search

ivadomed / Counterfactual_synthesis_for_lesion_segmentation

1 stars 0 forks source link

Testing on the brain, a less homogenenous area #7

Open BreziTasbi opened 3 months ago

BreziTasbi commented 3 months ago

Since the current model cannot do much more than finding the correct value to color the spinal coord with, i am going to try to find a Healthy brain dataset in order to train a new model on it and see if the method can be more relevant on a it. Thus i have to gather: -A healthy brain dataset with a couple hundreds of subjects. -A brain dataset with similar acquisition parameters (contrast, resolution) and visible lesions.

BreziTasbi commented 3 months ago

As the brain cannot be reduced to a 32x256x256 volume, I tried to reduce lower the latent space size by a factor 2 on each dimension shape in order to allow the management of 256x256x256 volumes. Event if this actually is the recommended latent space size in original paper, I have very blurry results that cannot be considered descent counterfactual. 30

I should stick to a larger latent space size and try to center on the brain better in order to try 128x128x128. Nevertheless, it means that the entire brain cannot be processed at once.