Discriminative Correlation Filter Network for Robust Landmark Tracking in Ultrasound Guided Intervention

Abstract

Due to uncertainties from breathing and drift in image-guided abdominal intervention, surgeon would add margins around target so that it can be adequately covered and treated. To mitigate the uncertainties and make motion management more effective, we develop a real-time and high accuracy algorithm for anatomical landmark tracking in liver ultrasound sequences. In this paper, we firstly generate a feature extractor based on an end-to-end network by embedding fully convolutional network (FCN) into discriminative correlation filter (DCF). Meanwhile, we reformulate traditional DCF as a differentiable neural layer (DCF layer) to guarantee generated convolutional features are tightly coupled to DCF. Then we train the end-to-end network by encoding millions of ultrasound images and optimizing an elaborate designed loss function. Finally, we utilize the tailored feature extractor and DCF tracker to perform online tracking. Proposed algorithm is evaluated on 85 landmarks across 39 ultrasound sequences by the organizers of the Challenge on Liver Ultrasound Tracking (CLUST), and yielding 1.11 ± 0.91 mm mean and 2.68 mm 95%ile tracking error. The processing speed for per landmark is about 44–47 frames per second with GPU implementation. Extensive evaluation is performed among proposed and published state-of-the-art algorithms, and results show our algorithm significantly reduces maximum error and achieves a leading performance. Ablation study further supports the benefit from the tailored feature extractor. Clinical application analysis proves our tracker can lessen the heavy burden on surgeon and reduce dependence on medical experience.