Automatic Head Model including 1010-System / Electrode Placement

[bryn]

But previously you managed to generate them, i.e. before you edited the transform?

[lucky_lin]

No, I only did these things after creating a new file.

[bryn]

I can reproduce your issue. If I set the transform to "0", i.e. Identity, the predictor fails. The head40 segmentation is also less accurate! We need to investigate.

A workaround would be:

• load image
• predict landmarks & segmentation
• compute inverse image transform
• apply this inverse to landmarks/segmentation/surfaces/etc

# assumes verts and labelfield are already predicted (without setting transform to "0")
inv_tx = img.Transform.Inverse()

# transform segmentation
labelfield.ApplyTransform(inv_tx )

# transform landmarks
for v in verts:
    v.ApplyTransform(inv_tx )

[bryn]

The issue is that our neural network was trained with the data in the RAS orientation (with some deviation +- 15 degrees, and flipping in all axes). If you manually edit the transform, you break assumptions used to pre-orient the data into RAS ordering.

Since RAS is a widely used convention in neuroscience, and medical images are always acquired with a direction (rotation) matrix and offset (translation), I think it is best you don't modify the transform.

For instance, if you try to assign DTI-based conductivity maps - you will need to rotate the grid AND the tensors accordingly. It can be done, but it will be more effort...

If this is to investigate if the fields are (nearly) symmetric, I suggest you

• find an approximate symmetry plane (wrt to the brain or skull or ...)
• align the plane of a slice viewer perpendicular to the symmetry plane

[lucky_lin]

@bryn Does this code only segment 40 types of tissues by default? I want to segment 16 types.

[bryn]

The default is 40 tissues. To be explicit you can specify this via

import ImageML

labelfield = ImageML.HeadModelGeneration([img], output_spacing=0.6, add_dura=False, version=ImageML.eHeadModel.head40)

For 30 (or 16) tissues you would specify the version head30 (or head16)

import ImageML

labelfield = ImageML.HeadModelGeneration([img], output_spacing=0.6, add_dura=False, version=ImageML.eHeadModel.head30)

But please note: the versions are an evolution. The head16 segmentation is not the same, with fewer tissues. It is also less accurate, as it was the first version we published (and trained on less training data).

[lucky_lin]

@bryn Thank you very much for your response! I have a question: What is the difference between constructing a head model using T1-weighted (T1W) and T2-weighted (T2W) images and constructing a head model using only T1W images? Why can only 16 types of tissues be segmented when using T1W and T2W images?

[bryn]

@lucky_lin In our first version of the head segmenation (head16) we trained with a smaller dataset, where T1w and T2w was available. We trained two networks, one with just T1w as input, one that gets T1w + T2w as input.

In our later work we extended the training data, but only have T1w images. Therefore, the head30 and head40 only needs a T1w image.

[lucky_lin]

@bryn Okay, I understand ^^

[lucky_lin]

@bryn Hello, if I use script, can I clone an already set up simulation and then make partial modifications to the settings?