Image matching approaches

[ChristophePhillips]

Image matching approaches

There are different ways to evaluate the match or overlap between 2 binary images, typically a proposed image (the segmentation we want to check) and a reference image (the segmentation we think is the truth). The methods described here under should be implemented in a function, such as image_overlap.m first introduced by @CPernet .

Here are a few possible measures:

• some direct and global 'image overlap' measure, e.g.Dice score or Jaccard index. Easy to calculate from 1/0 images and provides a measure of the whole volume agreement
• Standard sensitivity/specificity measures as derived from the confusion matrix (true/false positive/negative). These are not very useful because the tumor area is relatively small compared to the healthy volume, leading to over-estimated specificity value.
• corrected correlation/sensitivity/specificity measures that accounts for the unbalancedness in the data (much larger healthy volume than lesion), e.g. Cohen's kappa or Matthews correlation coefficient
• in case the lesion is split into multiple separate regions, the "True/False Positive Rate" (TPR/FPR) is defined as the ratio between the number of lesions, i.e. counted as regions NOT voxels, in the segmentation that do/do not overlap (by at least one voxel) with the reference image.
• surface matching is also useful, e.g the (average) Hausdorff distance. The Euclidean distance between each superficial voxel in a segmented region and that of its counter part in the reference image is calculated. These distance segment-reference and/or reference-segment can be averaged to give a global surface match measure.

Are currently available within the image_overlap.m:

• the modified Jaccard index
• true/false positive/negative
• Matthews correlation coefficient

References for segmentation comparison

• Menze et al., The Multimodal Brain Tumor Image Segmentation Benchmark (BRATS). IEEE TMI, 2015.
• Styner et al., 3D Segmentation in the Clinic: A Grand Challenge II: MS lesion segmentation, MICCAI 2008 Workshop.

[ChristophePhillips]

There is also another small issue about the brain volume definition: whenever we count the number of true negatives (TNs), we should count them over the brain volume and not the whole image! Otherwise we are sing an even larger number of TN voxels...
-> allow the possibility to enter a brain mask? This one could be derived during the processing of the images and provide a more reasonable volume over which to evaluate the match.

[CPernet]

Can we just input your ICV Image derived during normalization ?

Dr Cyril Pernet
Senior Academic Fellow
CCBS / Edinburgh imaging

Sent from my HTC mobile phone

----- Reply message -----
From: "Christophe Phillips" [email protected]
To: "CyclotronResearchCentre/USwLesion" [email protected]
Cc: "PERNET Cyril" [email protected], "Assign" [email protected]
Subject: [CyclotronResearchCentre/USwLesion] Image matching approaches (#3)
Date: Tue, May 31, 2016 15:11

There is also another small issue about the brain volume definition: whenever we count the number of true negatives (TNs), we should count them over the brain volume and not the whole image! Otherwise we are sing an even larger number of TN voxels...
-> allow the possibility to enter a brain mask? This one could be derived during the processing of the images and provide a more reasonable volume over which to evaluate the match.

You are receiving this because you were assigned.
Reply to this email directly, view it on GitHubhttps://github.com//issues/3#issuecomment-222700329, or mute the threadhttps://github.com/notifications/unsubscribe/AEjUDl1Apt07J5C0fFpy9K9-M9eeuxg-ks5qHEFcgaJpZM4Iqe8a.

The University of Edinburgh is a charitable body, registered in
Scotland, with registration number SC005336.

[ChristophePhillips]

Yep, that's exactly the idea! :-)
But it should be available in subject space...