Compared with slice-by-slice approaches for SPECT reconstruction, thre
e-dimensional iterative methods provide a more accurate physical model
and an improved SPECT image. Clinical application of these methods, h
owever, is limited primarily by their computational demands. This pape
r investigates methods for approximate 3D iterative reconstruction tha
t greatly reduce this demand by excluding from the reconstruction the
smaller magnitude elements of the system matrix. A new method is descr
ibed which is designed to control the resulting bias in the SPECT imag
e for a given reduction in computation. The approximate methods were c
ompared to fully 3D iterative reconstruction in terns of SPECT image b
ias and visual quality. All methods were incorporated into the ML-EM a
lgorithm and applied to data from 3D mathematical and experimental bra
in phantoms. The SPECT images reconstructed by the approximate methods
exhibited a positive bias throughout the image that was in general sm
aller with the new method (in the range of 2%-6%). The bias was smalle
st in locally hot regions and largest in locally cold regions. The hig
h quality brain phantom images demonstrated the capability of the new
method in realistic imaging contexts. The time per iteration for an en
tire 3D brain phantom on a modern workstation using the approximate 3D
method was 7.9 s. (C) 1997 American Association of Physicists in Medi
cine.