Trajectory estimation of ultrasound images based on convolutional neural network. (September 2022)

• Record Type: Journal Article
• Authors: Mikaeili, Mahsa; Bilge, Hasan Şakir
• Is Part Of: Biomedical signal processing and control. Volume 78(2022)
• Publication Date: 2022-09
• Subjects: Ultrasound imaging -- Deep neural network -- Position tracking; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2022.103965
• Languages: English
• ISSNs: 1746-8094

Abstract

Abstract: Estimating ultrasound probe position and consequently its image position is challenging in image registration and reconstruction concept. On the other hand, in the last few years, the convolutional neural network has had a huge impact on the image processing concept. In this study, we proposed a new network modality. The proposed model attempts to employ the benefits of both the densely connected network and FlowNet for estimating the position information of the ultrasound images. Furthermore, for reducing implementation costs, the inertial measurement unit was utilized. Definition of the images to the networks' input relies on stack manner. Evaluation of the network's results was completed in three stages: in the first stage, the comparison is made between proposed network performance with two different image sequences, whereby the proposed model has better performance with a stack of three sequences; in the second stage, the network performance was compared with the conventional method, whereby results indicate better performance, especially in rotation angels; and finally, in the third stage, we attempted to answer how is that the network performance if instead of inertial measurement unit, transformation matrix computed with conventional feature extracting methods. According to the acquired results—utilizing conventional methods with three and five sequence networks performance reduces the amount of absolute mean square error in comparison to stage two results. Especially the amount of this reduction is significant in Euler angels' estimation. However, the network has better performance while the transformation matrix is computed with IMU's information.