3D Ultrasound Anatomical Landmark Tracking via Fast Normalized Cross Correlation Network and Robust Outliers Rejection

Anonymous

Reviews Visibility: The authors agree that reviews are made publicly visible, if the submission is accepted.

Abstract: In image-guided abdominal radiotherapy, accurate localization of targets can minimize damage to crucial structure. Due to abdominal movements caused by heartbeat and breathing, however, margins would be added around target by surgeons to ensure target can be covered and treated, which would cause additional trauma. To alleviate motion uncertainties and minimize trauma, we propose an accurate algorithm based on a novel deep tracker and outliers rejection method for anatomical landmark tracking in 3D liver ultrasound sequences. Firstly, we couple normalized cross correlation filter (NCC) with fully convolutional network (FCN) and reformulate NCC as a differentiable layer to generate a novel and effective deep tracker. Meanwhile, we introduce the channel attention mechanism to generate the effective features. Second, we derive a fast implementation form of NCC, which enables the algorithm to track in real time. Finally, a robust outliers rejection method with the prior knowledge of physiological movement is employed to further improve the tracking performance. The organizers of the Challenge of Liver Ultrasound Tracking (CLUST) evaluate proposed algorithm, which yields mean and 95%ile tracking error of 1.70 ± 0.98 mm and 3.05 mm, on 22 landmarks across 10 3DUS sequences. Comparison between our and published algorithms shows our algorithm achieves state-of-the-art performance. Moreover, it is proved by ablation study that the leading tracking results significantly benefit from fast NCC and channel attention mechanism.

Keywords: Fast normalized cross correlation filter, Channel attention module, Outliers rejection, Abdominal intervention therapy