Analysis of an artificial neural network approach to quantitative SPECT reconstructio

An artificial neural network (ANN) for quantitative single photon emission computed tomography (SPECT) reconstruction is presented. The effects of collimation, attenuation and scatter introduce errors in SPECT images. Accurate SPECT reconstruction must therefore include compensation for degradations. The ANN technique presented in this document simultaneously reconstructs an image while compensating for degradations. This is the first ANN approach to solve the complete SPECT inverse problem.;Initial studies were performed to quantify the extent of degradations and to test the feasibility of an ANN approach to SPECT reconstruction. The first ANN found weights that were used as a filter prior to backprojection. Both sets of weights were similar to the ramp filter and could be used to reconstruct arbitrary data.;The initial studies supported forming a complete reconstruction technique using an ANN. The ANN was trained to minimize the mean-squared error between the actual and ideal outputs. This ANN was spatially-variant by translation, rotation and depth, and compensated for the effects of collimation, attenuation and scatter. Comparisons were made between the ANN and filtered backprojection (FBP) methods with noiseless, 200,000 counts/slice, and 50,000 counts/slice simulated data as well as clinical data. The ANN produced improved images compared to the FBP techniques.;ANN reconstruction kernels were interpreted in terms of ANN weighting and the underlying physics of the system. Comparisons were made between ANN reconstruction kernels and analytical compensation kernels. Similarities between the ANN and analytic kernels led to the formation of a spatially-variant compensation technique based on analytically inverted point spread responses.;Finally, a post-reconstruction compensation technique was developed using an ANN. The input to the post-reconstruction technique was a FBP image. This network was also capable of compensating for the effects due to collimation, attenuation and scatter. The complete reconstruction ANN gave superior images compared to the post-reconstruction ANN, but the post-reconstruction approach outperformed FBP techniques.;ANNs were shown to give superior reconstructions compared to images formed by FBP. The improvement shown by the ANN techniques supports the use of ANNs for quantitative SPECT reconstruction.