A spatially-variant SPECT reconstruction scheme using artificial neural networks

A quantitative, spatially varying, weighted backprojection has been developed for single photon emission computed tomography (SPECT) using artificial neural networks (ANNs). The network has been trained to compensate for collimator effects and attenuation. The required ramp filtering is also learned by the ANN. A supervised training scheme was utilized that implemented the generalized delta rule. After training, the backprojection weights were held constant and could be used to reconstruct source distributions other than those used while training. A noiseless Hoffman brain phantom reconstruction using the proposed technique has a 82.5% reduction in mean-squared error (MSE) compared to standard filtered backprojection (FBP) when collimator and attenuation effects were present. For noisy data, if standard noise reduction filters were implemented prior to reconstruction, the ANN images has a lower MSE than standard FBP images that used the same noise filter. For example, Wiener-filtered, 200000 count Hoffman brain projection data reconstruction by the present network had a 50% lower MSE than standard FBP images reconstructed with the same Wiener-filtered data