Figure S3.
3D U-net architecture, examples of network behavior during training, and F1 as a metric to compare ground truth annotations with model predictions. (A) Schematic representation of the steps used to train the 3D U-net encoder-decoder neural network. The input for the neural network mode are 3D blocks consisting of a stack of consecutive FIB-SEM images (size 204 × 204 × 204 voxels). The 3D block is subjected in the encoder to three cycles, each consisting of consecutive 3 × 3 × 3 convolutions without padding (purple) and downsampling operators with 2 × 2 × 2 max-pooling (pink). The feature maps from the encoder are then upsampled in the decoder by a factor of 2 (yellow), followed by concatenation with previous feature maps from the downsampling branch that had been exposed to central cropping and finally subjected to consecutive 3 × 3 × 3 convolutions without padding (purple); these steps are repeated three times. The output of the neural network model is a probability map (size 110 × 110 × 110 voxels) of two channels, representing the foreground (FG) and background (BG = 1- FG) probability maps, respectively. Number of featured maps are denoted in red, spatial dimensions at the indicated steps in the neural network in black. Figure designed based on PlotNeuralNet (https://github.com/HarisIqbal88/PlotNeuralNet; adapted from Sheridan et al., 2022). (B–D) Examples of plots showing validation cross entropy loss used to evaluate the predicting behavior of the indicated neural network models for (B) mitochondria, (C) Golgi, or (D) ER periodically obtained during training using FIB-SEM volume data of cells prepared by chemical fixation obtained at ∼5 nm resolution. Cross entropy values were obtained using hold-out ground truth annotations from the training set not used during training or from naïve cells, respectively. The gray area shows the first appearance of relatively stable cross-entropy loss and absence of major spikes obtained by the models during 20,000 consecutive training iterations; these models were then used for prediction. (E) Ground truth annotations consist of true positive (TP) and false negatives (FN) voxels and define the presence or absence of a perfect match with the subcellular structure of interest. The output of the model consists of true (TP) and false positives (FP) voxels, depending on whether the predicted voxels are part or not of the ground truth. F1, as defined in the figure, is used as a practical metric to evaluate the prediction accuracy of the neural network to identify the structure of interest. A perfect model prediction would yield F1 = 1 with FP = 0, FN = 0.

3D U-net architecture, examples of network behavior during training, and F1 as a metric to compare ground truth annotations with model predictions. (A) Schematic representation of the steps used to train the 3D U-net encoder-decoder neural network. The input for the neural network mode are 3D blocks consisting of a stack of consecutive FIB-SEM images (size 204 × 204 × 204 voxels). The 3D block is subjected in the encoder to three cycles, each consisting of consecutive 3 × 3 × 3 convolutions without padding (purple) and downsampling operators with 2 × 2 × 2 max-pooling (pink). The feature maps from the encoder are then upsampled in the decoder by a factor of 2 (yellow), followed by concatenation with previous feature maps from the downsampling branch that had been exposed to central cropping and finally subjected to consecutive 3 × 3 × 3 convolutions without padding (purple); these steps are repeated three times. The output of the neural network model is a probability map (size 110 × 110 × 110 voxels) of two channels, representing the foreground (FG) and background (BG = 1- FG) probability maps, respectively. Number of featured maps are denoted in red, spatial dimensions at the indicated steps in the neural network in black. Figure designed based on PlotNeuralNet (https://github.com/HarisIqbal88/PlotNeuralNet; adapted from Sheridan et al., 2022). (B–D) Examples of plots showing validation cross entropy loss used to evaluate the predicting behavior of the indicated neural network models for (B) mitochondria, (C) Golgi, or (D) ER periodically obtained during training using FIB-SEM volume data of cells prepared by chemical fixation obtained at ∼5 nm resolution. Cross entropy values were obtained using hold-out ground truth annotations from the training set not used during training or from naïve cells, respectively. The gray area shows the first appearance of relatively stable cross-entropy loss and absence of major spikes obtained by the models during 20,000 consecutive training iterations; these models were then used for prediction. (E) Ground truth annotations consist of true positive (TP) and false negatives (FN) voxels and define the presence or absence of a perfect match with the subcellular structure of interest. The output of the model consists of true (TP) and false positives (FP) voxels, depending on whether the predicted voxels are part or not of the ground truth. F1, as defined in the figure, is used as a practical metric to evaluate the prediction accuracy of the neural network to identify the structure of interest. A perfect model prediction would yield F1 = 1 with FP = 0, FN = 0.

This Feature Is Available To Subscribers Only

or Create an Account

Close Modal
Close Modal