Using a 3D dataset

[Jeremy Lynch]

Hi, 

I'm evaluating a few toolkits to analyse some 3D datasets, specifically volumetric cranial MRI scans. Each scan is composed of multiple sequential image slices representing an axial cut through the brain. We would be trying to identify normal vs abnormal scans. Would the Sony NN console be able to handle 3D data in this format?

Kind regards,

Jeremy. 

[Yoshiyuki Kobayashi]

Yes, it becomes possible by updating installed Neural Network Libraries.
On the command prompt, Follow the procedure below to update Neural Network Libraries.

SET PATH=%PATH%;C:\(NNC v1.2 install directory)\libs\Miniconda3;C:\(NNC v1.2 install directory)\libs\Miniconda3\Scripts
pip uninstall nnabla
pip uninstall nnabla-ext-cuda
pip install nnabla
pip install nnabla-ext-cuda

To handle 3D data, first divide the data for each 2d data and then prepare a dataset CSV in the following format.

(In the case of 10 × 20 × 30 data)

x__0,x__1,x__2, ... ,x__9
0_1.png,0_1.png,0_2.png, ... ,0_9.png
1_1.png,1_1.png,1_2.png, ... ,0_9.png
2_1.png,2_1.png,2_2.png, ... ,0_9.png
...

Each png files are monochrome with 20 x 30 pixels

You can input variable "x" in the above dataset by using Input layer and setting Size property to "10,1,20,30"
In order to do 3D Convolution, first set the size to "1,10,20,30" using Reshape layer.

Or you can prepare each data in CSV format and then use Reshape layer to change the shape to the original shape.

x
0.csv
1.csv
2.csv
...

Each CSV files contains 10 x 20 x 30 rows, and values are described in each row.

You can input variable "x" in the above dataset by using Input layer and setting Size property to "6000"
In order to do 3D Convolution, first set the size to "1,10,20,30" using Reshape layer.