A. Criminisi, J. Shotton, S. Bucciarelli and K. Siddiqui

http://research.microsoft.com/en-us/projects/medicalimageanalysis/

One-click visual navigation

Better visualization (class-driven col. transfer functions)

Initialization for organ-specific processing

Content-driven image search

Applications

One-click visual navigation

Setting up the ideal 3D view for diagnosing problems with heart valves is laborious.

Applications One-click visual navigation

Better visualization (class-driven col. transfer functions)

Initialization for organ-specific processing

Content-driven image search

If we know where the liver is then we can start an automatic process for detecting calcifications.

Applications One-click visual navigation

Better visualization (class-driven col. transfer functions)

Initialization for organ-specific processing

Content-driven image search

If we know where the liver is then we can start an automatic process for detecting calcifications.

Applications One-click visual navigation

Better visualization (class-driven col. transfer functions)

Initialization for organ-specific processing

Content-driven image search

No contrast agent

…

Considerable geometric variations. Conventional atlas-based techniques would not work.

Labelling via axis aligned 3D bounding boxes.

Classes = heart, liver, l. kidney, r. kidney, l. lung, r. lung, l. eye, r. eye, head, background

Positive and negative training examples for organ centres.

Node optimization functionNode optimization function

Training a single tree

class

class

class

class

class

S

S1 S2

During training each node “sees” only a random subset of all available features

Each tree is training independently, using the same procedure

Posterior output of classifierPosterior output of classifier

Organ detectionOrgan detection

Organ localizationOrgan localization

Testing

Using multiple trees has been shown to improve generalization.

Context-rich visual features, a 2D illustration

Feature response Feature response

Lots and lots of randomly generated features. Out of those the most discriminative ones are selected automatically during training.

Long-range spatial context is captured bythe displaced integration regions.

Results of automatic organ detection and localization for three different patients.

• Our algorithmOur algorithm

• Gaussian Mix. ModelGaussian Mix. Model

• Template matchingTemplate matching

(multiple runs on multiple train/test (multiple runs on multiple train/test splits)splits)

More anatomical structures

Hierarchical -> Finer structures

Spatial priors for greater robustness to noise

Larger training database

http://research.microsoft.com/en-us/projects/medicalimageanalysis/