Post on 22-Feb-2016
description
Lesion Penumbra Selection
Last Time
• Previously, Brian suggested using a z-score threshold of > 2 as perhaps a way to select lesion penumbra/DAWM
• We saw that this was effective and would at the very least reduce a lot of the work required by a radiologist reviewing the original lesion segmentation
• Not sure yet if Michael approves of the overall DAWM selection quality with this technique
Last Time
• We saw that all of the thresholding methods (before or after warping) yield very similar lesion boundaries– Notably, thresholding at 0.5 after warping the “> 4 masks”
edited by Hagen to patient space is very similar to the results with the more “review-board-safe” method of thresholding the z-score maps in patient space
– This similarity allows us to recover many of Hagen’s edits by applying the penumbra selection method to them
– We decided that warping the z-score maps made the most sense rather than warping the mean and standard deviation maps and calculating z-score patient space
Penumbra Selection Method
• Brute force approach – used for recovering Hagen’s edits– Input: 2 masks (lesion core and penumbra, or edited
>4 MNI and new >4 patient space) – Assign a number label to each contiguous region in
the masks (bwlabeln)– Go through all the labels in one mask and see if
there’s an entity in the other mask that overlaps with it• If there is, keep it; else, discard.
Penumbra Selection Method• Faster intelligent approach – used for keeping penumbra with
lesions cores– Can take advantage of the fact that >2 masks will entirely
encompass >4 masks– Assign a number label to each contiguous region in the masks
(bwlabeln)– For each labels in one mask, find the overlapping entity in the other
mask (this is no longer an if)• Keep it and subtract the entity from the first mask, i.e. remove all cores
contained in that penumbra• Then relabel the first mask and restart the for loop
– About a 2x speed increase, greater for patients with large abnormalities
Results – Penumbra Selection
• Color coding:– Yellow– core– Red – selected penumbra with cores– Green – left over penumbra
P009– Low CIS – FLAIR
P009– Low CIS – Penumbra
P015 – Low CIS – FLAIR
P015 – Low CIS – Penumbra
Results – Hagen’s Edits
• Color coding:– Red– new lesion core from thresholding in patient
space that overlapped with one of Hagen’s– Blue – this should be out-of-brain tissue
P009– Low CIS – FLAIR
P009– Low CIS – Hagen’s Edits
P015 – Low CIS – FLAIR
P015 – Low CIS – Hagen’s Edits
Discussion
• Overall, I think the selection method has merit• It does a decent job at carrying over Hagen’s
work but clearly there are some spots excluded that should not be
• Penumbra selection works as imagined, still need expert eyes to determine if it’s worthwhile
Future
• 3D visualizations?• A tool to easily choose which lesions are good
or bad with one-click