Research On Medical Image Segmentation Method Of Knee Cartilage Based On Deep Learning

Among joint diseases,knee osteoarthritis（OA）often endangers the health of middle-aged and elderly people,mainly manifested as cartilage degeneration.Therefore,it is very important to adopt an accurate and efficient diagnosis method for osteoarthritis.Magnetic resonance imaging of the knee is able to diagnose cartilage degeneration by virtue of its high contrast,high resolution,and non-invasiveness.The accurate segmentation of cartilage has an important impact on doctors’ diagnosis and preoperative planning.Manual cartilage segmentation has disadvantages such as low efficiency and many subjective interventions;traditional image segmentation methods have limitations in cartilage segmentation due to individual differences in cartilage and different pathological forms;The segmentation method based on deep learning performs hierarchical analysis on the target features through the training datasets to achieve the purpose of automatic segmentation and avoid manual intervention.In view of the above situation,in order to obtain high-precision knee cartilage segmentation results,this paper uses the representative deep learning network U-Net as the basic model,and conducts the following research work:（1）In order to solve the problem of low accuracy of cartilage segmentation in the traditional U-Net model due to the single feature extraction structure of the encoder and the simple upsampling structure of the decoder,this paper improves the U-Net encoder and decoder.In order to improve the feature extraction capability of the encoder,this paper combines the characteristics and advantages of the Atrous Spatial Pyramid Pooling module and the Attention Mechanism,and then proposes the AS module,which can extract the deep features of the image while focusing on useful features and suppressing irrelevant features.This further improves the segmentation accuracy of fine cartilage.Then,the AS module is embedded in the encoding end of U-Net,and the number of layers of the encoding end is deepened with reference to the VGG16 network structure,and the AS U-Net model is proposed as the basic model for subsequent work;In order to improve the ability of the AS U-Net decoding end to upsample,to reduce the over-segmentation of cartilage edges,this paper replaces the deconvolution in ASD U-Net with Dense Upsampling Convolution,and further proposes the ASD U-Net network model.When training all the networks in this paper,because the ratio of the knee cartilage area and the background pixels is very different,the Dice Loss function is integrated on the basis of the Cross Entropy Loss function,which improves the training effect.For the new network using the VGG16 structure,the weights obtained by VGG16 training on the PASCAL VOC dataset are used to improve the training effect of the network model and shorten the training time.The ASD U-Net model is tested on the OAI-ZIB dataset.Compared with the baseline model U-Net,the Dice Similarity Coefficient,Precision,Recall,and Hausdorff Distance in the knee cartilage segmentation task are measured.They have improved and reached 90.51%,90.45%,90.58% and5.47 mm respectively.（2）In order to solve the problem that in the above ASD U-Net model,since the output prediction map is only from the last fusion,the model cannot make full use of the feature information of each level,this paper improves the output part of the ASD U-Net model.In this paper,a multi-scale fusion（MF）module is designed,which can integrate multi-level and multi-scale features,and use the residual module to learn the fused features,and then apply the MF module to the output part of the ASD U-Net network to obtain ASDMF U-Net model.It is proved by experiments that compared with the baseline model ASD U-Net,ASDMF U-Net has extended some training time,but the Dice Similarity Coefficient,Precision,Recall,and Hausdorff Distance obtained by segmentation are optimized by 0.93%,1.23%,0.62%,and 0.43 mm,respectively.（3）In order to further optimize the accuracy of ASDMF U-Net upsampling,this paper integrates atrous convolution into Dense Upsampling Convolution to obtain a larger receptive field,and finally proposes the ASDd MF U-Net network model.Experiments show that compared with the baseline model ASDMF U-Net,ASDd MF U-Net has improved all indicators in the segmentation task,and the Dice Similarity Coefficient,Precision,Recall,and Hausdorff Distance are 91.49%,91.51%,91.47%,and4.95 mm,respectively.