AbstractID: 7358 Title: Effect of Median Root Prior (MRP) regularization on the accuracy and convergence of a maximum likelihood expectation maximization algorithm applied to 3D scintillation dosimetry
A new method for simultaneous measurement of radiation dose in all points of a volume for the needs of radiotherapy has been previously proposed (Kirov. A.S. et al,
Med. Phys. 26,1069,1999, and Med. Phys.27,1156,2000). A liquid scintillation (LS) solution is used both as a phantom material and as a dose registration medium. The scintillation photons emitted from the volume are detected at multiple angles by a highly collimated image detector. A SPECT maximum likelihood expectation-maximization
(MLEM) algorithm is used for reconstructing the scintillation light emission density for each voxel of the solution. We have found, that when applied to experimental images corresponding to complex dose distributions and containing edges due to artifacts of the optical system, the shape of the reconstructed emission density depends on the number of iterations. We modified the MLEM code by adding Median Root Prior (MRP) regularization (Allenius S, et al, J. of Nucl.Med.24,258,1997) to include a penalty term, which is set according to the maximum allowed deviation of a pixel value from the local median. The penalty imposed on each pixel’s value is determined from a prior Gaussian distribution, centered at the median of a neighborhood of the pixel. The new algorithm converges in less than 20 iterations, which eliminates the lack of convergence previously observed and leads to improved accuracy of the 3D scintillation dosimetry method.
This work is supported by a research grant No. 00-0194 from the Whitaker
Foundation. A patent on the 3D scintillation dosimetry method is pending.