Fast 3D cell tracking with wide-field fluorescence microscopy through deep learning: supplementary material

The schematic architectures of our convolutional neural networks (CNNs) are illustrated in Fig. 1. Both architectures consist of several functional layers: convolution layer, pooling layer, fully connected layer, and output layer. The first 5 stages (denoted as” Layer”) are composed of convolutional layers and pooling layers. In each” Layer”, a convolutional layer (depicted as a green block) with the rectified linear unit (ReLU) extracts feature maps from the previous layer, and then a pooling layer (depicted as a blue block) downsamples the feature maps before the next convolution layer. We increase the number of features learned in each” Layer” by gradually increasing the number of channels for high level feature inference. The size of all the kernels (filters) throughout the convolutional layers is 3× 3. For the final classification, the extracted feature representation of” Layer 5” is fed into the fully connected layers (denoted as” FC1” and” FC2”), and the output layer gives a predicted response. The input of lateral detection CNN is a 128× 128-pixel patch cropped from the raw wide-field fluorescence image and the output is a binary response (y=±1). The positive value means there exist diffraction patterns at the central xy (lateral) position of the input image while the negative value is opposite. Regarding to the axial localization CNN, we focus on the predicted positive samples of lateral detection CNN and the output is a 50-element vector (z1 to z50). Each element, having a probability value between 0 and 1, represents an axial position and the whole vector covers an axial range from the focal plane to 100 microns above with 2-micron spacing. To be …