Ultrasound image motion tracking based on multi-level time information and unified feature

Ultrasound tracking is of great significance in assisting ultrasound imaging diagnosis. Due to the high noise, temporal continuity, and periodic motion patterns of two-dimensional ultrasound images, it is challenge for efficient ultrasound motion tracking management. In this work, we propose a motion tracking method based on multi-level temporal information and unified visual feature, allowing for effectively focusing on temporal contextual information and key region features. Firstly, the time adaptive CNN is utilized to dynamically adjust the weights of convolution kernels to understand the temporal variation of ultrasound sequence lesions, and a fixed time knowledge factor is introduced for feature extraction to prevent slow tracking speed caused by long continuous frame sequences. The time knowledge accumulation encoder is used to exploit the time information of the context, capture the dependency between dynamic frames, and use the parallel learnable visual center mechanism to further capture the local corner areas of the input, which can obtain comprehensive and discriminative feature representation at different scales. The overall model employs a top-down global unified regulation mechanism, adjusting the shallow features at the front end using explicit visual center information extracted from the deepest features. In addition, a template update module is introduced to maintain the accuracy and continuity of feature extraction by adaptively update the template image. The model proposed in this article was trained and evaluated using the MICCAI-2015 Challenge dataset. The tracking error is computed by calculating the Euclidean distance between the tracked point location and the ground truth. The results show that the average error was 0.79mm, the maximum error was 1.92mm, and the standard deviation was 0.31mm. and the tracking speed is 80-105 frames per second.