A new approach to image reconstruction in positron emission tomography using artificial neural networks

This study investigates the possibility of using an Artificial Neural Network (ANN) for reconstructing Positron Emission Tomography (PET) images. The network is trained with simulated data which include physical effects such as attenuation and scattering. Once the training ends, the weights of the network are held constant. The network is able to reconstruct every type of source distribution contained inside the area mapped during the learning. The reconstruction of a simulated brain phantom in a noiseless case shows an improvement if compared with Filtered Back-Projection reconstruction (FBP). In noisy cases there is still an improvement, even if we do not compensate for noise fluctuations. These results show that it is possible to reconstruct PET images using ANNs. Initially we used a Dec Alpha; then, due to the high data parallelism of this reconstruction problem, we ported the learning on a Quadrics (SIMD) machine, suited for the realization of a small medical dedicated system. These results encourage us to continue in further studies that will make possible reconstruction of images of bigger dimension than those used in the present work (32 × 32 pixels).