Quantification of Voxel-Wise Uncertainties Via Modified Bootstrapping in 3D Gamma-Ray Imaging

3D gamma-ray imaging is becoming common in both commercial and research applications. 3D images from simple techniques using spatially variant count rates, or more complex reconstruction methods like Compton or coded aperture imaging, require complex, often nonlinear reconstruction methods. Error propagation through non-linear methods is often difficult such that bootstrapping is an attractive alternative method. However, traditional bootstrapping can be computationally intractable in large 3D reconstructions with many source voxels and imaged events. A computationally efficient modified bootstrapping approach, using repeated event subsampling and starting from burned-in solutions, was tested on real-world datasets. No overfitting was observed across 104 MLEM iterations where the event subset was repeatedly shuffled every five iterations. Voxel-wise uncertainty distributions appeared Gaussian or lognormal due to non-negativity constraints. The modified bootstrapping approach will be benchmarked against traditional bootstrapping in the final presentation.