Validation of using artificial neural networks to correct crosstalk in simultaneous dual radionuclide imaging

ABSTRACT Simultaneous dual radionuclide SPECT imaging has the advantage of imaging two different radionuclides in a single acquisition. However, this emission tomography imaging technique has not been widely used in clinical application due to the crosstalk contamination - the photons emitted by one radionuclide can be detected in the energy window of the other radionuclide and vice versa. Many methods have been tried to correct crosstalk. or these methods, Artificial Neural Networks (ANNs) have demonstrated a promising ability to remove crosstalk from the photopeak energy windows. A new approach using ANNs to correct crosstalk in simultaneous dual radionuclide imaging was proposed by Zheng et al. (2004). It achieved a sound image quality. However, this approach has only been tested on limited experimental phantoms and has not been compared with the standard ANN approach. This thesis is to validate this new approac h h y us. 111g 1·I · ve p h an toms on t h e '>9111T c /12.· 1 I an cl -'7°11·1;'>9 111011 c rac 1· 1onuc 1· I( I c pa. irs. The SimSET sortware package is used to generate the projection data for the five digital p I iantoms. V · anous acu· · v1ty ra · tios o f. lJlJml, c /17-1· 1 an. d -,(JITl/'JtJniT c toget h er wit · h t h c d elector and collimator geometrics as used in clinical practice, arc configured as the SimSET . mpul parameters to Sl. !llll I ate t h e SPEC... -T acqu1. s1. ll. on. f' 'Or 'JlJrnT c /1 '7--11 , t h e pro . JcCL. !on d ata are acquired in the energy range of 79- 183 keV , which covers the 'J9111Tc/123 1 photopeak energies or 140 keV and 1 59 keV. For 201 Tl!9'J1nTc, the projection data are acquired in the energy range or 60- I SOkeV, which also covers the photopeak energies of 72 keV and 140 keV. These energy ranges are evenly divided into 24 subwindows. The projection data from the PHG torso phantom are used to train a MLP network. This MLP network consists of 24 input nodes, 32 hidden nodes and 2 output nodes. The trained network is used to correct for the crosstalk in the projection data acquired from t h e ot h er I. our p h antoms f'"I ]I e d wi. t I 1 l h e actt. v1. . ttcs 01 j" l I 1e ')tJmT c ;P--11 am 1 -'7()\r[']/9')niT c c. I ua l radionuclide pairs. In the crosstalk correct.ion procedures for using the new ANN approach, only the input nodes corresponding to the photopeak energy window or one or the dual radionuclide pair, are fed into the MLP network. The other inputs are set to zero. The procedure is repeated for the other radionuclide in the dual radionuclide pair. The outputs of the MLP network are used lo cross subtract the original photopeak windows, which result in the crosstalk corrected projection data. The ANN package is the Stuttgart Neural Network Simulator. The image reconstruction algor ith ms arc OSEM and FBP. The images resulting from these two reconstruction algorithms are compared. To quantitatively measure the performance or the two approaches, a percentage of crosstalk, P, is defined. For the 99111Tc/ 123 I radionuclide pair, the P values of the cylindrical striatal phantoms arc reduced from 300% to -20% and from 100% to O on the two slices, respectively. After using the new approach for crosstalk corrections, less than I 0% crosstalk contarninations arc found for the eight-small-cylinder phantom on various 99mTLi2 1I ratios except for the activity ratio or I :8. For the Zubal phantom, the New Approach is better to correct the crosstalk contaminations than the Standard Approach in most of the regions. For the slice-learn ing phantom, the new approach produces almost perfect correction results while the standard approach produces a perfect clearance at the '>'JmTc/1 23 1 ratio of 50:50.