AutoCellLabeler Live

Background

In freely moving C. elegans imaging, the nervous system deforms continuously and in a non-rigid manner. Thus, determining and maintaining neuron identity over time (and across animals) has been a longstanding and challenging problem. NeuroPAL drastically increased the ease of labeling, expressing combinations of fluorophores to give each neuron a "color." This foundational work enabled highly-trained scientists to label many neurons in a C. elegans brain, although many were often still hard to discriminate. Immobilizing the animal (at the end of a recording) is necessary to achieve a high SNR image and for most microscopes to perform full-color acquisition. To improve that, Atanas et al., described networks that (i) perform high-quality, non-rigid alignment and (ii) accurately annotate neuron identities from NeuroPAL volumes to form an incredibly useful trace-extraction pipeline:

• BrainAlignNet performed non-rigid registration to track ROIs across freely moving volumes, match them to immobilized frames, and enable freely moving trace extraction
• AutoCellLabeler identifies the neuron class of ROIs within the data-rich immobilized frames, allowing our lab to scale our recording volume and experimental bandwidth, without the bottleneck of hours of manual human labeling per dataset.

This process worked incredibly well, performing high-quality automated labeling and trace extraction, but although BrainAlignNet does clever registration graph solving to minimize the amount of computation required, the pipeline was still bulky and slow. Unfortunately, it was taken for granted that post-hoc trace analysis was inevitable, leading to experiments which had to be limited in scope, performed hoping the animal was in the correct behavioral state, or just discarded.

AutoCellLabel Live

If we wanted to get real time labeling working, the previous approach would simply not work. First, BrainAlignNet was a non-starter, and the computational difficulty of the non-rigid registration problem was not tractable to perform online. Secondly, we couldn't wait until the end of a recording to get our labels, and would need to move past immobilization, instead performing accurate inference on freely moving frames. Luckily, these problems lend themselves to an elegant mutual solution: if we can label each frame accurately, we no long need to match neurons across freely moving frames. (and, we can still use BrainAlignNet post-doc to perform ANTSUN labeling for verification or even just aggregate freely moving predictions and pick up more neurons than standard ANTSUN).

From that idea, I began working on AutoCellLabel Live. ACLL is a pipeline, mostly focused on the FreelyMoving AutoCellLabeler network, but also encompassing highly accelerated preprocessing and trace extraction procedures. Unfortunately, recording practicalities mean we only have one low-SNR fluorescence channel to work with, instead of four high-SNR channels in the immobilized images. Luckily, we also have 1600 times the number of volumes to train on, as each recording only contains one immobilized image. This, however, meant ACL training would have to speed up, as the published version took more than a week to train on just 81 frames. However, even once the network existed, its inference time would need to be drastically decreased, as it look just over a minute to run. We record a volume every 0.6 seconds, so the entire pipeline would need to complete in that time.