Make the background channel optional in the napari plugin

[aymuos15]

Follow-up to the core single-channel support (#619), extending it through the napari UI so users can run detection, curation, and training with only a signal channel. Also closes #636 by auto-selecting the single-channel model in the Python API.

What changed

• Detection (detect.py): the background image is now optional. A missing background is passed to core as None. Added a guard: running without a background requires a single-channel model, so the widget prompts the user to enable "Skip classification" or uncheck "Use pre-trained weights" and pick a single-channel model.
• Curation (curation.py): background layer is optional throughout. When absent, only the signal channel is written, and the training YAML marks bg_channel = -1 (which core training already reads as signal-only). Signal/background selections reset cleanly to None when deselected.
• Core classify (classify.py): skips background normalization when there is no background array.
• Core model selection (main.py): when background_array is None and the default multi-channel model is still selected, the API switches to resnet50_1ch, avoiding a dimension mismatch ([Feature] Choose our "resnet50_1ch" as the default if background array is None #636). An explicitly chosen model is left untouched.
• Containers: DataInputs.background_array is now Optional.

Tests

• Save single-channel training data produces single-channel cubes and a bg_channel: -1 YAML, verified consumable by core training.
• check_image_data_for_extraction accepts signal-only; missing-signal message updated.
• Deselecting the background resets the layer.
• DataInputs(background_array=None) passthrough and single-channel detection worker smoke test.
• Missing background auto-selects resnet50_1ch; an explicit resnet50_all is preserved.

[aymuos15]

@alessandrofelder I realize targetting two things at once is slightly messy but they were really intertwined so hence I put it this way. Happy to split your issue seperately if you that makes this revewing easier.

[alessandrofelder]

Thanks a lot for this @aymuos15

I've only skimmed the code changes (will come back to it later this week) and done some initial user-testing (largely all looks reasonable to me). Thought it might be useful to flag an initial specific comment before I have time to look more deeply - sorry if I misunderstood!

[alessandrofelder]

I think one should actually check "Use pre-trained weights" to use the default single channel model, no?

(If not, there is no easy way through the GUI to download and use the single-channel default model?)

(and therefore related logic needs to be adapted 🤔 )

[aymuos15]

This was a brain fog moment of mine. I think I was just doing everything locally and sided with that. Sorry about that! Thank you very much for catching it! :D

[aymuos15]

Have commited this.

[alessandrofelder]

sorry - I am still unsure about this 🤔
Will this not evaluate to True and return too early if

• I don't have a background channel
• I want to classify (i.e. I don't want to skip classification)
• I have trained my own single-channel model and therefore do not want to use the pre-trained model?

I wonder whether we want more something like (pseudo-code):

n_channels = 2 if background_image else 1
model_has_correct_number_of channels = use_pre_trained_weights or model.channel_dimension == n_channels

if not model_has_correct_number_of_channels:
     # sensible info message and return

[aymuos15]

Apologies once again, I think I confused myself here. I just wanna quickly make sure my understanding is right this time.

First, for context: How the default path works (pretrained ticked):

Step	File	Action
1	detect.py:457	pretrained → trained_model = None (Path.home() discarded)
2	main.py:215	no background → resnet50_tv → resnet50_1ch
3	main.py:218 → prep.py:35	no weights given → download_models(...)
4	tools.py:get_model	no custom model → build_model + load_weights(default)
5	classify.py:180	channel check

This guard was set up because the tick-box use_pre_trained_weights is the real switch, but trained_model starts as Path.home() (a folder), not empty, and only becomes empty when the tick-box is on. So when the tick-box is off, there's no clean way to tell "I picked a model" from "I picked nothing", both look like a path. The guard just patches over that. I think this was what led to my choice above.

Maybe, we can actually just collapse to pretrained / custom / skip into one exclusive choice, so "nothing picked" is a checkable state and the guard isn't needed. Current vs proposed behaviour:

Model choice	Background	Current	Proposed
Pretrained	yes	2ch default, proceeds	same
Pretrained	no	1ch default, proceeds	same
Custom file	yes	proceeds; core errors if ≠ 2ch	same
Custom file	no	guard blocks it	proceeds; core errors if ≠ 1ch
No model picked	yes	cryptic load_model(home) error	blocked, "select a file"
No model picked	no	guard blocks (masks that error)	blocked, "select a file"
Skip classification	yes	detection only	same
Skip classification	no	detection only	same

Since the logic you mentioned already exists there (classify.py:180-188), can we just follow the above and patch it there?

# detect_containers.py
class ModelSource(Enum):
    PRETRAINED = "Pretrained default"
    CUSTOM = "Custom model"
    SKIP = "Skip classification"

# ClassificationInputs: one field replaces the two flags
    model_source: ModelSource = ModelSource.PRETRAINED

# as_core_arguments(): translate the choice for the core
    args["skip_classification"] = self.model_source is ModelSource.SKIP
    if self.model_source is not ModelSource.CUSTOM:
        args["trained_model"] = None   # core downloads the default

# detect.py: derive skip, guard only the "custom but no file" case
    skip_classification = model_source is ModelSource.SKIP
    if model_source is ModelSource.CUSTOM and not (
        trained_model and Path(trained_model).is_file()
    ):
        show_info("Select a trained model file.")
        return

[alessandrofelder]

No worries at all, and thanks for dissecting this @aymuos15 - the fact that we use both a Path and a checkbox to encode one thing is confusing currently for sure! (See related 4-year issue I made that never went anywhere 😂 #328)

I agree with the "proposed" column in your table above, I think that's what the behaviour should be. My only minor note would be that the error message should also not the option to tick "pretrained"?

# detect.py: derive skip, guard only the "custom but no file" case
    skip_classification = model_source is ModelSource.SKIP
    if model_source is ModelSource.CUSTOM and not (
        trained_model and Path(trained_model).is_file()
    ):
        show_info("Select a trained model file or check the `Use pre-trained weights` to use the default model.")
        return

[aymuos15]

My bad, Shouldve spotted that issue in advance haha. Thank you for pointing it out, I have added it to the todo list in our projct and will PR that after this.

In my latest commit, I have done what we discused in the table along with your proposed error message. Thank you!