Improved PET image quantification via iterative denoising and partial volume correction

Objectives To design an automated, accurate, robust, and efficient image enhancement algorithm for PET image to facilitate accurate diagnostic quantitative measurements of lesions

Methods We analyzed 20 PET-CT images of NEMA phantoms. An overview of our proposed framework is as follows (Fig 1A): (i) optimal segmentation based on an affinity propagation machine learning algorithm was applied for identifying the number of clusters in the PET images; (ii) segmentation/region based information was utilized to conduct edge-preserving denoising method; (iii) then, partial volume correction was performed by characterizing the geometric interaction using the uptake regions. These steps are utilized in an iterative manner until the clustering/segmentation convergences. For quantitative evaluation of image enhancement, boundaries from CT images were used as ground truths of uptake region definitions. The activity concentrations within each uptake regions were evaluated accordingly by comparing with true value from phantom parameters. Moreover, signal-to-noise ratio (SNR) was estimated from manually defined ROIs before and after enhancement.

Results For SNR of all high uptake regions before and after image preprocessing, some qualitative results were shown in Fig 1B, as shown from the three sample lines, red as before and blue as after enhancement, the system effectively promoted the image quantification accuracy for PET images. Quantitative results were given in Table 1. An average improvement of 32% was achieved. SNR was significantly promoted from 11.7 before enhancement to 360.5 after enhancement, for ROIs within the phantom body (exclude hot spheres).

Conclusions The proposed comprehensive automated framework helps to achieve better quantification of PET images in an efficient and robust manner. It provides an effective tool for PET image enhancement, and accurately delineates the uptake regions.