What can SPECT learn from autoradiography?

For SPECT, where noise and systematic degradations are severe, Bayesian reconstruction approaches have been advocated for their ability to effectively model the degradations, and to model, through prior density functions, the expected local spatial structure (smoothness) of the class of objects to be reconstructed. These priors are chosen subject to the constraints of mathematical tractability and belief as to the nature of the object. The authors propose to use autoradiography as a source of “ground truth” functional objects, and show how these may be used as training data to compute a smoothing hyperparameter in a commonly used form of prior in which differences between adjacent pixels are penalized as the sum of the squares of their differences. A discussion of problems in conditioning autoradiographic data for use as ground truth data in SPECT is included, as is a brief description of the image formation process in the autoradiography of radiopharmaceuticals. The approach to hyperparameter learning applies to any data, not just autoradiography, deemed representative of the class of objects to be imaged