Research On Bone Segmentation And Reconstruction Of Ultrasound Image Based On Deep Learning

Fracture is a common clinical disease.For some serious cases,if effective treatment measures are not taken in time,it may lead to the deterioration of the disease and various complications,such as bedsore,lung infection,venous thrombosis,etc.,which is not conducive to the health of the patients,and even life-threatening.At present,the clinical treatment for fractures is surgical treatment.The accuracy of locating the affected area in fracture reduction surgery directly affects the accuracy of fracture reduction.Therefore,the computer-assisted orthopedic navigation surgery(CAOS)system can be widely used in orthopedics.Currently,the main intraoperative imaging method used in CAOS system is fluoroscopic imaging.However,two-dimensional(2D)fluoroscopy imaging is lack of depth information,therefore doctors need to rely on experience to perform surgical operations;three-dimensional(3D)fluoroscopy imaging can better solve this problem,but its high cost has not been widely used.And whether it is 2D or 3D fluoroscopy imaging will produce a certain amount of ionizing radiation,which is harmful to the health of doctors and patients.To solve the above problems,this article explores the use of ultrasound imaging to provide real-time bone 3D imaging without radiation for CAOS.However,due to the low imaging quality of ultrasound imaging,extracting bones from ultrasound images has become a challenging task,which hinders its development in orthopedic surgery navigation system.Therefore,a key work of this paper is to propose a novel convolutional neural network Bone Net for bone segmentation tasks in ultrasound images.The accuracy of the network has reached a higher level in the same field-under.Under the unified data set,Bone Net’s DICE metrics reaches 93%,and the speed also meets the real-time requirements,for the input size of320*288 pictures,the network’s processing speed reached 0.0044s/piece.On this basis,this article also uses binocular optical positioner and device registration technology to register the auxiliary probe and the ultrasound probe under a unified world coordinate system.In this way,according to the image coordinates of the bone pixels in the ultrasonic image,the spatial coordinates of the bone pixels can be calculated to realize the function of bone positioning.Finally,we performed a surface rendering of the pixel point cloud on the 3D bone surface converted to the world coordinate system to get a more intuitive result.In order to verify the performance of this system,a model experiment was conducted.The system is used to scan,segment and reconstruct the fractured bone model,and then register it with the CT image of the model.The distance error of the corresponding position is calculated,and the average distance is 0.3057 mm.The experimental results show that the system has reached a good level of accuracy,and the processing speed also meets the real-time requirements.