33
MedIX – Summer 07 Lucia Dettori (room 745) [email protected] u

MedIX – Summer 07 Lucia Dettori (room 745) [email protected]

Embed Size (px)

Citation preview

Page 1: MedIX – Summer 07 Lucia Dettori (room 745) ldettori@cti.depaul.edu

MedIX – Summer 07

Lucia Dettori (room 745)

[email protected]

Page 2: MedIX – Summer 07 Lucia Dettori (room 745) ldettori@cti.depaul.edu

Projects

Contrast EnhancementA brief summary of what has been doneThings I would like to explore next

Texture classification Evaluations of segmentation algorithms

Page 3: MedIX – Summer 07 Lucia Dettori (room 745) ldettori@cti.depaul.edu

Broad goal – The big picture

Manipulate (medical) images to facilitate the radiologists’ job of recognizing features and pathologies in radiological images

Improve the visual quality of an image and automatically “highlight” certain features

Give them a way to focus on subsets of the image that are of interest to them

Page 4: MedIX – Summer 07 Lucia Dettori (room 745) ldettori@cti.depaul.edu

Contrast is all we “see”

Human eye identifies details by contrasting an object (foreground) and its background

Improve the quality of the image by creating (color) contrast

In our case we are talking about CT scan images with different levels of grey

Page 5: MedIX – Summer 07 Lucia Dettori (room 745) ldettori@cti.depaul.edu

Contrast enhancement

Take the gray level intensities of an image and proportionally redistribute them

Some mapping is necessary anyway since the images are based on 12 bits of information (gray levels ranges from 0 to 4095) and on these monitors we can only display 8 bits (gray levels from 0 to 255)

Can we do a better job?

Page 6: MedIX – Summer 07 Lucia Dettori (room 745) ldettori@cti.depaul.edu

Example

Page 7: MedIX – Summer 07 Lucia Dettori (room 745) ldettori@cti.depaul.edu

Techniques implemented

Linear binningLinearly redistribute the intensities from a

range of 0 - 4095 to a range of 0-255 over a chosen number of bins

Nonlinear binningFirst identify clusters of intensities then use

those to guide the redistribution of gray levels Histogram equalization

Page 8: MedIX – Summer 07 Lucia Dettori (room 745) ldettori@cti.depaul.edu

Soft tissues or bones?

Radiologist might be interested in only some parts of the image:Soft tissuesLungsBones

Each of these correspond to a different range of grey levels

Page 9: MedIX – Summer 07 Lucia Dettori (room 745) ldettori@cti.depaul.edu

Local contrast enhancement

Instead of trying to enhance the entire picture, concentrate the enhancing power in the range of intensities you are interested in

Window enhancement Multiple windows enhancement

Page 10: MedIX – Summer 07 Lucia Dettori (room 745) ldettori@cti.depaul.edu

Where are we now?

Previous students have created an application (C#) that allows the user to select windows and technique and display the enhanced image

Page 11: MedIX – Summer 07 Lucia Dettori (room 745) ldettori@cti.depaul.edu

Things I’d like to explore

Find an objective way to measure the improvement resulting from the contrast enhancement

See how these techniques perform when using different image modalities beyond CT-scans

Explore additional contrast enhancement techniques

Page 12: MedIX – Summer 07 Lucia Dettori (room 745) ldettori@cti.depaul.edu

Projects

Contrast Enhancement Texture classification

A brief summary of what has been doneThings I would like to explore next

Evaluations of segmentation algorithms

Page 13: MedIX – Summer 07 Lucia Dettori (room 745) ldettori@cti.depaul.edu

The big pictures

Given a pre-segmented organ region, can you tell me what it is: kidney, heart etc?

It depends … on its texture Identify image features that give texture

information Find rules that distinguish the texture

features of one organ from another

Page 14: MedIX – Summer 07 Lucia Dettori (room 745) ldettori@cti.depaul.edu

Texture Classification Process at a glance

Apply filterTo the image

Physician annotatedOrgan/Tissue

TextureDescriptors

Classifier(Decision Tree)

Classification rules for tissue/organs in

CT images

Liver Kidney Bone Spleen Heart

Page 15: MedIX – Summer 07 Lucia Dettori (room 745) ldettori@cti.depaul.edu

Image Data Set

Organ Qty Image Organ Qty Image

Kidney 223   Aorta 66  

Liver 260   IP Fat 59  

Spleen 95   Muscle 198  

Trabecular Bone 39   SQ Fat 157  

Lung 15  

Total Images

1112  

Page 16: MedIX – Summer 07 Lucia Dettori (room 745) ldettori@cti.depaul.edu

Step3 – Texture features extraction

Apply Gabor filters to the

image

TextureDescriptors

For example:Mean, standard deviation, energy, entropy etc..

Array of texture descriptors

[T1, T2, T3, …, Tn]

Physician annotatedOrgan/Tissue

Liver

Page 17: MedIX – Summer 07 Lucia Dettori (room 745) ldettori@cti.depaul.edu

Step4 - Classification

Apply Gabor filters to the

image

TextureDescriptors

Classification rules for tissue/organs in

CT images

The process of identifying a region as part of a class (organ) based on its texture properties.

Decisiontree

Predicts the organ from the values of the texture descriptorsTraining / Testing

Classificationperformance

measures

Physician annotatedOrgan/Tissue

Page 18: MedIX – Summer 07 Lucia Dettori (room 745) ldettori@cti.depaul.edu

Step5 – Evaluating the classifier

Page 19: MedIX – Summer 07 Lucia Dettori (room 745) ldettori@cti.depaul.edu

Apply Gabor filters to the

image

TextureDescriptors

Classification rules for tissue/organs in

CT images

Decisiontree

Things I would like to explore

Wedgelet transforsFractal Dimensions

Performancemeasures

Test Gabor texture descriptors on additional images and natural images

Physician annotatedOrgan/Tissue

Page 20: MedIX – Summer 07 Lucia Dettori (room 745) ldettori@cti.depaul.edu

Projects

Contrast enhancement Texture classification Evaluations of segmentation algorithms

Brief summary of what has been doneThings I would like to explore next

Page 21: MedIX – Summer 07 Lucia Dettori (room 745) ldettori@cti.depaul.edu

Texture segmentation Given an image, can you tell me

how many organs you have? That was easy enough. Can you

tell which organs they are?

Identifying regions with similar texture Identifying which texture it is to label the

organ

Page 22: MedIX – Summer 07 Lucia Dettori (room 745) ldettori@cti.depaul.edu

A couple of key questions

Can you do it better by varying a parameter? How do you choose the values of your segmentation parameters?

If it looks better is it really better?

Page 23: MedIX – Summer 07 Lucia Dettori (room 745) ldettori@cti.depaul.edu

A couple of key questions

Parameter optimization

Performance evaluation

Page 24: MedIX – Summer 07 Lucia Dettori (room 745) ldettori@cti.depaul.edu

1 3

4 2

0.87 0.56

0.50 0.75

Page 25: MedIX – Summer 07 Lucia Dettori (room 745) ldettori@cti.depaul.edu

Increasing value of a segmentation parameter

GroundTruth

Regionskey

Machine

Segmentations

How do I decide what the optimal value of the parameter is?

How good a segmentation is it?

Page 26: MedIX – Summer 07 Lucia Dettori (room 745) ldettori@cti.depaul.edu

The “goodness” metric

A single value that assigns a rating to a particular segmentation based on how well the machine segmented regions “match” the regions in the ground truth images

Page 27: MedIX – Summer 07 Lucia Dettori (room 745) ldettori@cti.depaul.edu

Region Categories

Ground Truth vs. Machine Segmented Correctly Detected Over Segmented Under Segmented Missed Noise

GT

MS

Page 28: MedIX – Summer 07 Lucia Dettori (room 745) ldettori@cti.depaul.edu

CORRECTLY DETECTED

OVER SEGMENTED

UNDER SEGMENTED

A Missed region is a GT region that does not participate in any instance of CD, OS, or US

A Noise region is an MS region that does not participate in any instance of CD, OS, or US

Index for each region

Page 29: MedIX – Summer 07 Lucia Dettori (room 745) ldettori@cti.depaul.edu

The “Goodness” Metric good = Correct Detection Index bad = 1-Correct Detection Index goodness = good-bad*weight

1.0

-1.0

Ceiling = CDind

Floor = 2*CDind-1

Weight Range = CDind-1

Page 30: MedIX – Summer 07 Lucia Dettori (room 745) ldettori@cti.depaul.edu

How can we use the metric?

Create a set of ground truth mosaic using radiologist-labels images of pure patches of organ tissues

Apply segmentation algorithm Optimize the segmentation parameters using the

metric Apply optimized algorithm to the “real” image

Page 31: MedIX – Summer 07 Lucia Dettori (room 745) ldettori@cti.depaul.edu

Ground Truth Region key

T=1000; GM= - .94

T=4000; GM= .74

T=3000; GM= .73T=2000; GM= - .02

T=5000; GM= .75 T=6000; GM= .08

Page 32: MedIX – Summer 07 Lucia Dettori (room 745) ldettori@cti.depaul.edu

Done so far

Used the metric on a block-wise walevet-based segmentation algorithm on some sample mosaic

Page 33: MedIX – Summer 07 Lucia Dettori (room 745) ldettori@cti.depaul.edu

To be done

Fully test the metric on a wide range of segmentation algorithms

Decouple the various components of the metric and test the individual performance measures instead of the overall score

Extend the metric to measure one region vs background segmentation