism are proposed to ensure that image will not lose information while the network is deepened,and extract the image feature information that is more relevant to the segmentation task.Therefore,the multi-scale adaptive attention network shows good segmentation performance.