Researchers published a research paper detailing their project, BigNeuron. This initiative seeks to establish standard methods for the accurate and swift automated reconstruction of neurons, using deep learning algorithms.
BigNeuron serves as an open platform for benchmarking automatic neuron tracing. A diverse collection of image volumes was gathered to represent the data obtained in neuroscience laboratories interested in neuron tracing. Gold standard manual annotations were generated for a subset of available imaging datasets, and the tracing quality of 35 automatic tracing algorithms was quantified. The primary objective is to advance the development of tracing algorithms and facilitate generalizable benchmarking.
To achieve this, an interactive web application has been developed, pooling the data along with image quality features. This empowers users and developers to perform principal component analysis, t-distributed stochastic neighbor embedding, correlation and clustering, as well as visualization of imaging and tracing data. Additionally, it allows benchmarking of automatic tracing algorithms within user-defined data subsets.
The data analysis reveals that image quality metrics explain the majority of variance in the data, followed by neuromorphological features related to neuron size. It is observed that diverse algorithms can provide complementary information, contributing to accurate results. A method has been devised to iteratively combine algorithms and generate consensus reconstructions. These consensus trees offer estimates of the neuron structure ground truth, outperforming single algorithms in noisy datasets.
It is crucial to note that specific algorithms may outperform the consensus tree strategy under particular imaging conditions. Finally, to assist users in predicting the most accurate automatic tracing results without the need for manual annotations as a reference, we have implemented support vector machine regression. This powerful technique enables the prediction of reconstruction quality based on image volume and a set of automatic tracings.
