Hallucination Analysis of Score-Based Diffusion Models for CT Denoising Through Spatial Frequency Decomposition

While score-based diffusion models are powerful denoisers, they present currently non-quantifiable hallucinations in their reconstructions. In this work, we apply quantitative analysis to the hallucinations resulting from denoising diffusion models. Specifically, we sampled the generative reverse dif-fusion process multiple times and calculated standard deviation maps to compare the variability among reconstructions. We further decomposed the spatial frequency of the reconstructions and calculated the error metrics of each frequency band. Lastly, we associated the image quality analysis and frequency decomposition with a classification observer study. We found that diffusion based reconstructions had lower root-mean-squared error than the original measurements but re-sulted in lower accuracy in the classification observer study. These preliminary results demonstrate the need for further in-vestigation into hallucinations before widespread adoption of generative image processing in medical imaging.