Denoising PET via non-local means with entropy-based regulation.

Background: The noise levels prevalent within PET images render the delineation of structures difficult; thus it is customary to reduce noise post-reconstruction. While there are many methods of doing this, non-local means (NLM) filtering has the advantage that it better preserves the quantitation of the original image. However, the conventional NLM implementation reduces edge integrity and consequently the level of detail surrounding structures of interest. Existing dual-modality methods, which integrate morphological information from an associated MR or CT scan in order to adapt the level of filtering within those regions, assume image alignment, which is not always the case. In this work, the image entropy is utilized to suppress the strength of the filter in such regions, preserving detail while suppressing noise where needed.

Methods: The following 3D NLM algorithm was applied to five whole-body ¹⁸F-FDG PET scans, acquired with a Siemens Biograph Vision 600: 1. Calculate the image entropy map, using empirically optimized parameters. 2. Normalize and invert the image entropy map. 3. Apply the NLM filter to the image using the values obtained in the previous step to suppress the strength of the filter in highly-detailed areas while maximizing the strength within homogeneous regions. A qualitative evaluation was performed by way of a blinded assessment of lesion delineation by an expert nuclear medicine physician. The evaluation included original images as well as filtered images with and without entropy-based regulation. To test the preservation of quantitation, mean SUV lesion-to-liver ratios were obtained from the original and filtered images. To assess edge preservation, line profiles were obtained from two spheres of the NEMA IQ phantom (foreground-to-background ratio of 10:1), on both the original and filtered images.

Results: An overall improvement in lesion delineation with the entropy-regulated filter was observed in the blinded clinical evaluation. The filter was found to preserve mean SUV lesion-to-liver ratios to within an average percentage difference of 0.01% and to preserve edge contrast.

Conclusions: Non-local means filtering with entropy-based regulation shows promise at removing image noise while preserving edge contrast and quantitation.

• Download figure
• Open in new tab
• Download powerpoint