Is it possible to run inference using a BiaPy model from within ImageJ ?

[cyrilbz]

Hello there,

I am satisfied with a model I trained using BiaPy, and wanted to start including it into an image processing workflow.

From this page : https://biapy.readthedocs.io/en/latest/for_developers/library_examples.html I understand how to do using Python, but I was wondering if it is possible (or will be!) to use a BiaPy model from within ImageJ ? for example by launching it as a sub-process (like the Cellpose-BIOP plugin), or by exporting the model to the BioImageModelZoo format and then using it with DeepImageJ?

Thank you for your help, and thanks again for this amazing tool!

Best regards,

Cyril

[danifranco]

Hi,

We have no plugin for ImageJ/Fiji for BiaPy yet, but it is in our roadmap and hopefully will be developed soon. Now you can proceed as you were describing: expert the model into BMZ format and reuse it using DeepImageJ.

Best,

Dani

[cyrilbz]

Sounds good, I'll try the BMZ export and DeepImageJ option! 👍

Thank you!

[danifranco]

Great! Let me know if you have problems and I will be more than glad to help!

[cyrilbz]

Hi @danifranco

I have an issue when trying to export my model to BMZ format.

I have the following error :
Traceback (most recent call last): File "/home/cbozonnet/Documents/image_processing/BiaPy/script_biapy.py", line 43, in <module> biapy.export_model_to_bmz(bmz_folder, bmz_cfg=bmz_cfg) File "/home/cbozonnet/miniconda3/envs/biapy-env/lib/python3.12/site-packages/biapy/_biapy.py", line 984, in export_model_to_bmz with open(arch_file_path, "w") as f: ^^^^^^^^^^^^^^^^^^^^^^^^^ FileNotFoundError: [Errno 2] No such file or directory: './bmz_model/model.py'

And my script is the following:

#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
Created on Thu Oct 16 2025

@author: cbozonnet
"""
from biapy import BiaPy

# --- Set up the parameters 
config_path = "config.yaml"            # Path to your YAML configuration file
result_dir = "./results"                 # Directory to store the results
job_name = "ResidualUnet_Clermont_Lesvos-testtoBMZ"                       # Name of the job
run_id = 1                                      # Run ID for logging/versioning
gpu = "0"                                       # GPU to use (as string, e.g., "0")

# --- Create and run the BiaPy job
biapy = BiaPy(config_path, result_dir=result_dir, name=job_name, run_id=run_id, gpu=gpu)
biapy.run_job()

# --- export to BioimageModelZoo format

# Create a dictionary with all BMZ requirements
bmz_cfg = {}
bmz_cfg["description"] = "Plant segmentation for field arabidopsis"
bmz_cfg["authors"] = [{"name": "Cyril Bozonnet", "github_user": "cyrilbz"}]
bmz_cfg["license"] = "CC-BY-4.0"
bmz_cfg["tags"] = ["fieldplants", "residualUnet"]
bmz_cfg["model_name"] = "arabidoJ"
bmz_cfg["cite"] = [
        {"text": "training library", "doi": "10.1038/s41592-025-02699-y"},
        {"text": "architecture", "doi": "10.1109/LGRS.2018.2802944"},
        {"text": "data", "doi": "10.48550/arXiv.1812.06024"},
        ]
bmz_cfg["data"] = {
        "name": "blablabla",
        "doi": "blablabla",
        "image_modality": "blablabla",
    }

# Export model
bmz_folder = "./bmz_model"
biapy.export_model_to_bmz(bmz_folder, bmz_cfg=bmz_cfg)

the 'bmz_model` folder does exist. I do not understand what I am missing to get the full export working? I already updated my BiaPy env to the last biapy version (using conda).

Thanks a lot!

Cyril

[danifranco]

Can you set up the full path instead of the relative path for bmz_folder? E.g. /home/user/bmz_model

[cyrilbz]

Yep, it was the issue, thank you very much! I had to set up the full path to the results folder in addition to the bmz folder and everything went well.

Tried to run it with DeepImageJ, it works! There are some slight differences in the output between DeepImageJ and Biapy, see screenshot below. I do not know if this is expected or not, nor if we have ways to do better in terms of similarity? In any case, I guess a better trained model could help!!

[danifranco]

Glad that it worked!! I will check the code for the upcoming release so it does not crash when using relative paths.

Regarding the differences, by looking at the first white blob's shape at the top, is something related to the inference by patches. In deepImageJ seems that no padding was done as the shape of that blob is jagged and seems to be cut. Donno how you configure the workflow in BiaPy but a minimum padding is aolways done, so maybe, depending on the input image shape and the patch size selected, that minimum padding was enough to solve these border effects. I will try to check if I can modify something in the BMZ package that BiaPy builds so I can add a bit of padding by default too.

[danifranco]

The problem with the relative paths solved here: 44dbef6 !

[cyrilbz]

config.yaml

Here is my config.yaml file. I guess I did not change anything regarding the padding compared to the first config file generated by the GUI. There is some padding for the test (32,32), but not for the training. Would adding padding during training solve (or improve) the discrepancy? Thank you!

[danifranco]

Your config file is perfect as it is. The discrepancy comes from the test image prediction, not from the training itself. It seems that the padding and sliding-window operations used to reconstruct the final test image, when the patch size is smaller than the input test image, which is common, are implicitly defined in the BMZ package that gets created.

I’ll check how I can configure this so that all models exported in BiaPy include a bit of padding by default, ensuring consistency when used in external software such as deepImageJ. I assume deepImageJ follows the procedure defined in the BMZ package and executes it as provided. Currently, I don’t think I’m adding any padding during those steps in the BMZ package, which would explain the output difference observed in deepImageJ.

Thank you so much for the details Cyril !!

[cyrilbz]

Sounds good, thank you for these detailed explanations!!

I think we can mark the issue as closed. Let me know if you manage to improve the export to BMZ as you said, I could try to reproduce the exact same case to see if it improves things. I can now go on with improving my training and easily integrate it in my existing ImageJ workflows!

Thanks again for your kind help,

Cyril

[danifranco]

Thank you so much! I will ping you back when I manage to solve this so you can check. Also, I would like accelerate the process of integrating BiaPy in Fiji so you can remove a few steps in your pipeline. Thanks again!