Improve the predict output file options

Issue description

Currently, predict/deepreg_predict saves input images/labels for different label pairs, with both png and nifti file formats. This leads to sizable written files.

1 - Add options that can turn off the following for space-saving/speed:

• Save input images/labels (useful for debugging or research);
• Save png files
• Save nifti files (is this useful to be able to turn off?)

2 - Avoid saving the same input files multiple times.

Type of Issue

• New feature request
• Documentation update
• Test request

What’s the expected result?

What’s the actual result?

Additional details / screenshot

-

[NMontanaBrown]

Also, I was reviewing this for the edu challenge and at this line we call:

# predict
 # (batch, f_dim1, f_dim2, f_dim3)
        pred_fixed_label = outputs_dict.get("pred_fixed_label", None)
        pred_fixed_image = (
            layer_util.resample(vol=moving_image, loc=fixed_grid_ref + ddf)
            if ddf is not None
            else None
        )

Which I found quite confusing: the warped image is called pred_fixed_image, which I associated to the fixed image. I would suggest renaming this and the save file for clarity to warped_moving_image, or something along the lines?

@YipengHu @mathpluscode @s-sd

[mathpluscode]

so ideally, the output should be like following (this has been discussed before lol)

• one folder per image pair (not related to label)
  • input images
  • one subfolder per label
    • input label
    • outputs

then for the following items we can choose to save or not

• input images/labels, default false
• png, default false
• nifti, default true

@YipengHu what do you think?