Feature Request: Add SparseDice, SparseClDice, CenterlineDice Losses and 3D TransUNet

Hello MONAI team,

I would like to request the addition of several features to MONAI that could benefit the community for 3D medical image segmentation tasks:

SparseDiceLoss – A Dice loss variant that supports sparse labels, useful for multi-class segmentation with incomplete annotations.

SparseClDiceLoss (Centerline Dice for sparse data) – Combines clDice and sparse label support for tubular structures such as vessels, fibers, or centerlines.

CenterlineDiceLoss (clDiceLoss) – Measures topological agreement between predicted and ground-truth structures by comparing their soft skeletons; very useful for thin elongated objects in 2D/3D.

3D TransUNet – Extend the current 2D TransUNet to full 3D, to handle volumetric medical images directly.

These additions would greatly help users working on complex 3D segmentation tasks, especially with sparse annotations and tubular structures.

Thank you for considering this feature request!

Review MONAI's existing 2D TransUNet and loss implementations first, then separate the request into sparse-label losses, centerline Dice, and 3D TransUNet. Define completion against the stated needs: sparse annotations, topological agreement for tubular structures, and direct volumetric medical-image handling; the issue does not name files or tests.