Application of deep convolutional and long short-term memory neural networks to red blood cells motion detection and velocity approximation

Abstract

The paper deals with processing data obtained using nailfold high-speed videocapillaroscopy. To detect the red blood cells velocity two approaches are used. The deterministic approach is based on pixel intensities analysis for object detection and calculation of the displacement and velocity of red blood cells in a capillary. The obtained data formulate targets for the second approach. The stochastic approach is based on a sequence of artificial neural networks. The semantic segmentation network UNet is used for capillary detection. Then, the classification network GoogLeNet or ResNet is used as a feature extractor to convert masked video frames to a sequence of feature vectors. And finally, the long short-term memory network is used to approximate the red blood cells velocity. The results demonstrated that the accuracy of the mean velocity approximation in the time range of several seconds is up to 0.96. But the accuracy at each specific time moment is less accurate. So, the proposed algorithm allows the determination of the RBCs mean velocity but it doesn't allow determination of the RBCs pulsations accurate enough.