Table 1. Glossary of key concepts.

Term	Explanation
Architecture	The general layout of the neural network (layer type, sequence, etc.).
Hyperparameter	Design parameters of the neural network that are set externally by the user before training.
Input (data)	The data to be analyzed, e.g. an optical microscopy image of a fluorescently labeled cell.
Output (data)	The parameter to be extracted from the input data, e.g. the outline of the cell or the position of the centroid of the cell.
Layer	An individual layer within a neural network constitutes one hierarchical level when the data is processed by the neural network.
Loss	A metric to determine how strongly the prediction of the neural network differs from the ground truth.
Neural network	A universal function approximator, consisting of connected neurons and other building blocks, loosely inspired by neurons and synapses in the brain.
Node/Neuron	Inspired by biological neurons, a node/neuron in a neural networks is a function that receives a certain number of scalar inputs, processes them, and returns a single scalar as output.
Testing	The final assessment of neural network performance after all hyperparameter optimization etc. is completed and the final training has been completed.
Test set	The dataset that is set aside at the very beginning of the project, which is used for final testing. It is not part of the training set, and, in contrast to the validation set, is also not used for hyperparameter optimization and other tasks like that. This is done to ensure robustness and avoid focusing on the validation set as benchmark.
Training	The process of modifying the parameters (weights and biases) of a neural network in such a way that it yields a result closer to the correct output.
Training set	The dataset that is used to train the neural network.
Validation	The process of accessing the performance of the trained neural network.
Validation set	The dataset that is used to perform the validation. Crucially, this dataset was not part of the training set, i.e. the validation set was not used to train the neural network.