Improving Reconstruction Speed of Positron Emission Particle Tracking by Efficient Gradient Calculation

Abstract

The development of molecular imaging algorithms and tools has advanced our understanding of the molecular dynamics in complex systems, such as tracking cells in vivo. One of these advancements, the positron emission particle tracking (PEPT) algorithm, allows particles to be tracked through a positron emission tomography (PET) scanner. The spatiotemporal B-spline reconstruction (SBSR) method of the PEPT algorithm is capable of tracking a single particle, such as a cell using PET with high accuracy. However, its slow computational speed, particularly with large data, results in time-intensive hyperparameter tuning, which is a limitation in real-world applications. This study introduces a novel approach, employing the backpropagation algorithm, commonly used in deep learning, to enhance the efficiency of gradient computation during particle trajectory reconstruction. Comparisons of the computational speed of the previous and current algorithms on a PEPT benchmark dataset show that the novel approach significantly increased the computational speed without compromising the tracking accuracy. Notably, we found that the difference in computation time between the current and previous algorithms increased as the size of the data increased. In conclusion, we have improved the SBSR method by efficiently computing the gradients, making it faster and more efficient. Even with bigger data, our approach keeps up, showing an improvement in computational speed.