Tracking of Microbubbles with a Recurrent Neural Network for Super-Resolution Imaging

For the clinical application of ultrasound localization microscopy, the complexity of the microvasculature in combination with low frame rates and high microbubble concentrations poses problems for the tracking algorithms that rely on linear motion models. Therefore, we implemented a multitask-learning neural network to learn the nonlinear motion patterns of microbubbles flowing through vessels. With the trained network the next positions of a track can be predicted and the likelihood of candidate detections to continue tracks can be estimated. In two different simulation scenarios, we show that the prediction of the next position is more accurate for the trained recurrent neural network compared to a linear Kalman filter especially for small curvature radii of the vessels (RMSE between ground-truth positions and predicted positions 20.8 μm vs. 31.8 μm). In case of large and changing curvature radii, the neural network and the Kalman filter perform comparably well (RMSE of 25.2 μm vs. 26.7 μm). Furthermore, an exemplary likelihood of candidate positions is shown. Generally, the performance of neural networks depends on the training data that should represent the later applications as well as possible. Therefore, more realistic vessel models should be developed in a next step. Furthermore, the motion prediction and likelihood estimation are to be integrated into a tracking algorithm.