[Auto3dSeg] Inference for new cases

[imruljubair]

Hi,

Question about: Auto3dSeg.

I used AutoRunner for algo=segresnet. There were 4 folds. I finished training for all folds and got results in ensembled_output. Now I want to run inference for a new set of test files (may be in different directory) with my trained ensemble model.

I believe there is an easy solutions to do that, which I can't find in tutorial. May be I missed something.

Can anyone guide me please?

Jubair

[KumoLiu]

Hi @imruljubair, thanks for your interest here.
You can use EnsembleRunner directly.

MONAI/monai/apps/auto3dseg/ensemble_builder.py

Line 397 in 50f9aea

class EnsembleRunner:

Also here is a tutorial about how to customize your ensemble pipeline by defining a new ensemble class.
https://github.com/Project-MONAI/tutorials/blob/main/auto3dseg/notebooks/ensemble_byoc.ipynb

Hope it helps, thanks.

[imruljubair]

@KumoLiu
Thank you.
I need to use this: ensemble_runner = EnsembleRunner(data_src_cfg_name, work_dir, ensemble_method_name, mgpu=device_setting['n_devices']>1, **kwargs, **pred_params)

Could you please tell me what is data_src_cfg_name?

Thanks.

[KumoLiu]

It is the str that points to the data source path.
An example here:
https://github.com/Project-MONAI/tutorials/blob/main/auto3dseg/tasks/btcv/input.yaml

[imruljubair]

@KumoLiu
Is it possible to perform predictions on the files which are not added in the folds.json. For instances, in my folds.json I have -

    "testing": [
        {
            "image": "./imagesTs/pat43.nii.gz"
        },
        {
            "image": "./imagesTs/pat57.nii.gz"
        },
        {
            "image": "./imagesTs/pat29.nii.gz"
        },
        {
            "image": "./imagesTs/pat60.nii.gz"
        }
               ]

But I have new test cases which are in different directory and different names. Can I do that if I do not want to regenerate the folds.json?

[KumoLiu]

Hi @imruljubair,
Currently, the functionality you're interested in doesn't have a straightforward implementation without making some alterations to the code. If you would wish to make these modifications, refer to this section of the codebase:

MONAI/monai/apps/auto3dseg/ensemble_builder.py

Line 387 in f6f9e81

ensemble.set_infer_files(self.data_src_cfg["dataroot"], self.data_src_cfg["datalist"])

Additionally, we have a tutorial on preparing the data, which could prove beneficial for you: MSD Datalist Generator.
Let us know if you need further assistance.
Best,

[che85]

@KumoLiu Hi. Why is there no easy way to run auto3dseg on new data by just providing a list of files and the original work_dir of the training results? The Slicer MONAIAuto3DSeg module only supports running for a single model (no ensemble is supported). If ensemble is needed, developers recommend using auto3dseg API. Now I am learning, that there is no easy way to run inference for new unseen data that was not part of the original data.

How is auto3dseg intended to be used post-training if these functionalities are not provided? Is the sole purpose to train and validate a model and then move on?