Abstract
1375
Objectives To improve image quality in PET, a correct identification of true coincidences, while discarding as many random coincidences as possible, is mandatory. To this aim, we have developed a new coincidence sorting method based on artificial neural networks (NN). The objective of this work is to asses the performance of the NN approach compared to a conventional sorting technique.
Methods The feed-forward NN used is composed by 2 hidden layers of 4 units each, and a single unit in the output. As input, information about time, energy and position for each possible coincidence was used. For training, we used simulated data (GATE) of an extended cylindrical source (1mCi) in a small animal PET scanner. For testing, several phantoms and concentration activities were simulated. A conventional sorting method using a time window and geometrical constraints was also developed.
Results The efficiency (E), defined as the ratio of correctly identified true coincidences to all true coincidences, and the random fraction (RF) were studied. For the NN, the compromise between E and RF was tuned by adjusting the threshold used; for the conventional method E and RF depend on the time window. Given a fixed E, the NN technique yielded RF values always lower than the conventional method. In the NN sinograms, a significant reduction of random coincidences was visible. Correlation coefficients (CC) for the sinograms were calculated, using as reference the sinogram containing only true coincidences. For an off-centered line source (1mCi) and E=92%, CC was 0.98 (NN) and 0.91 (conventional method and time window twice the FWHM).
Conclusions Neural networks are a promising tool for coincidence sorting in PET, capable of reducing the number of randoms while maintaining or even improving the efficiency. We can foresee an improvement in image quality by inspecting the resulting sinograms (dedicated reconstruction algorithms are in progress)