Neurocounter - A deep learning framework for high-fidelity spatial localization of neurons.

BACKGROUND: Many neuroscientific applications require robust and accurate localization of neurons. It is still an unsolved problem because of the enormous variation in intensity, texture, spatial overlap, morphology, and background artifacts. In addition, curating a large dataset containing complete manual annotation of neurons from high-resolution images for training a classifier requires significant time and effort. In this work, we presented Neurocounter, a deep learning network to detect and localize neurons.

NEW METHOD: Neurocounter contains an encoder, a decoder and an attention module. It is trained on images containing incompletely-annotated neurons having highly varied morphology, and control images containing artifacts and background structures. During training, Neurocounter progressively labels the un-annotated neurons in the training data. It detects centers of neuron soma as the output.

RESULTS: Neurocounter's self-learning ability reduces the need for time-intensive complete annotation and ensures high accuracy in the localization of neurons across various brain regions (approximately 94 % F1 score). Comparison with existing methods Neurocounter shows its efficacy over the state of the arts by significantly reducing false-positive detection (by at least 3 %).

CONCLUSIONS: Neurocounter offers precise neuron soma detection in various scenarios, such as with background artifacts, clutter and overlapped cell soma. This tool can be potentially used to reconstruct brain-wide 3D maps of activated neurons from 2D localization of neurons.