# Object fingerprinting and tracking - Workshop 2

# Attendees
* Evie Corcoran
* Ben Evans :notebook:
* Andrew Fleming
* Chris Soelistyo :headphones: :video_camera: :shallow_pan_of_food:
* Arianna Sallili-James :bat: :bird: :cake: :ant:
* Krisina Ulicna :female-technologist: :muscle:
* Evangeline Corcoran :leaves:
* Hiris Gu :bat: :four_leaf_clover: :palm_tree: :whale: :cat2:
* Marjan Famili :dragon_face: :tea: :tea: :tea:
* Nathan Day :8ball: :kissing_cat: :egg: :sunflower:
* Alan Lowe :dizzy_face: :email: :email: :email: :fire:
* Bea Costa Gomes :coffee: :coffee: :coffee:
* Louisa Van Zeeland :croissant::tea:
* Ania Babula :biohazard_sign: :microscope:

## Agenda
### Updates/Intro

Alan's 'trackathon' summary
Project outline - Kristina
Priscilla
Chris background
Bea
Ania
Sanson
Hiris
Nathan
Ben
Anyone else - updates/questions etc.

### Discussion Points

* Dataset interoperability - consolidated archive of segmentations from different use-cases?
* Geospatial capability (inc. vector-format)
* Multiple children and lineages
* Options for motion models and how to implement
* Evaluation of track quality (HOTA?)
* Most generalisable options for shape description and matching - what are people using and why?
* Scaling tracks to large datasets/across image tiles

### Notes

#### Alan
Tracking metrics:
Janelia - traccuracy
-pip installable
Hopefully some of Kristina's metrics to be incorporated

btrack now outputs candidate graphs

Working on a graphs layer for Napari

People using btrack to follow bats in 3d within thermal image data - Centre for Advanced Study of Collective Behaviour

Also pollen tracking

motile - graph tracking optimisation, but doesn't build the graph itself. Can add constraints then solve cf. btrack which is more tracklet-based (1st pass then ILP on top of that) cf tracking full tracks within ILP.
motile specifically set up to use time (e.g. Bea's problem doesn't have a time domain)

Data transfer discussions: 'experts' were pushing them towards Apache Arrow (being used in GRACE)
can't store a tuple

Bat tracking
https://thejasvibr.github.io/

Pollen tracking
https://www.cedarwarman.com/

traccuracy
https://github.com/Janelia-Trackathon-2023/traccuracy

motile
https://funkelab.github.io/motile/

#### Kristina

Cell trajectory analysis
NEUBIAS symposium (cell tracking/lineaging)
Janelia trackathon repo
Repurposing Bea's GRACE project into a tracking tool.

Discussion of error sources:
branching/merging
Identity preservation
Segmentation

Evie overlap with Kristina's latent phase representation/Dynamic time warping classification.
Can pass any features into btrack and track on those alone (no motion).

#### Priscilla
MRI images of tumours (kidney)
Stacks of images - extract part with tumour
Analyse images from CT and MRI together - merge data from both modalities
Mobility of whole kideny within body between acquisitions
Need to co-register the locations of the tumours if they are to analyse the modalities together.

Arianna - problem screams LDDMM* to her. (-- there is new work on LDDMM for registration in 3D imaging which could be interesting to play with.)
Priscilla - using gradient based methods.

*Large Deformation Diffeomorphic Metric Mapping

#### Chris
Derive rules governing a system from its behaviour
DNN for patten-finding to internalise rules
DNN to predict whether a cell is going to A- divide or B- die
Take lots of examples of mitosis and apoptosis, truncate before event and predict.
https://www.nature.com/articles/s42256-022-00503-6

#### Shameless Bea Plug = PINT OF SCIENCE is APPROACHING! :beers:
https://www.turing.ac.uk/events/pint-science-0

* [name=Kristina] Also on Twitter - buy your tickets SOON! [Link here](https://twitter.com/pintofscience/status/1647877808109268992)

#### Ania

#### Hiris
Hopefully able to give a fulller outline at next meeting.

#### Nathan

Loves Cat Stevens

32 gb arrays. (75, 2, 3, 6480, 6480)
using DASK for lazy loading of tiled images - would like someone experienced with DASK to look over code
Track lengths increase when he rescales the FOV
34% of errors ID switches - often segmentation-related
For less dense areas 70% well tracked
Recolour segmentation function in btrack now
https://btrack.readthedocs.io/en/latest/api/btrack.utils.update_segmentation.html#btrack.utils.update_segmentation
Morphological traits including 'foaminess' may relate to mtb persistence.
scale option in btrack - so can keep seg the same but coarsen the internal representation within btrack without having to edit the raw data - will then map back to the original image size

prob_not_assign parameter.... Set really low in default models. See what happens if set it much higher e.g. 0.1

Can also specify which hypotheses you're generating - e.g. cut out the cell division hypotheses. #