fix loadtiffd spatial shape

[fanweiya]

Description:

This PR addresses a bug in the vista2d/components.py:LoadTiffd transform that causes RuntimeError: each element in list of batch should be of equal size during data collation when using pad_list_data_collate or similar functions.

Problem:

The current implementation calculates and stores the ImageMetaKey.SPATIAL_SHAPE metadata before normalizing the image array's dimensions and channel count. Input image files can have different original dimensions (e.g., 2D grayscale (H, W), 3D channels-last (H, W, C), or 3D channels-first (C, H, W)). This leads to the stored spatial_shape metadata being a tuple of inconsistent lengths (e.g., length 2 or length 3) across different samples in the dataset.

When the DataLoader attempts to collate a batch containing samples with different spatial_shape lengths, the default collation logic for metadata (monai.data.utils.collate_meta_tensor_fn -> torch.utils.data._utils.collate.default_collate) fails because it cannot stack tuples/lists of varying sizes.

Solution:

The fix implemented in this PR modifies LoadTiffd as follows:

1. The calculation of the shape to be stored in ImageMetaKey.SPATIAL_SHAPE is moved to after all image array manipulations (loading, transposing, adding/repeating/selecting channels).
2. The code now ensures the image array is consistently represented in a 3D (C, H, W) format internally before metadata creation.
3. It explicitly extracts only the spatial dimensions (the last two dimensions, H and W) from the final processed 3D array.
4. This extracted (H, W) tuple, which always has a length of 2, is stored as the value for ImageMetaKey.SPATIAL_SHAPE in the MetaTensor.

Outcome:

This change guarantees that the spatial_shape metadata associated with loaded images has a consistent length (always 2) across all samples, regardless of the original file's dimensions. This resolves the downstream collation error. Robustness for handling different input dimensions (2D, 3D channels-last/first, >3D) has also been slightly improved with added logging.