Statistical Image Reconstruction: 2D PET Simulation
Information Weighted Smoothing Splines
My group is developing statistical methods
for tomographic image reconstruction.
In collaboration with physicians
in the Nuclear Medicine Division
of the Radiology Department at UM,
one emphasis in my group is reconstruction methods
for PET and SPECT imaging.
These are functional imaging methods
that can provide unique diagnostic information
for detecting cardiac disease,
detecting cancer,
and monitoring the progress of treatment
for patients with these ailments.
Unfortunately, the signal to noise ratios
in PET and SPECT measurements
are often pretty poor
since the systems work by counting photons
one at a time as they are emitted
from trace amounts of radioactively-labeled tracers
within the patient.
The classical method for image reconstruction
is filtered backprojection.
This method is fast and simple,
but it does not use any statistical information
about the measurements.
In fact it treats all measurements as if they are equals.
Statistical image reconstruction methods
like those we are developing
are based on statistical models for the measurements
and can give more weight to the "good" measurements
and less weight to the "bad" measurements,
thereby resulting in images with less noise.
The following 3 images
give a computer simulated example.
The image on the left represents
a simplified digital brain scan.
The middle image was reconstructed
from simulated noisy PET measurements
using conventional FBP image reconstruction.
Note the large amount of noise and "streaks"
that can confound diagnosis.
The image on the right
was reconstructed using
information weighted smoothing splines,
a simple non-iterative statistical method
for PET image reconstruction
developed by our group at UM.
Note the reduction in streaking artifacts.
Here are links to papers describing the algorithms.
Summary
- PET and SPECT imaging can diagnosis cancer and other diseases
- Classical image formation methods for PET
make image artifacts that hinder diagnosis
- Using accurate statistical and physical models
leads to improved image quality
- Trade-off: more computation. But computers are cheaper
than misdiagnoses...
This research is a highly interdisciplinary collaboration
involving people and/or principles from
EE, BME, Radiology, Medical Physics, and Statistics.
Back to Fessler home page