Research On Feature Detection And Stitching Of Musculoskeletal Ultrasound Image

The incidence of musculoskeletal diseases is increasing year by year with the aggravation of aging society.Musculoskeletal diseases great variety,and its pathogenesis is complicated.It has a negative impact on the physical and mental health and daily life of the elderly.Musculoskeletal ultrasound(MSKUS)is an important diagnosis method commonly to assist in the diagnosis and treatment of musculoskeletal diseases.The MSKUS panoramic image can make up for the shortcomings of traditional ultrasound image with a narrow field of view,and comprehensively and clearly display the anatomical images of muscles,joints and other parts.It is helpful for the diagnosis and treatment of musculoskeletal diseases.However,the contrast of the ultrasound image is low,and speckle noise and image artifacts are presented in the MSKUS images.These limitations negatively affect the extraction of the feature points of MSKUS image.And the feature detection of MSKUS image plays an important role in image stitching.How to improve the effectiveness and repeatability of the feature detection of MSKUS image?This issue has been of great research value.Therefore,this study investigates the feature detection algorithm based on the combination of FAST and SIFT,and the deep learning feature detection method based on the combination of improved MagicPoint and CycleGAN,respectively.The feature detection algorithm based on the combination of FAST and SIFT(FAST-SIFT)uses FAST algorithm to detect feature points and determine the coordinates of the feature points,and then applies SIFT descriptors to describe the detected feature points.In this way,the feature points and their corresponding feature point descriptors can be obtained.Then,the Nearest Neighbor Distance Ratio and the Random Sample Consensus algorithm are applied to achieve rough feature matching and fine feature matching,respectively.The projection transformation matrix is taken as the basic model to estimate the optimal deformation matrix between the two images.According to the projection transformation model,multiple images are mapped to the same coordinate system.Finally,the MSKUS panorama is post-processed using the maximum flow minimum cut algorithm and the multiband blending to realize image stitching.The MSKUS panorama with a wide field of view is eventually obtained.The experimental results show that compared with SIFT,SURF,ORB algorithms,FAST-SIFT algorithm is able to extract more uniformly distributed feature points and detect most of the end points of the muscle fibers.The correct rate of feature matching in FAST-SIFT is the highest.Furthermore,the mean value of mutual information and normalized cross-correlation coefficient of FAST-SIFT are higher than those of the other three feature detection algorithms,indicating that the accuracy of image registration is higher.Moreover,the stitched panorama of MSKUS by using the FAST-SIFT algorithm is natural and smooth,without no obvious misalignment and fracture of anatomical structure.It has good image stitching and small visual error.The feature detection algorithm based on the combination of improved MagicPoint and CycleGAN uses CycleGAN network for style transfer,which converts the linear composite image in the synthetic shapes dataset from the virtual image domain to the MSKUS image domain.This method gives the synthetic shapes dataset some certain features of ultrasound images.Then,by increasing the number of convolutional layers of the front-end encoder in MagicPoint network,the MagicPoint network structure is improved.With the aim of extracting feature points with high repeatability and good effectiveness,the fine-tuned synthetic dataset is used to train the improved models,which promotes the models to be better applied to the feature detection of MSKUS images.The experimental results show that among all the improved models,the MagicPoint model having nine convolutional layers and using CycleGAN network for data augmentation has good performance in feature detection.More widely distributed feature points are extracted and more the end points of the muscle fibers are detected.It is superior to the traditional ORB algorithm.However,compared with FAST and SIFT algorithms,the repeatability of feature points of this model is still lower.Although it did not show obvious advantages in this study,this method has potentials to obtain good image features with improvement in the network architecture.