Research On Spinal MR Image Segmentation Algorithm Based On Deep Learning

Disc herniation,spinal canal stenosis and intervertebral disc degeneration are common spinal diseases in real life.Magnetic Resonance(MR)of the spine is a multi-directional imaging method that can reflect the true state of the spine.Doctors can more accurately diagnose and analyze lesions by segmenting MR images.However,the shape difference of spine images is small,multiple scales alternate,and the cost of image annotation is high,which makes the accurate segmentation of spine images very challenging.In this thesis,spine image segmentation technology is studied based on deep learning method.The main work is as follows:(1)A multi-scale information aggregation network model(MIAU-Net)for spine images is proposed.This model is an improved multi-scale semantic segmentation model based on the classical convolutional neural network U-Net.It achieves better segmentation performance through redesigning multi-scale information fusion convolution module and jump connection module.Specifically,the proposed multi-scale information aggregation convolution module is used to obtain receptive fields of different sizes and capture multi-scale feature information.At the same time,the improved jumping connection module can alleviate the problem of gradient disappearance.The model was evaluated using a publicly available spine data set.The model intersection over Union(IoU)can reach 82.11%.Compared with other advanced network models,the effectiveness of this method is verified.(2)Based on the MIAU-Net module,a MIAUNet++ module based on UNet++ is proposed.It solved the semantic gap problem of the skip connection module.The model improves the convolutional structure of the skip connection module into residual convolution while using dense skip connections.The residual convolution is able to reduce the semantic differences between layers by adding branches.The semantic and location information of the feature map can be better preserved and more accurate segmentation results can be obtained.Meanwhile,the scSE attention module is added to enhance more valuable features and weaken the weights of unimportant features after each set of convolution operations in the global.It is demonstrated that the IoU of the improved network model can reach 83.25%,which improves the accuracy of spine MRI image segmentation.(3)A dual channel and spatial attention network DAUNet++ based on UNet++for spine image segmentation is proposed in order to reduce the model complexity and prevent the model overfitting problem.The coding module is constructed by using residual convolution blocks thus mitigating the gradient disappearance.In addition,two novel dual channel and spatial attention modules are proposed,which can emphasize feature-rich regions,enhance useful information,and improve the segmentation performance of the network by performing feature calibration in both spatial and channel dimensions.It is experimentally demonstrated that the improved DAUNet++ model achieves higher segmentation accuracy by improving the IoU of segmentation results by about 2% compared to the original UNet++ model with relatively small number of parameters.