Mutual Information as a Measure for Image Quality of Temporally Subtracted Chest Radiographs...

Preview:

Citation preview

Mutual Information as a Measure for Image

Quality of Temporally Subtracted Chest

Radiographs

Samantha PassenSamuel G. Armato III, Ph.D.

Introduction Commonly, radiologists compare multiple

chest radiographs side-by-side

Current Previous

Introduction Kano et al. (1994)- Temporal Subtraction

Detect ribcage edges and denote “lung mask”

Current

Introduction ROIs involved in nonlinear geometric warping

to align previous image to current

Current Warped Previous

Introduction Temporal Subtraction

Image

Related Work Difazio et. Al (1997) demonstrated improved

radiologist diagnostic accuracy with temporal subtraction images

Ishida et. Al. (1999) used local cross-correlation method to maximize alignment

Armato et al. (2006) – automated identification of registration accuracy Feature-based linear discriminant analysis Based on radiologist ratings of images

Motivation

While temporal subtraction images effectively enhance areas of pathologic change, misregistration of the images can mislead radiologists in diagnosis by obscuring or creating interval change

Mutual information as a metric to quantify misregistered cases

Related Work

Mutual Information confirmed to:

Coselmon et al. (2004)- register volumetric image data

Sanjay-Gopal et al. (1999) – register mammograms

Pluim et al. (2003) – transformation technique to align images

Mutual Information (MI) Joint Histogram:

Lower Entropy------------------Higher Entropy

(misregistration) Mutual Information:

Materials Radiologists rated 138 temporal

subtraction images from 1.0-5.9

Rating= 1.0 Rating= 5.8

Previous Methods

Calculate the correlation of the two radiologists’ ratings= 0.785

Calculated correlation coefficient of NMI values and radiologist’s ratings Correlation = 0.649

Good Rating: 5, Bad MI: 1.135

• Clear difference between the two images, not due to misregistration but to interval change

Good Rating: 5, Bad MI: 1.135

• Clear difference between the two images, not due to misregistration but to interval change

Motivation for New Data•Calculate the NMI on portions of the bottom removed

•Pathologic change•Positioning of the body affects diaphragm •Inaccurately defining inferior bottom ribcages

Previous Key Results

0.500

0.550

0.600

0.650

0.7000.750

0.800

0.850

0.900

0.950

1.000

0% 10% 20% 30% 40% 50% 60%Percent Lung Mask Cropped

Co

rre

lati

on Full Resolution

256 Gray Levels

128 Gray Levels

64 Gray Levels

32 Gray Levels

• Maximum correlation = 0.785

New Methods Same radiologist

rated left and right lungs on subtraction image separately

Calculate NMI on right and left lung

Results- Correlation Coefficient

Right max: 0.746 Left max: 0.752 Average max: 0.782

Averaged Lung NMI with Averaged Rating

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

0 10 20 30 40 50 60

Percent Cropped

Co

rre

lati

on

Co

eff

icie

nt full resolution

256 gray levels

128 gray levels

64 gray levels

32 gray levels

New Methods Randomly divide 69 patients into 2 sets

Training Set: 34 patients, ~66.5 pairs of images

Testing Set 35 patients, ~71.5 pairs of images 20 different trials Calculated NMI on all 35 combinations of

parameters of testing set Determine correlation coefficient Choose 3 trials with maximum correlation

coefficient

New Method

Apply these parameters to testing sets and calculate: Correlation coefficient Calculate predicted rating

Use regression line from training set and substitute MI value from testing set

ROC analysis

Results

  0% 10% 20% 30% 40% 50% 60%

Full Resolution - - - - - - -

256 Gray Levels - - - - - - -

128 Gray Levels - - - - - 2 3

64 Gray Levels - 1 - - 2 10 10

32 Gray Levels - 1 1 - 4 11 15

Results

Sensitivity = TP/(TP + FN)

Specificity = TN/(TN+FP)

TP = Calculated Rating < 3, True Rating < 3

TN = Calculated Rating ≥ 3, True Rating ≥ 3

Results

  Specificity SensitivityCorrelation Coefficient

Max 0.936 0.850 0.864

Min 0.696 0.440 0.632

Average 0.851 0.667 0.785

ResultsConventional Binormal ROC

Curves

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

0 0.5 1False Positive Fraction

Tru

e P

osi

tive

Fra

ctio

n

60%, 32 GrayLevels, Az = 0.909

60%, 32 GrayLevels, Az= 0.900

50% 32 GrayLevels, Az= 0.915

New Method Calculate normalized cross-correlation to

compare usefulness of MI technique

Only compute 1 cross-correlation value for each pair of image when directly aligned

Results

All cross-correlation values range from 0.999-1.0

Correlation with Radiologist’s ratings = 0.035 – 0.180

No Information Gained

Conclusion

Successfully demonstrated a correlation between MI and radiologist evaluation

Calculating the NMI on the top 50% of the lung mask and scaling to 128 bins has a correlation of 0.785, comparable to that of the two radiologists

Conclusion

Maximum Az value for 60 testing sets = 0.915

For training set, cropping 50% of the lung mask and scaling to 32 gray levels maximum correlation and Az

Mutual information gives complimentary information to that of cross-correlation

Future Work

Mutual information can be incorporated into existing temporal subtraction algorithm

Calculate NMI on warped previous and current images

Determine if predicted rating < 3 Re-warp or inform radiologists

Questions?

Recommended