Figure 4
Segmentation of F-actin rings and longitudinal fibers using a fully convolutional neural network. (a) Architecture of the fully convolutional network (FCN) (\(\mathrm {h}_{\rm {d}}\)), which is a modified 2D U-Net (see Materials and Methods for specific implementation details, Conv: Convolution, Batchnorm: Batch Normalization, ReLU: Rectified Linear Unit). \(\mathrm {h}_{\rm {d}}\) is trained with images labeled for F-actin rings (green) and fibers (magenta). It generates scores between 0 and 1 for each pixel to create prediction maps for both structures. Independent thresholds are applied for rings (0.25) and fibers (0.4) to obtain two segmentation maps (see Materials and Methods and Supplementary Fig. 7). (b) Comparison between the labeling of an expert (middle) and the corresponding FCN segmented image (right) on a representative image from the testing dataset. MAP2 (yellow) and phosphorylated neurofilaments (cyan) immunostaining and corresponding confocal images are used to identify dendrites and axons, respectively. Quantification of F-actin rings and fibers was performed within a dendritic mask generated from the MAP2 channel (white line, right). (c) Representative input image analyzed with the FCN. The segmented area for F-actin rings (green) and fibers (magenta) is calculated inside the dendritic mask (white line) (right image) for each image. Scale bars (a, b) 1 \(\upmu\)m, (c) 2 \(\upmu\)m. For the raw images without overlay see Supplementary Fig. 5.
