Automatic partial volume correction in cardiac PET without anatomical images: a preliminary study

Aim: The partial volume effect (PVE) is one of the main sources of error in quantification of cardiac PET imaging. Most of the algorithms for partial volume correction (PVC) are based on recovery coefficients (RC) derived from phantom measurements and applied to well defined regions-of-interest (ROIs). Such methods require the precise knowledge of the thickness of the structure to be corrected, and this is usually obtained by anatomical imaging (MR, CT or US). However, anatomical images may not be always justified or available for the specific patient. Here, we describe a novel method for automatic PVC of ECG-gated PET cardiac images, which does not require anatomical images. The presented method simply requires a curve describing RC that is obtained, for the specific scanner and reconstruction protocol in use, with a phantom of known geometry. Materials and methods: The PVC algorithm consists in four steps: 1) 3D segmentation of the myocardial walls on the gated PET image, 2) transformation of the segmented image in a 3D thickness map using a local thickness transform (LTT), 3) transformation of the LTT image in a 3D correction mask using the phantom-based RC curve, and 4) application of the correction mask to the original image. PET cardiac image is segmented by a thresholding method used in conjunction with an iterative deblurring method. RCs are measured in a static phantom, simulating cardiac left ventricle (LV), made of two off-centered and independently fillable cavities that reproduced a variable wall thickness (2 mm to 19 mm max.). The LV walls in the phantom were filled with an initial activity concentration of 133 kBq/mL of 18F, whereas all other structures were filled with water. Images were acquired with a PET/CT (GE Healthcare Discovery VCT Rx) and an RC curve was derived and fitted with a 4th degree polynomial curve. Results: The PVC method has been preliminary used on PET images. The corrected activity concentration resulted accurate within 5% and 10% for wall thicknesses > 9 mm and 4 mm respectively. Thinner walls, having a thickness comparable to the voxel size, cannot be recovered because of the limitations of the segmentation method in use. Conclusions: The proposed PVC method has an acceptable accuracy and does not necessitate any additional information by anatomical imaging modalities. Furthermore, it can be implemented in an automatic analysis workflow, thus avoiding manual ROI definition and/or contouring, which are both time-demanding and operator dependent.