There are many approaches for training CNN on 3d data, but the decision to use a particular architecture is heavily dependant upon the format of your dataset.
If you are using 3d point cloud data, I would suggest you go through PointNet and PointCNN.
But training a CNN on 3d point clouds is very tough.
There is also a way to train CNNs by posing the 3d structure from different viewpoints (Multiview CNNs).
But remember that training CNN on 3d data is really a tough task.
If you plan to use a voxelized input data format, I suggest going through VoxelNet.
Since you are mentioning deconvolution, the most relevant paper I can come across is 3D U-Net: Learning Dense Volumetric Segmentation from Sparse Annotation.
But deconvolution in its own right is a very expensive operation, which acting on 3d data makes it very hard, so I would suggest you check for alternate methods.