TOMOGRAPHIC IMAGE RECONSTRUCTION OF FAN-BEAM PROJECTIONS WITH EQUIDISTANT DETECTORS USING PARTIALLY CONNECTED NEURAL NETWORKS

Learning and Nonlinear Models - Revista da Sociedade Brasileira de Redes Neurais, Vol. 1, No. 2, pp. 122-130, 2003
© Sociedade Brasileira de Redes Neurais

filtering network. When more coefficients are used the FIR filter approaches the IIR (infinite impulse response) ideal filter and
a better quality reconstruction results. In practice there is a limit after which no improvement is noticed in quality. Using more
coefficient than that leads to a slight decrease in PSNR. Reconstruction with interpolation results in higher PSNR meaning
better quality when more than 50 coefficients are used

Figure 9: Plot of PSNR values versus number of (FIR) filter coefficients, showing that the best image quality is obtained using
interpolation.

The following simulation was carried out to verify the influence of fan-beam geometry in the reconstruction process.
The projection of Shepp-Logan synthetic phantom was generated assuming a fan-beam configuration with the ray source
located D=20 units apart from origin. This sinogram was reconstructed by a neural network assembled for parallel geometry
and by another network build for fan-beam geometry. The reconstructed images with 256x256 pixels had their PSNR
computed and plotted for different distances D used to build the fan-beam reconstruction network. Figure 10 shows the
resulting points and a polynomial fit to help the visualization of the region of maximum PSNR. This maximum quality is
obtained when reconstruction if performed for values of D close to 20 units as expected.

Figure 10: Plot of PSNR values of reconstructed image under the geometry fan-beam with distance D varying at defined
range.

129

<<   5   6   7   8   9   >>

More intriguing information