External Email - Use Caution

Hello,

I am testing SAMSEG on my group of MS patients. Unfortunately, even when dropping threshold to 0.005 I have large number of false negative lesions left behind by SAMSEG (80% in about 75 cases I tried), which are clearly visible lesions. I ran/checked registration between inputs, which is typically very good to excellent. Images are all 3D acquisitions , mprage, flair, and t2 with ~1 mm3 isotropic voxel. Here is the command, which I ran:

run_samseg --input mprage_reg.nii.gz flair.nii.gz t2_reg.nii.gz --pallidum-separate --lesion --lesion-mask-pattern 0 1 1 --threshold 0.005 --output samseg_threshold_0.005

and example of the lesion in occipital white matter in sagittal and axial orientations.

[image: image.png] [image: image.png]

I would very much appreciate any leads as to how to improve the segmentation.

Thanks,

Martin