A new saliency metric for precise denoising PET images for better visualization and accurate segmentation

Summary

Nasim Souly, Georgios Z. Papadakis, Uygar Teomete, and Ulas Bagci presented this work orally at the Radiological Society of North America (RSNA) 101st Scientific Assembly and Annual Meeting, held November 29–December 4, 2015, in Chicago. The proceedings describe a saliency-based approach that combines contrast maps across image scales to denoise PET images. Its retrospective evaluation used 20 PET-CT images of NEMA phantoms, with CT correspondence as ground truth.

Evaluation and scope

The RSNA proceedings report comparisons with non-local means denoising. They state that the evaluation found no statistically significant difference between methods for signal-to-noise ratio (SNR), relative contrast ratio, and measured standardized uptake values (SUVs), while reporting an average SNR improvement of more than 30% and preservation of SUVmax. The reported evaluation was on phantom images; it does not establish performance in prospective clinical use.