Knuckle Segmentation Based On U-Net Network Model And Bone Densitometry

Osteoporosis is a systemic bone disease caused by external factors,and even can cause bone cancer in severe cases.The bone mineral density measurement method is used to measure the bone mineral content of the human body,the degree of osteoporosis can be judged according to the bone mineral density value,so as to take preventive measures in advance to prevent the occurrence of fractures.With the development of information technology,computer vision combined with clinical medical diagnosis reduces the burden of medical staff and improves work efficiency.Quantitative CT method(Quantitative Computed Tomography,QCT)combined with image segmentation method is used for medical diagnosis.Using special equipment to obtain the image of the measured area,segment this image,and then cut out the region of interest in the segmentation results,then calculate the bone mineral density,according to the calculation results,determining the possibility of individual osteoporosis.So the segmentation of medical images is particularly important.At the same time,the accuracy of medical image segmentation determines whether QCT method can be widely used or not.UNet is an image segmentation method based on deep learning,which is widely used in the field of medical image segmentation.However,the model will produce a lot of redundancy and slow training speed.Due to the small amount of medical image data,time-consuming manual labeling and high segmentation accuracy requirements,in order to better adapt to the characteristics of medical image segmentation,this paper made a series of improvements based on the U-Net network model,Quantitative CT method was used to measure the bone mineral density of the segmentation results of the improved model.The main improvements are as follows:(1)In order to reduce the noise problem of the image,the wavelet transform is used to enhance the hand image,reducing the detail noise and highlighting the imaging effect of the foreground area.(2)The improved network model includes two parts: the initial segmentation module and the optimization module.The initial segmentation module adds maximum pooling and transposed scaling convolution in the upsampling process,the main feature information is retained while the data dimension is reduced,the computational complexity is reduced,and the computational speed is improved.Using group normalization instead of batch normalization can reduce the impact of high error rate on the accuracy of the network when the batch size is small.The initial segmentation result often loses edge detail information,and the segmentation accuracy is low,so optimization processing is required.The structure of the optimization module is similar to the initial segmentation module,which also includes the encodingdecoding structure,and each input in the decoding process is also cascaded.Fusion context feature information,while making details more complete.(3)In this paper,the mixed loss function is used to optimize the model,and it judges the segmentation effect of the model both image block and pixel levels.The manually annotated dataset in this paper is divided into three data sets: training set,validation set,and test set,select appropriate evaluation indicators,comparing the experimental results of different models,it can be seen that the improved algorithm proposed in this paper has obvious improvement in segmentation accuracy and precision compared with other models.