Preventing resampling when using RandCropByPosNegLabeld?

[MJJdG]

Hi MONAI team, I'm new to the framework and running into a problem where including RandCropByPosNegLabeld in my data preprocessing pipeline seems to yield volumes resampled to 1mm isometric, whereas I want to keep the sampled VOI's in the original spacing. I've looked through the docs but can't seem to find where this is toggled.

I have tried setting resampling=False in SaveImaged but the resampling already seems to happen in RandCropByPosNegLabeld. Removing it from the pipeline yields the entire volume in the original spacing.

crop_lesion_volume = Compose(
                    [
                        RandCropByPosNegLabeld(
                            keys=["img", "seg"], # img and seg have the same size and spacing
                            spatial_size=(
                                self.height_width, # = 256
                                self.height_width, # = 256
                                self.depth, # = 128
                            ),
                            label_key="seg",
                            neg=0,
                            num_samples=self.num_samples,
                        ),
                        SaveImaged(
                            keys=["img"],
                            resample=True,
                            mode="trilinear",
                            squeeze_end_dims=True,
                            folder_layout=img_out_path,
                        ),
                        SaveImaged(
                            keys=["seg"],
                            resample=True,
                            squeeze_end_dims=True,
                            folder_layout=seg_out_path,
                        ),
                    ]
                )

[wyli]

The code looks good to me and the crop_lesion_volume(data) shouldn't change data's resolution, probably the resolutions have been changed somewhere else..

[MJJdG]

Hmm interesting, when I examine the img metadata directly before this step I get:

{'sizeof_hdr': array(348, dtype=int32), 
'extents': array(0, dtype=int32), 
'session_error': array(0, dtype=int16), 
'dim_info': array(0, dtype=uint8), 
'dim': array([  3, 512, 512,  13,   1,   1,   1,   1], dtype=int16), 
'intent_p1': array(0., dtype=float32), 
'intent_p2': array(0., dtype=float32), 
'intent_p3': array(0., dtype=float32), 
'intent_code': array(0, dtype=int16), 
'datatype': array(8, dtype=int16), 'bitpix': array(32, dtype=int16), 
'slice_start': array(0, dtype=int16), 
'pixdim': array([1.      , 0.976562, 0.976562, 5.      , 1.      , 1.      ,
       1.      , 1.      ], dtype=float32), 
'vox_offset': array(0., dtype=float32), 
'scl_slope': array(nan, dtype=float32), 
'scl_inter': array(nan, dtype=float32), 
'slice_end': array(0, dtype=int16), 
'slice_code': array(0, dtype=uint8), 
'xyzt_units': array(0, dtype=uint8), 
'cal_max': array(0., dtype=float32), 
'cal_min': array(0., dtype=float32), 
'slice_duration': array(0., dtype=float32), 
'toffset': array(0., dtype=float32), 
'glmax': array(0, dtype=int32), 
'glmin': array(0, dtype=int32), 
'qform_code': array(0, dtype=int16), 
'sform_code': array(2, dtype=int16), 
'quatern_b': array(0., dtype=float32), 
'quatern_c': array(0., dtype=float32), 
'quatern_d': array(0., dtype=float32), 
'qoffset_x': array(0., dtype=float32), 
'qoffset_y': array(0., dtype=float32), 
'qoffset_z': array(0., dtype=float32), 
'srow_x': array([0.976562, 0.      , 0.      , 0.      ], dtype=float32), 
'srow_y': array([0.      , 0.976562, 0.      , 0.      ], dtype=float32), 
'srow_z': array([0., 0., 5., 0.], dtype=float32), 
affine: tensor([[   0.9766,    0.0000,    0.0000, -124.9999],
        [   0.0000,    0.9766,    0.0000, -124.9999],
        [   0.0000,    0.0000,    5.0000, -320.0000],
        [   0.0000,    0.0000,    0.0000,    1.0000]], dtype=torch.float64), 
original_affine: array([[0.97656202, 0.        , 0.        , 0.        ],
       [0.        , 0.97656202, 0.        , 0.        ],
       [0.        , 0.        , 5.        , 0.        ],
       [0.        , 0.        , 0.        , 1.        ]]), 
'as_closest_canonical': False, 
spatial_shape: array([512, 512,  13], dtype=int16), 
space: RAS, 
original_channel_dim: nan, 
'filename_or_obj': 'ct01.nii.gz'}

Where the pixdims and affine seem to still show the original spacing. Before the RandCropByPosNegLabeld step I've used the following functions LoadImaged -> EnsureChannelFirstd -> Orientationd -> BorderPadd. But those shouldn't have changed the resolution either as far as I'm aware...