PET Slides
Introduction to Image Reconstruction
Topic Title: Image Reconstruction: Data Models and Reconstruction Methods
Presenter: Prof Andrew J. Reader, School of Biomedical Engineering and Imaging Sciences
Focus on Positron Emission Tomography (PET) as a primary example, applicable to general medical imaging.
PET Image Reconstruction Basics
Image Acquisition: Fundamental process in PET imaging, involving 3 data formats:
Backprojected images
Sinograms
List-mode data
Data Models: Two primary models in use:
Convolution Radon Transform (or x-ray transform)
Key Components in PET Imaging
Radiotracer: Commonly used radioactive isotopes, e.g., 18F, injected into the patient.
Detection Mechanism: PET scanners utilize high-density crystals to detect annihilation photons at 511 keV, which allows for imaging of various organs including the brain and heart, and diagnostic applications such as dementia and cancer.
Medical Imaging Process: Steps involved from data acquisition to image reconstruction, delineating the measured data and resulting radiotracer distribution.
The Image Reconstruction Problem
The goal is to reconstruct the object function f(r) from detected signals.
Image Dimension: Standard pixel dimension is 256x256 with over 65,000 parameters.
PET scanner configuration includes a detector ring that captures photon pairs, defining a line of response (LOR).
Data Representation Models
Point Source Representation:
Convolution Model: Converts each point source into a point spread function (PSF), modeled mathematically as:
f(r) = ∫ f(r')δ(r - r')dr'
g(r) = ∫ f(r')h(r - r')dr'
Assumptions: The convolution assumes a linear shift-invariant (LSI) system, essential for modeling point sources effectively.
Types of Data Used
List-mode Data: Includes the parameters:
D1, D2 (detector positions)
t (time)
E1, E2 (energy)
∆t (time differences)
These parameters assist in detailed modeling of radiotracer information.
Advanced Reconstruction Methods
First Reconstruction Method: Based on backprojected images with the equation g(r) = ∫ f(r')h(r - r')dr' involving deconvolution for improving accuracy.
Second Reconstruction Method: Utilizing sinogram data, this method is termed Filtered Backprojection (FBP) which involves:
Fourier transformation of sinograms
Application of a ramp filter
Inverse Fourier transformation followed by backprojection.
Sinogram and Radon Transform
Sinograms: Modelled using 2D Radon transform, characterizing that counts in a sinogram are proportional to summed activity along a line.
Important equations:
m(s,φ) ≈ ∫ f(x,y)dy
Relationships illustrating conversions between angle and position:
x = s cos φ - l sin φ
y = s sin φ + l cos φ
Corrections in Medical Imaging
Data Corrections Needed: Essential steps include subtracting randoms and scatter from sinograms, followed by multiplicative correction for attenuation factors and normalisation respectively.
Steps defined:
Subtract randoms (r) and scatter (s)
Apply attenuation correction factors (ACF)
Correct for non-uniform detector efficiency.
Iterative Reconstruction Techniques
Maximum Likelihood Expectation Maximization (ML-EM): A statistical approach designed for accurate modeling, utilizing a Poisson noise model.
Need for Iterative Reconstruction: More accurate compared to FBP, it allows for enhanced modeling of noise and signal behavior in PET images.
Conclusion
Overview of reconstruction methods, including improvements made over the years from basic FBP to advanced ML-EM techniques, emphasizing iterative approaches that lead to greater accuracy in medical imaging.
Closing remarks on ongoing innovations in the field and the importance of statistical models in PET.
End of Lecture: Review of all topics covered, leading to a deeper understanding of image reconstruction techniques in biomedical imaging.