Research On Vertebrae CT Image Segmentation Algorithm Based On Deep Learning

In clinical practice,computed tomography（CT）technology has the advantages of high speed and good imaging quality,becoming one of the most commonly used imaging methods for the detection of spinal diseases.Vertebrae segmentation from medical images is critical for spinal disease diagnosis and treatment.Whether in segmentation accuracy or efficiency,traditional segmentation approaches or machine learning methods are hard to meet the demand of clinical application.Deep learning,especially convolutional neural network,has the powerful recognition that makes it able to obtain segmentation results significantly better than other methods.This paper researches the automatic segmentation algorithm of vertebrae in CT image based on deep convolutional neural network AM-UNet.Firstly,a series of pretreatment process is needed,including bilateral filtering and window adjustment.Besides,in order to avoiding over-fit of the subsequent segmentation neural network,four kind of data augmentation is employed.Secondly,a segmentation network AMUNet is proposed,which integrate the multi-level feature enhancement and multi-scale feature extraction.First,the multi-scale feature extraction is mainly composed of two cascaded Res2 net module,which make full use of different combinations of receptive field sizes and significantly improve the network’s perception of the shape and size of the vertebrae.Second,in order to learn essential feature representation and suppress interference information,a feature representation enhancement mechanism,including a channel attention module and a dual attention module,is proposed.Before the skip connection,the channel attention module is used to explore and emphasize the interdependence between channel graphs of low-level features.After the skip connection,The dual attention module is used to enhance features along the location and channel dimensions.The third is to train AM-UNet by means of deep supervision.Finally,this segmentation neural network is applied to vertebrae image segmentation.Evaluated on publicly available dataset of CSI Vertebral Segmentation Challenge,our method achieved mean Dice similarity coefficient of 92.18±0.45%,Intersection over Union of 87.29±0.58% and 95% Hausdorff Distance of 7.7107±0.5958,outperforming other algorithms.The ablation study results show that the Dice coefficient of the multi-scale feature extraction module is 91.37 ± 0.65%,which can help to improve the segmentation accuracy;The 95% HD of multi-level feature enhancement module reached 9.1597 ± 0.7115,which played a greater role in maintaining boundary and suppressing interference;Deep supervision will fully integrate and complement the two.