DEEP LEARNING METHOD FOR PROBABILISTIC PARTICLE DETECTION AND TRACKING IN FLUORESCENCE MICROSCOPY IMAGES

Analyzing particles in fluorescence microscopy images is important to obtain insights into viral and cellular processes. We introduce a deep learning method for probabilistic detection and tracking of fluorescent particles. For particle detection, we integrate temporal information for regressing a density map and determine sub-pixel particle positions. Detections close to particles are rewarded during training and highly non-linear direct regression of positions is avoided. For tracking, we introduce a fully Bayesian neural network that emulates classical Bayesian filtering and exploits both aleatoric and epistemic uncertainty. The method considers uncertainty information of individual particle detections. Experiments based on the Particle Tracking Challenge data show that the proposed method outperforms previous methods. We also applied the method to fluorescence microscopy images of hepatitis C virus proteins.