Pattern Recognition And Automatic Tracking Of Myotendinous Junction In Muscular Ultrasound Images

Muscle-tendon unit(MTU)that consists of muscle and tendon plays an important role in human movement.In recent years,the displacement of the myotendinous junction(MTJ)obtained by ultrasound imaging is frequently used to quantify the changes of fascicles and tendinous tissues,thus helping to understand the mechanics and pathological conditions of the MTU during motion.However,the lack of a robust automatic detection method limits its application in human movement analysis.Therefore,based on the feature of MTJ structure on the muscular ultrasound image,the topic of automatic MTJ measuring and tracking in muscular ultrasound image is discovered from the perspectives of motion estimation,structure recognition and deep leaning respectively,thus providing effective methods for the quantitative analysis of MTU during motion.First,based on the prior knowledge of tendon,the phase congruency(PC)is combined with the localized Radon transform(LRT)to gain an effective MTJ region,on which the Lucas-Kanade(LK)optical flow method is applied to track the points artificially labelled on the tendinous tissues,thus avoiding the interference brought by the non-tendinous tissues and tracking the MTJ among consecutive muscular ultrasound images effectively.The automatic method is evaluated mainly by using the coefficient of multiple correlation(CMC).As shown,the MTJ tracking results obtained by proposed method are consistent to the manual measurements(CMC = 0.97 ± 0.02),gained on the ultrasound image dataset recording gastrocnemius muscle movements of 8 healthy subjects.Furthermore,based on the same dataset,a superiority is shown when compared to the classical LK optical flow method(CMC = 0.79 ± 0.11).Considering the limits brought by artificial labelling of tendinous structure,based on the MTJ structure information,the topic is further discovered and a multi-scale MTJ searching method applied on the corresponding PC maps is proposed,by combing the correlation matrix(CM)and Hessian matrix(HM).As a result,based on the same ultrasound image dataset,this fully automatic method is proved to be more consistent with the manual measurements(CMC = 0.97 ± 0.01)when compared to the last proposed method(CMC = 0.97 ± 0.02).Considering the boom of deep learning in recent years,the topic is preliminarily explored from the perspective of data-driven,and an effective MTJ region is segmented by using a deep convolutional neural network(DCNN).The region is then applied to calculate the displacement of MTJ among consecutive ultrasound images.A good result is gained by implementing the training process on the above ultrasound image dataset,proving the great potential of deep learning for MTJ measuring in ultrasound images.With the focus on the topic of MTJ localization in muscular ultrasound images,the mature methods in the fields of both computer vision and image processing are extensively researched and applied.New methods for the automatic MTJ measurement are finally achieved.Apart from providing assistance for the quantitative analysis of MTU during motion,the newly developed methods are also able to support the theory analysis and algorithm development implemented for the researches about object recognition,segmentation and tracking in muscular ultrasound images.