Fig. 1

The workflow of the Open Source DeepNeuron toolbox, which has five deep learning-based modules. Each DeepNeuron module has one or more processing components. Neurite signal detection module (Sect. 2.1) uses convolutional neural networks (CNNs) to do foreground/background classification. Neurite connection module (Sect. 2.2) uses a revised Siamese network [21, 22] to connect neurite structure from detected neuron signals. Smart pruning module (Sect. 2.3) refines a neuron’s morphology by using CNN models to filter out false positives. Manual reconstruction evaluation module (Sect. 2.4) uses the output of CNNs as quality scores to evaluate reconstructions. Finally, dendrites/axons classification module (Sect. 2.5) uses CNNs to perform multiclass classification to differentiate axons, dendrites, and background. Note that all the actual deep learning networks in our five modules can be replaced with other network models or user’s own design