Help with improving the training of an alternative atlas for medial temporal lobe subregions
External Email - Use Caution
Hi all, We are aiming at training an alternative atlas of some subregions in the medial temporal lobe using FreeSurfer. As training examples, we manually segmented the subregions using tkmedit and tksurf. In this process we deleted the original segmentation labels around our new regions of interest (see Figure 1). Then, we trained the atlas and parcellated new data as follows: # Train atlases for hemi in lh rh; do mris_ca_train -n 2 \ -t ./doc/LUT.txt \ ${hemi} \ sphere.reg \ CUSTOM_ROIs.annot \ S01 S02 S03 S04 S05 S06 \ ./atlas/${hemi}.CUSTOM_ROIs.gcs done
# Create annotations for hemi in lh rh; do for ID in S10 S11 S12 S13; do mris_ca_label \ -orig white \ -novar \ -t doc/ LUT.txt \ ${ID} \ ${hemi} \ sphere.reg \ ${hemi}.CUSTOM_ROIs.gcs \ $SUBJECTS_DIR/${ID}/label/${hemi}.CUSTOM_ROIs.annot done done
[cid:image001.png@01D89C87.200FE420]This lead to a very inaccurate parcellation (see Figure 2), also when we filled the unlabeled region with a dummy label (see Figure 3).
Figure 1. Example for manual segmentation of the 4 subregions (encircled) with deleted original segmentation labels around.
[cid:image003.png@01D8A190.FBE5BED0]
Figure 2. Parcellation example using the trained atlas without dummy label for unlabeled regions.
[cid:image006.png@01D8A192.42B1DF00]
Figure 3. Parcellation example using the trained atlas with dummy label for unlabeled regions.
Any comments or suggestions on how to deal with our problem would be greatly appreciated. Many thanks for having taken the time to read and reflect Best, Nicolas
Hi Nicolas I'm not sure what is going wrong as your basic procedure looks correct. I would start by building a gcs from a single subject, then labeling that same subject and seeing if you recover the manual label (more-or-less)
Cheers Bruce
From: freesurfer-bounces@nmr.mgh.harvard.edu freesurfer-bounces@nmr.mgh.harvard.edu On Behalf Of Henzen Nicolas Sent: Wednesday, July 27, 2022 2:24 AM To: 'freesurfer@nmr.mgh.harvard.edu' freesurfer@nmr.mgh.harvard.edu Subject: [Freesurfer] Help with improving the training of an alternative atlas for medial temporal lobe subregions
External Email - Use Caution Hi all, We are aiming at training an alternative atlas of some subregions in the medial temporal lobe using FreeSurfer. As training examples, we manually segmented the subregions using tkmedit and tksurf. In this process we deleted the original segmentation labels around our new regions of interest (see Figure 1). Then, we trained the atlas and parcellated new data as follows: # Train atlases for hemi in lh rh; do mris_ca_train -n 2 \ -t ./doc/LUT.txt \ ${hemi} \ sphere.reg \ CUSTOM_ROIs.annot \ S01 S02 S03 S04 S05 S06 \ ./atlas/${hemi}.CUSTOM_ROIs.gcs done
# Create annotations for hemi in lh rh; do for ID in S10 S11 S12 S13; do mris_ca_label \ -orig white \ -novar \ -t doc/ LUT.txt \ ${ID} \ ${hemi} \ sphere.reg \ ${hemi}.CUSTOM_ROIs.gcs \ $SUBJECTS_DIR/${ID}/label/${hemi}.CUSTOM_ROIs.annot done done
[cid:image005.png@01D8A1A0.79D98F10]This lead to a very inaccurate parcellation (see Figure 2), also when we filled the unlabeled region with a dummy label (see Figure 3).
Figure 1. Example for manual segmentation of the 4 subregions (encircled) with deleted original segmentation labels around. [cid:image001.png@01D8A1A0.79E10920]
Figure 2. Parcellation example using the trained atlas without dummy label for unlabeled regions. [cid:image002.png@01D8A1A0.79E10920]
Figure 3. Parcellation example using the trained atlas with dummy label for unlabeled regions.
Any comments or suggestions on how to deal with our problem would be greatly appreciated. Many thanks for having taken the time to read and reflect Best, Nicolas
you can also try train-gcs-atlas (a front end for mris_c_train that might make things easier)
On 7/27/2022 10:06 AM, Fischl, Bruce wrote:
Hi Nicolas
I’m not sure what is going wrong as your basic procedure looks correct. I would start by building a gcs from a single subject, then labeling that same subject and seeing if you recover the manual label (more-or-less)
Cheers
Bruce
*From:* freesurfer-bounces@nmr.mgh.harvard.edu freesurfer-bounces@nmr.mgh.harvard.edu *On Behalf Of *Henzen Nicolas *Sent:* Wednesday, July 27, 2022 2:24 AM *To:* 'freesurfer@nmr.mgh.harvard.edu' freesurfer@nmr.mgh.harvard.edu *Subject:* [Freesurfer] Help with improving the training of an alternative atlas for medial temporal lobe subregions
* External Email - Use Caution *
Hi all,
We are aiming at training an alternative atlas of some subregions in the medial temporal lobe using FreeSurfer.
As training examples, we manually segmented the subregions using tkmedit and tksurf. In this process we deleted the original segmentation labels around our new regions of interest (see Figure 1). Then, we trained the atlas and parcellated new data as follows:
# Train atlases
for hemi in lh rh; do
mris_ca_train -n 2 \
-t ./doc/LUT.txt \
${hemi} \
sphere.reg \
CUSTOM_ROIs.annot \
S01 S02 S03 S04 S05 S06 \
./atlas/${hemi}.CUSTOM_ROIs.gcs
done
# Create annotations
for hemi in lh rh; do
for ID in S10 S11 S12 S13; do
mris_ca_label \
-orig white \
-novar \
-t doc/ LUT.txt \
${ID} \
${hemi} \
sphere.reg \
${hemi}.CUSTOM_ROIs.gcs \
$SUBJECTS_DIR/${ID}/label/${hemi}.CUSTOM_ROIs.annot
done
done
This lead to a very inaccurate parcellation (see Figure 2), also when we filled the unlabeled region with a dummy label (see Figure 3).
Figure 1. Example for manual segmentation of the 4 subregions (encircled) with deleted original segmentation labels around.
Figure 2. Parcellation example using the trained atlas without dummy label for unlabeled regions.
Figure 3. Parcellation example using the trained atlas with dummy label for unlabeled regions.
Any comments or suggestions on how to deal with our problem would be greatly appreciated. Many thanks for having taken the time to read and reflect
Best,
Nicolas
Freesurfer mailing list Freesurfer@nmr.mgh.harvard.edu https://mail.nmr.mgh.harvard.edu/mailman/listinfo/freesurfer
freesurfer@nmr.mgh.harvard.edu
-
Douglas N. Greve -
Fischl, Bruce -
Henzen Nicolas