DEVOE-1

CSE5810

Research Topics in Computer Science and Medical Imaging

Gordon Devoe - 2014

DEVOE-2

CSE5810

A Nonlocal Maximum Likelihood Estimation Method for Rician Noise

Reduction in MR Images

Lili He & Ian GreenshieldsFebruary 2009

DEVOE-3

CSE5810

Magnetic Resonance (MR) Images MRI is the newest most versatile medical imaging

technique available (very minor fluctuations in chemical composition can be determined).

High resolution images of a patient’s interior body without surgery.

Strong magnets and pulses of radio waves used to manipulate the natural magnetic properties in the body.

Especially useful for capturing images of the brain, spine, organs and soft tissue.

DEVOE-4

CSE5810

Magnetic Resonance (MR) Images

DEVOE-5

CSE5810

Noise in Medical Resonance Images

Sources Thermal noise from the conductivity of the

system’s hardware. Inductive losses from the conductivity of the

object being images. Factor Trade-offs

Resolution Signal-to-noise ratio (SNR) Acquisition speed

DEVOE-6

CSE5810

Denoising MR Images

By denoising magnetic resonance (MR) images we can improve medical processes including: Clinical Diagnosis

the determination of the nature of a disease distinguishing one disease from another.

Tissue Classification Systematic arrangement of similar tissues on the

basis of certain differing characteristics. Image Segmentation

Coming Soon…

DEVOE-7

CSE5810

Research Topics In Medical Imaging Image Registration Image Segmentation

DEVOE-8

CSE5810

Image Registration Methods: A Survey

Zitová, Barbara and Flusser, Jan. Image and Vision Computing, ISSN 0262-8856, 2003,

Volume 21, Issue 11, pp. 977 - 1000

DEVOE-9

CSE5810

Image Registration Image registration is one of the most fundamental

concepts when using computer science techniques to compare images with one another.

Image registration is the process of overlaying two or more pictures/images of the same scene and comparing them from different viewpoints, times and/or by different sensors.

The intersection of these images involves a set of changes influenced by different imaging conditions or various data sources such as image fusion, change detection and multi-channel image restoration.

DEVOE-10

CSE5810

Image Registration - Applications Multispectral Classification Environmental Monitoring Change Detection Image Mosaicking Weather Forecasting Creating super-resolution images Geographic information systems (GIS) Cartography Computer Vision

DEVOE-11

CSE5810

Image Registration – Applications Medical Comparing nuclear magnetic resonance spectra data

(NRM) with computer tomography (CT) data. Monitoring tumor growth Comparing patient data with anatomical atlases

(digital libraries of anatomy information). Brain mapping.

DEVOE-12

CSE5810

Image Registration – 3 Types Multi-Temporal Image Registration

Images captured at different times/conditions Evaluate changes in the scene Tumor evolution monitoring Healing therapy

Multi-Modal Image Registration Images from same scene are captures with sensors

with different characteristics. Integrate the information obtained from different

source streams to gain more detailed scene representation.

Combine MRI, ultrasound or CT, magnetic resonance spectroscopy for radiotherapy & nuclear medicine applications.

DEVOE-13

CSE5810

Image Registration – 3 Types Scene-to-model Image Registration

Images of a scene and models are compared. Models are computer generated and images are

overlaid on the computer model. Comparing a patient’s image with a digital model.

DEVOE-14

CSE5810

Image Registration - Steps Feature Detection

Salient and distinctive objects (close-boundary regions, edges, contours, edges, lines, corners, etc.)

Features are then represented by their controls points.

Feature Matching Correspondences between the features detected in

the sensed and reference image are established. Transform Model Estimation

Type and parameters of mapping-functions aligning the sensed/reference image are established.

Image Resampling and transformation Sensed image is transformed based on the

established mapping functions

DEVOE-15

CSE5810

Image Registration - Steps

Feature Detection

Feature Matching

Transform Model Estimation

DEVOE-16

CSE5810

Image Registration – Feature Detection Features

Salient structures Regions (forests, lakes, fields) Lines (region boundaries, coastlines, roads, rivers) Points (region corners, lines intersections, points on

curves) Stable in time and stay in fixed position Distinct and spread all over the image.

Features - Medical Domain Line detection to compare anatomical structures. Interactive selection. Introduction of extrinsic features (screw markers,

dental adapters, etc.)

DEVOE-17

CSE5810

Image Registration – Feature Matching Sensed and references images can be compared using

Intensity values Feature Spatial Distributions Feature Symbolic Descriptions

Feature Mapping can be divided in to two main categories: Area-based Feature Detection Feature Based Feature Detection

DEVOE-18

CSE5810

Image Registration – Area Based Correlation “Template Matching”

Image intensities are directly matched without any structural analysis.

Sensitive to noise, varying illumination, and/or by using different sensor types.

Useful for real-time applications and easy to implement hardware.

figure, imshowpair(peppers(:,:,1),recovered_onion,'blend')

DEVOE-19

CSE5810

Image Registration – Area Based Mutual Information

Statistically dependency between two data sets. Comparing anatomical and sensed images of

patient’s body.MI criterion (bottom) computed in the neighborhood of point P. Maximum of MI shows correct matching position (point A). Point B indicates the false matching position selected by human operator.

DEVOE-20

CSE5810

Image Registration – Transform Model Est Once features have been discovered in the

sensed/referenced images a mapping function can be constructed to overlay the sensed image onto the reference image.

The mapping function is designed in a way to ensure the correspondence of the control points from the sensed/reference images are as close as possible.

Global mapping models use all control points for estimating one set of mapping function parameters for the entire image.

Local mapping models break the images in to sections making the function parameters depend on the location of their support in the image and the mapping function is defined for each section separately.

Changes in medical images tend to appear locally.

DEVOE-21

CSE5810

Image Registration – Transform Model Est

Similarity Transform (top left) Affine Transform (top right) Perspective Projection (bottom left) Elastic Transform (bottom right)

DEVOE-22

CSE5810

Current Methods in Medical Image Segmentation

Dzung L. Pham, Chenyang Xu, Jerry L. PrinceJohn Hopkins University

Department of Electrical & Computer Engineering

DEVOE-23

CSE5810

Image Segmentation

What is image segmentation?

Images are divided up in to smaller pieces or groups of pixels.

Segments are used to simplify or change the original image to a format easier to analyze.

Used to identify objects, lines, curves, etc. within an image.

Non-Trivial problem - noise, missing data, overlap of features, etc.

DEVOE-24

CSE5810

Image Segmentation - Medical

In the medical domain image segmentation is extremely important when using computers to facilitate the growing number of images in biomedical-imaging applications.

Quantify tissue volumes. Reconstruction of soft tissues such as the cerebral

cortex. Study of anatomical structures. Treatment planning Computer-integrated surgery Growing number of use cases…

DEVOE-25

CSE5810

Image Segmentation - Methods

Image segmentation can be roughly divided in to eight categories:

Thresholding Approaches Region Growing Approaches Classifiers Clustering Approaches Markov Random Fields Models (extension) Artificial Neural Networks Atlas-Guided Approaches

DEVOE-26

CSE5810

Image Segmentation - Thresholding

Image segmentation can be achieved by assigning an intensity value to every pixel of an image. Pixels can then be categorized based on whether or not they meet this threshold value.

DEVOE-27

CSE5810

Image Segmentation - Thresholding Medical Uses

Digital Mammography Limitations

Only two classes are generated (cancerous/healthy).

Cannot be applied to multichannel images. Doesn’t take in to account spatial characteristics. Sensitive to noise and intensity inhomogeneities

DEVOE-28

CSE5810

Image Segmentation – Region Growing

A ‘seed point’ is used to extract pixels from a particular part of the image and then transformations are made to the seed to perform further analysis. Separate regions that have the same properties. Provide original images with good segmentation

results. Choose multiple criteria at the same time.

DEVOE-29

CSE5810

Image Segmentation – Region Growing

Original Figure 155 ~ 255

190 ~ 255 255 ~ 255

DEVOE-30

CSE5810

Image Segmentation - Clustering

Image segmentation using clustering algorithms is very similar to segmentation classification methods except clustering algorithms do not require any training data.

The K-means algorithm iteratively computes a mean intensity for each class and classifies each pixel of the image to the class with the closest mean.

DEVOE-31

CSE5810

Image Segmentation - Clustering Advantages

Do not require training data (self-learning). Lack of spatial modeling provides fast

computation. Disadvantages

Sensitive to noise and intensity inhomogeneities. Do not directly support spatial modeling.

Potential Algorithms K-means Algorithm Fuzzy c-means Algorithm Expectation-maximization Algorithm

DEVOE-32

CSE5810

Image Segmentation – MRF Models Statistical model that can be used within segmentation

methods such as clustering. Models spatial interaction between nearby pixels.

K-means Segmentation With MRF

DEVOE-33

CSE5810

Image Segmentation – MRF Models

Medical images are a good candidate for Markov Random Field (MRF) models because pixels are generally in the same area as others from the same class and are widely used in digital mammograms.

• A depends on B and D• B depends on D and A• C Depends on E• Etc.

DEVOE-34

CSE5810

Image Segmentation – Deformable Models Deformable models delineate region boundaries by

using closed parametric curves or surfaces. These curves and surfaces deform when under the

influence of internal or external forces

DEVOE-35

CSE5810

Image Segmentation – Deformable Models

Advantages

Ability to directly generate closed parametric curves or surfaces from images.

Incorporation of a smoothness constraint that provides robustness to noise and spurious edges.

Disadvantages

Manual interaction is required to place initial model and choose appropriate parameters.

Deformable models exhibit poor convergence to concave boundaries.

DEVOE-36

CSE5810

Image Segmentation – Deformable Models

Deformable surface used for the reconstruction of the cerebral cortex.

Intersection between deformable surface and orthogonal slices of the MR image.

DEVOE-37

CSE5810

Image Segmentation – Atlas Guided When an atlas or template is available atlas-guided

image segmentation methods can be employed to anatomically separate features.

A template can simply be a collection of information about the anatomy of an image being segmented.

A Template can continually be re-used to segment other images of the same type.

By using an original image and the pre-segmented template a ‘warped’ image can be generated by using linear and non-linear transformations.

DEVOE-38

CSE5810

Image Segmentation – Atlas Guided

Template Image Target Image Warped Template

DEVOE-39

CSE5810

Image Segmentation – Atlas Guided

Three slices from a MR brain volume image overlaid with the atlas.

DEVOE-40

CSE5810

Image Segmentation – Future Work Future work

Improving accuracy. Incorporating prior information from atlases.

Increasing precision tolerances. Combine discrete & continuous spatial-domain

segmentation methods. Improve computational speed of segmentation

methods. Multiscale processing Parallelizable methods

Reduce amount of manual interaction.

DEVOE-41

CSE5810

Questions

DEVOE-42

CSE5810

ReferencesMedical Image Registration

Image registration methods: a survey by Zitová, Barbara and Flusser, Jan. Image and Vision Computing, ISSN 0262-8856, 2003, Volume 21, Issue 11, pp. 977 – 1000

Q. Chen, M. Defrise, F. Deconinck, Symmetric phase-only matched filtering of Fourier–Mellin transform for image registration and recognition, IEEE Transactions on Pattern Analysis and Machine Intellingence 16 (1994) 1156–1168.

Multimodality image registration by maximization of mutual information by Maes, F; Collignon, A; Vandermeulen, D; Marchal, G; Suetens, P. IEEE transactions on medical imaging, ISSN 0278-0062, 04/1997, Volume 16, Issue 2, pp. 187 – 198

Medical Image Segmentation 

Current methods in medical image segmentation by Pham, D L; Xu, C; Prince, J L Annual review of biomedical engineering, ISSN 1523-9829, 2000, Volume 2, p. 315

http://en.wikipedia.org/wiki/Image_segmentationhttp://en.wikipedia.org/wiki/Region_growinghttp://en.wikipedia.org/wiki/Markov_random_field

Magnetic Resonance Imaging (MRI) Noise Reduction A nonlocal maximum likelihood estimation method for Rician noise reduction in MR images by He, Lili and Greenshields, Ian R IEEE transactions on medical imaging, ISSN 0278-0062, 02/2009, Volume 28, Issue 2, pp. 165 - 172