would be great if anyone checks if this download works https://s3.amazonaws.com/neuro.datasets/challenges/neurofinder/neurofinder.04.00.zip
No, the coordinates represent where the neuron is located, and this location is assumed to be constant throughout the movie. The data structure is as follows: You have a movie with, for example, 512x512 pixels per frame. The field of view of the movie is constant throughout all movie frames. Within this field of view, you have many neurons, whose location is also assumed to be constant. The goal is to figure out where the neurons are, i.e. to map pixels to neurons. The coordinates in the regions.json provide training data for this, i.e. for each region (=neuron), they list the coordinates of the pixels that were determined to be part of that region. So these coordinates index into the 512x512 spatial pixels of the movie.
To figure out when a neuron is firing, you might calculate, for each frame separately, the average intensity in all pixels belonging to that neuron. This will give you one datapoint per frame, i.e. a timeseries of fluorescence values. The brighter the fluorescence, the more the neuron is firing at that point in time.
@freeman-lab Hi! http://datasets.codeneuro.org/ is down for me and I'm trying to understand the exact dimensions of the data. As I understand some codeneuro challenge data are multi-slice time series and others are multi-slice series, where can I find which are which? (and also how many slices are in each)
the structure of the neurofinder data is quite simple and they're all the same
each dataset is just a collection of 2d TIFF images
the only thing that differs is the exact xy dimensions of the images, and the number of images
that info is available in the
info.json file that comes with each dataset when you download it
also FYI we'll be posting two new datasets later today
and at that point the core neurofinder datasets will be finalized
@jwittenbach / steve neurwin once i'm done maybe you can rerun yours to make sure everything is functioning with the new data
datasets
04.00 and 04.01
reran the "demo submissions" on them and everything went through fine
would be great if someone downloaded the actual data to double check formatting and stuff
@jovo there's no 3d data in the neurofinder datasets and we probably won't add any
from this point forward the datasets are more or less final, at least for this challenge
so people can start submitting algorithms!
the "Suite2P" algorithm from @marius10p is doing the best, but isn't dramatically different from the NMF-based approaches
@d-v-b nmf and cnmf were run on the cluster and only took a couple minutes per dataset running on approximately 10 nodes.
or is there something about the algorithm that crucially leverages GPUs
I expect most of the run time you got was data reads and writes, as well as the SVD decomposition, which you should lower the parameters for, because the datasets are very short.
The number of clusters also will increase the run times of the cell detection part in Suite2P. The defaults assume several hundred (active) ROIs, and is robust in that range. These datasets have an order of magnitude fewer ROIs.