Estimating Rotation Angle and Transformation Matrix Between Consecutive Ultrasound Images Using Deep Learning

Abstract

Image registration plays a crucial role in biomedical imaging, especially in image-guided surgery. Obtaining real-time images with an Ultrasound Imaging System (US) makes it possible to register them with magnetic resonance (MR) or computed tomography (CT) images and increase the accuracy of imageguided surgery. Differences in the resolution and intensity of these images motivated us to register ultrasound images with each other. Ultrasound images suffer from low contrast and resolution in comparison to other image modalities such as MR. By acknowledging the fact that the transformation matrix is the building block of the registration concept. Also, given the success of deep learning in classification, we choose to apply it to identify the angle difference and rotation matrix of three consecutive ultrasound images. This paper attempts to find the Euler angles and rotation matrix of three consecutive ultrasound images by applying a deep learning method. At the end of the study, we attain promising results when our learning rate is 0.00002 and the scaling factor is 64× 32. Furthermore, the comparison of positive and negative angles demonstrates that the overall network performs better in predicting positive angles.