Refining tracking performance when using own masks

[raeesahhy]

Hi, thanks for this tool!

I have 3D+t datasets with cells very densely compacted so have had little success using Ultrack's native segmentation approaches.

I have managed to feed in masks from Cellpose/CPSAM into Ultrack and use this to ask Ultrack to auto detect contours from the input labels.

However I am running into issues with the tracking.
When Ultrack creates its 'segments' layer from my labels I see colour switching between frames which I take to mean that the tracks it produces are too short and it is not linking cells between frames
I have tried a number of parameters below with little success.
The problems may also be due to Ultrack splitting cells which were segmented well in the original segmentation by CPSAM as per the below image.

Untitled 1.pdf

Any advice on what next I can try would be much appreciated.

sigma = 0, 1, 2

config.segmentation_config.min_frontier = 0.35
config.linking_config.max_neighbors = 5
config.linking_config.max_distance = 50

config.tracking_config.appear_weight = -0.001, -0.05, -1.0
config.tracking_config.disappear_weight = -0.001, -0.05, -1.0
config.tracking_config.division_weight = -0.001, -0.05, -0.1

config.tracking_config.window_size = 15
config.tracking_config.overlap_size = 3
config.tracking_config.solution_gap = 0.01

[JoOkuma]

Hi @raeesahhy,

I these the parameters which you have already tested?

In these cases my recommendation is to increase the min_frontier which you already have done and reduce the range between min_area and max_area to match your input (these are pixel counts). Have you tried that?

Have you read this section? It provide some useful tips to optimize the results.

[raeesahhy]

Thanks @JoOkuma!

Yes, all those parameters I've tested. I have read the section you've linked to but testing each of the parameters one-by-one is time-consuming so I wanted to ask.

Reducing the range between min_area and max_area has helped improve it a little - I'll keep going trying to optimise it.