Partition Models in the Analysis of Autoradiographic Images

SUMMARY A new method is proposed for the analysis of electron microscope autoradiographs. The method uses Bayesian principles to formulate a model, where a Poisson process describing the autoradiographic grain distribution is combined with prior distributions quantifying knowledge about the underlying truth. In common with an earlier model, the new model directly exploits assumptions of homogeneous subregions within the specimen and gives a full treatment of edge effects. However, the new model does not require accurate information about which physical boundaries lead to discontinuities in intensity or whether regions spatially separated have distinct intensities. The key extension over previous models is the inclusion of region labels, in addition to the region intensities. A Bayesian approach is used which also allows the inclusion of other prior information. Monte Carlo methods, based on Metropolis dynamics, are described to produce both point and interval estimates of the model parameters. Several examples illustrating the new procedure are reported. These examples use both real autoradiograph data and simulated data based on the morphology of the real specimen.