MIA_Atlas_Patch_based_Segmentation

Main idee of atlas-patch based segmentation?

The use if registration to align to an atlas (which is pre segmented) à segmentation of target structure

Challenges of atlas-based registration?

You need a working registration Needs calculation time à clinical process things should go fast What you see on atlas and what you see on patient à if you have an disease, anatomy will change à so aligning two images, that may not have a full correspondence in terms of structure Atlas is average brain à some patients have a very different anatomy à trying to register two different images

What is the aim of an multi atlas?

-       Cover more of the anatomical differences in a population Want to cover ages à change in anatomic (some structures decreasing)

What are methods to fuse data for multi atlas segmentation?

Global fusion strategies Majority voting (for example 10 atlas then you look for each pixel what the majority have for that pixel Weighted voting à if atlas is more reliable Local fusion strategies: à looking to patches around pixels Locally weighted fusion Simultaneous truth and performance level estimation (STAPLE) Shape based averaging à Based on averaging the distance transforms of the segmentations

Challenge of multi atlas- segmentation?

Very time consuming à if you have 100 atlas images you need to do 100 registrations If you just have 5 atlases, the majority voting is not good à false voting because of 2 against 3 à

What is STAPLE?

Simultaneous truth and performance level estimation Probabilistic approach Try to predict the underlying ground truth segmentation Try to estimate the performance of the predictions Looks for: Sensitivity p: true positive fraction Specificity q: true negative fraction EM algorithm is used to estimate the underling ground truth and to estimate the rating of the experts

Basic ideas of patch-based segmentation

-       only coarse registration required (e.g. rigid or affine registration) use the same concept as label fusion (different patches, find the similarities, weight them and fuse them)

Concept of patch- denoising?

The final intensity is a weighted average of the intensity across the different images à weight is going to relate on the similarities Uses a neighborhood averaging strategy called non-local means (NLM) assumes that every patch in a natural image has many similar patches in the same image. Weight function: looks how different intensities are in a patch Similar intensities should have similar labels

The main idea of atlas-based segmentation is to…. Use the atlas as reference intensity values Segment an image via registration to an atlas Segment and image and compare labels of an atlas Use the atlas to initialize a voxel-wise segmentation

Segment an image via registration to an atlas

The direction of the atlas-based registration is…. Atlas is registered to the image Image is registered to the atlas Both ways and then average the transformation

Atlas is registered to the image à atlas is moved to the image space à segmentation on image space

Fusion is used in multi-atlas-based segmentation in order to… Fuse atlases and then segment Remove outliers in the segmentation Fuse individual atlas results

-       Fuse individual atlas results Remove outliers is not completely wrong but not the main reason

STAPLE algorithm produces: Most probable true segmentation and performance metrics of segmentation Fuses segmentation by geometrical average Most probable true segmentation and dice estimations

Most probable true segmentation and performance metrics of segmentation

-       The basic concept of patch-wise segmentation is…. Averaging labels over a patch Assign similar labels from similarity of voxel intensity Assign labels based on the variability within a patch Assign labels as a random selection within a patch

Assign similar labels from similarity of voxel intensity