Research Based On FC-DenseNets Algorithm And Its Application In The Diagnosis Of Spinal Canal

In recent years,spine diseases are one of the high-incidence diseases that are currently attracting social attention.The fast-paced life and high work pressure have led to irregular living habits,making spinal stenosis and intervertebral disc herniated diseases more common.The segmentation of the spinal canal area and the boundary of the vertebral body has become an important criterion for the diagnosis of such diseases.At present,doctors only rely on clinical experience to manually outline the target area,and use fixed-point measurement to detect the anteroposterior diameter distance.For some medical image sequences with fuzzy boundaries and overlapping tissues,accurate segmentation will cost a lot of manpower and material resources.Therefore,the design of spinal canal segmentation algorithm can provide doctors with effective auxiliary diagnosis,which has important clinical value.With the development of medical imaging methods,CT and MRI(Magnetic Resonance Images)images are commonly used as the basis for diagnosis.On the basis of getting in touch with the hospital,this thesis obtains a complete medical CT image as a data set,and uses a positioning algorithm based on YOLOv3 and an improved FC-DenseNets algorithm to segment the CT data set.Compared with the previous segmentation effects of traditional segmentation algorithms,deep convolutional neural networks can quickly deal with a large number of situations with complex features,with higher accuracy and robustness,which can save a lot of medical resources.Based on the above considerations,the work and innovations done in this thesis are as follows:(1)In view of the characteristics of traditional semantic segmentation networks,the optimized FC-DenseNets network structure is used in conjunction with the target detection network YOLOv3 to deeply extract the features of the target area,reducing the interference of non-main features on the segmentation process.Make its segmentation measurement index Dice reach 96.31%,which is better than the accuracy of other algorithms.(2)Aiming at the lack of samples in the target detection process,after data enhancement,the YOLOv3 transfer learning method can be used to achieve the training purpose faster.In the whole training process,compared with other models,the adaptive learning rate attenuation strategy is adopted to optimize the training parameters and improve the network performance.(3)Semantic segmentation models are usually built on a large number of training samples.Not only is the amount of medical CT image data far less than the above-mentioned order of magnitude,but also requires higher segmentation accuracy.The lightweight improvement of the FC-DenseNets network structure makes it more suitable for this medical CT data set and reduces the calculation of redundant information in the image.(4)Combine the Dice index with the Jaccard index to obtain a new evaluation standard(JD index).Using the JD index to define the loss function for training,the obtained JD is 94.83%,and the Jaccard is 93.87%.After experimental comparison,they are better than other models,and the trained model has stronger robustness.