Complete HNC segmentation implementation

[wbwakeman]

Complete remaining work to have a working implementation of HNC segmentation algorithm

Tasks:

• [x] Read the underlying paper
• [x] Revisit their code repository to determine if it may be reusable for our purposes
• [x] Either wrap their implementation or write our own implementation of their approach
• [x] If wrap their code, write in this ticket a proposal for how we would productionize it.
• [x] Integrate with the rest of our proto-segmentation pipeline

Validation criteria:

• [x] code is merged
• [x] can show running code and plotted outputs from the segmentation

[djkapner]

I have wrapped the HNCcorr code in this development branch: https://github.com/AllenInstitute/ophys_etl_pipelines/pull/255

This implementation:

• [x] uses our generica seeder (same seeder Scott's segmenter is now using)
• [x] replaces their ROI candidate class and outer-most segmentation loop with close relations that improve the explicit nature of the algorithm.

Analysis

This works. We've gotten a lot of good ideas from their paper and their approach. From what I've seen so far, we wouldn't simply use the results coming out of this. We'd want to clean up the results, probably by passing through the merging tool that Scott is developing. We'd want to resolve the non-deterministic nature of it. Given that Scott's approach is also yielding fruitful results, and, we know/own that code in and out, I'd be more inclined to invest our precious time in that branch of development.

TODO (if desired):

• [ ] Paralleliize. We are using ImageMetricSeeder which is an iterator that serves a single seed at a time. We also created a class called BatchImageMetricSeeder which would serve up batches of seeds. The candidate class and outer segmentation loop would need modification so that they could be run in spawned processes. Perhaps passing in a instantiate Patch object.
• [ ] The results are non-deterministic. We think this comes from the use of ApproximatePCA in their sparsecomputation repo. This could be replaced with their non-approximate PCA but it would require stripping away another few layers of their code we are currently using.