This example demonstrates training a model using Vertical Partitioning on CT
scans in DICOM format with synthetic tabular data to predict patient age.

The image data is CT scans in DICOM format from:
https://www.kaggle.com/kmader/siim-medical-images?select=overview.csv.  The
first step is to pull that into the TripleBlind Package format and generate
synthetic data in csv format. The DICOM images are originally sized at 512x512,
and they are not adjusted before uploading -- resizing and other preprocessing
happens at model train/inference time. The tabular data has 100 features
(or columns) with 99 rows to match the 99 images which we will be using in
training.

Next, the DICOM images for testing and training are positioned on the second
organization while the tabular data is positioned on the third.

In the third script, the model structures are defined.  The model must be
defined with a "split" with one training against the DICOM images; a second
model for training against the tabular data, and a final model which takes the
outputs of the previous two models as input in a single layer, outputting the
age. The server model resides on the first organization's Access Point, as they
are the initiator of the training operation.

In the final script the first organization performs distributed inference
against data found on the second and third organizations.
