Title: Deconvolution of noisy radiological images
1Deconvolution of noisy radiological images
2Medical image processing
- Image formation described by Poisson statistics
- Signal-to-noise ratio increases with radiation
dose - Radiological images are obtained under
conflicting requirements - ALARP principle
- X-ray machine design and performance
considerations - Radiological images are obtained in an inherently
noisy environment - Enhancement or Restoration?
- Spatial or frequency filtering, wavelets,
deconvolution methods - Restoration methods derive from an image
degradation model and attempt to undo the
effects blurring and noise.
3Focal spot size, magnification and blur
4Application 1 Mammography
- Focal spot size determines the blurring
- Fine focal spot produces less blurring, but
- Slower tube heat dissipation so smaller currents
are used - Longer exposures and chance of patient motion
blur - Geometry determines the image magnification
- Higher magnification allows detection of smaller
microcalcifications, but - More image blurring
- Modern X-ray units have evolved to live with
these conflicting design and performance
requirements. - Alternative Deblurring, to obtain images as good
as those which would be obtained with a perfect
(ideal) focal spot
5The degradation/restoration model
True image
Observed image
Restored image
Geometric Unsharpness
Quantum noise
In spatial domain
In frequency domain
6The deconvolution problem
Solution?
7(No Transcript)
8The deconvolution problem
Solution?
If H?0 then the restored image is devastated with
artefacts, even when H is well defined.
The answer is to use regularised deconvolution,
where constraints are placed on the acceptable
restoration.
9Maximum Entropy Method
- Maximum entropy method (MEM) is a suitable
deconvolution technique.
- Guess a solution. Generate trial data using
the forward map. - Chi-Squared Goodness of Fit statistic for the
observed data and trial data. - When this is minimised then we have a good fit
between our trial and observed images? - No! It is equivalent to solving the matrix
equation
h is ill conditioned for a given g there are
many not necessarily close together.
10Maximum Entropy Method
- We impose a constraint on the solution Choose
the with the maximum entropy, S.
- (Choice of the entropy function is a result of
treating the restoration process as a statistical
inference problem). - Numerical procedure is to minimise the objective
function
11Mammography experiments
- Obtain images with conventional set-up
- 1.8 magnification, fine focal spot
- Typical Radiographic factors (28kV, 40mAs)
- Realistic scatter
- Obtain images taken with unconventional set-ups
- 1.8 magnification, broad focal spot
- 3.0 magnification, fine focal spot
- These unconventional modes are not used in
practice blurring introduced is unacceptable - Use MEM to restore these images
- Can we make 1.8 BF images as good as 1.8 FF?
- Can we reduce blurring in the 3.0FF case?
12TORMAM Phantom Scoring
13The system PSF
Brass foil with pinholes
141.8FF Conventional Setup
151.8BF Not used in practice
16Pixel luminance profiles
171.8BF as good as 1.8FF?
183.0FF Better shape resolution
19Image scores
Image features scored with 3,2,1,0 depending on
perceived visibility
Average scores of two experienced, independent
observers
20Discussion
- Better visibility of all features under all three
set-ups. - Improvement in the signal-to-noise ratio
- Improvement in spatial resolution
- General acceptance amongst radiologists that
traditional (Fourier based) de-blurring
techniques are of little value - MEM deconvolution addresses issues of noise
amplification and artefact introduction - MEM offers a way of weakening the link between
focal spot size and geometric blurring
21Paper
- Collaboration with RMPG at Newcastle General
Hospital - Paper submitted to Physics in Medicine and
Biology in May 2004
22Application 2 Linear tomography
- X-ray tube and image receptor have linear but
opposing movements - Only a focal plane remains in focus regions
above and below are blurred - High pass filtering will remove blur (and low
frequencies in the focal plane) - Deconvolution models are a better way forward
23Linear tomography machine
24Simple 3-plane tomography model
Tomograph
Object
25Linear tomography model
26Simulated 3-plane restoration
275 planes through a skull phantom
28What next?
- Magnification mammography
- Repeat, using a more complicated, realistic
phantom - Experiments with reduced radiation doses (higher
levels of noise)
- Linear tomography
- Different image priors?
- Post processing with high frequency filters?
- Deconvolution of scatter
- To improve radiographic contrast
29The End