Siamese Spatial Pyramid Matching Network with Location Prior for Anatomical Landmark Tracking in 3-Dimension Ultrasound Sequence

Abstract

Accurate motion tracking of the liver target is crucial in image-guided intervention therapy. Compared with other imaging modalities, ultrasound is appealing choice as it provides accurate and real-time anatomical information surrounding lesions. Besides, compared with 2-dimensional ultrasound (2DUS) image, 3-dimensional ultrasound (3DUS) image shows the spatial structure and real lesion motion pattern in patient so that it is an ideal choice for image-guided intervention. In this work, we develop Siamese Spatial Pyramid Matching Network (SSPMNet) to track anatomical landmark in 3DUS sequences. SSPMNet mainly consists of two parts, namely feature extraction network and decision network. Feature extraction network with fully convolutional neural (FCN) layers is employed to extract the deep feature in 3DUS image. Spatial Pyramid Pooling (SPP) layer is connected to the end of feature extraction network to generate multiple-level and robust anatomical structure features. In decision network, three fully connected layers are used to compute the similarity between features. Moreover, with the prior knowledge of physical movement, we elaborately design a temporal consistency model to reject outliers in tracking results. Proposed algorithm is evaluated on the Challenge of Liver Ultrasound Tracking (CLUST) across 16 3DUS sequences, yielding \( 1.89\; \pm \;1.14 \) mm mean compared with manual annotations. Moreover, extensive ablation study proves that the leading tracking result can benefit from hierarchical feature extraction by SPP. Besides proposed algorithm is not sensitive to sampled sub-volume size. Therefore, proposed algorithm is potential for accurate anatomical landmark tracking in ultrasound-guided intervention.