MemoCard lernen mit System

Why semi-automatic approaches are used for segmentation?

Practice semi-automatic approaches shouldn't be discarded as the tradeoff between robustness and operator’s time is for many clinical scenarios very convenient. In the next we will be discussing one such semi-automatic approach

Applying a max operator to a distance map enables us to have a….

Incredibly good segmentation result

An awesome volumetry of the structure

A skeleton representation of the structure

A hausdorff distance estimator

- A skeleton representation of the structure

How image segmentation can be validated?

Digital phantoms Acquisition and careful segmentation Autopsy/histopathology Clinical data

What does the semi-automatic segmentation method?

- Finds the optimal path between starting and ending point Optimal path = minimal cost between points = cumulative sum between local segments Not only considers the intensity of pixels but also incorporates priors on the type of contours it yield. Main concept is to cast the segmentation problem as an optimization problem, where the object function to optimize is a minimal cost bath between start/ end point defined by the user

Why not just automatic segmentation methods are used?

- Lack of robustness

Dicom tags contain:

- Personal patient information only

- Image information only

- Manufacture Information only

- All above and more

Advantages of live-wire segmentation:

- Monitor segmentation Correct on the fly No fix-path in advance Speed over other path searching approaches. Enables real time

Orientation information is helpful since it allows to…..

Know the manufacture type

Understand what is left and right, anterior and superior

Defined pixel size

- Understand what is left and right, anterior and superior

Live wire is:

User-assisted minimal path finding algorithm

A fully automated segmentation algorithm

- User-assisted minimal path finding algorithm

Goal of segmentation:

Create labeled image from (maybe several) input images. The label image can represent a semantic separation or classification of pixels.

Different segmentation methods

Simple methods (Thresholding, Region-Growing) Classification + Clustering (based on machine learning and pattern recognition techniques) Deformable models / level sets (shape prior, but not very rigorous, solely based on some geometric properties like curvature) Active shape models (very strong shape prior, based on actual geometry of the structure to be segmented à model-based method) Random Fields/graph-cuts (one of the most recent methods)

What differentiate the methos for segmentation?

What mostly differentiates these methods is in the amount of prior knowledge they use to yield a segmentation result Most advanced methods introduce a notion of the structure being segmented, as well as local or global properties that the final segmentation should have.

Name a semi-automatic segmentation algorithm

- Live-wire segmentation

How many elements are needed to translate from pixel coordinates to real world coordinates:

Depends on the property of orientation

3 Image Origin; m6ust not always be (0,0,0); (0,0)

Image dimension

Voxel/Pixel spacing (voxel/pixel)

size

More live-wire segmentation:

Classification of Voxels:

Different voxels to different structures depending on their intensity value

Info cartes d'apprentissage

MIA_Segmentation_Intro