What input type for image3dInputLayer

Hi all,

I'm trying to build a classification network using 3-D images as input. However, I can't get the image3dInputLayer to work.

So far, i've used a folder structure with images sorted per classolution in subfolders. As file types I tried a layered tif (which trows an error since it is recognized as a 2-D image), or .mat files each containing a single 570x570x356 matrix.

Could anyone give me a suggestion on how to prepare the input data?

Thanks in advance,

Daan

[Tomaso Cetto]

Hey Daan,

The image3dInputLayer expects data with 3 spatial dimensions and one channel dimension - the inputSize argument ot the layer is in the format [h w d c], where [h w d] are the sizes of the spatial dimensions of the images (which, if I understand correctly is [570 570 356], for your data). If your imags are grayscale, then you'll have c = 1.

That said, the best way for you to input your images into a network will be using datastores (the images are big and so depending on how many you have it'll most probably not be feasible to hold them in-memory). I suspect the datastore you'll want to be using is the imageDatastore.

Hopefully the documentation for this function helps you get started - we also have a documentation example which shows a semantic segmentation workflow using 3-D images. If you're not planning on doing segmentation, there is a bunch of stuff in this example you won't need to worry about, but I mention it because it starts by creating an imageDatastore from the 3D images (and shows how to read them from MAT-files), so that might be of help to you as well.

I hope this helps!