Characterization of PET image using global and local entropy

Abstract

In the clinical practice PET imaging provides semi-quantitative information about metabolic activities in human body, using the Standardized Uptake Value (SUV). The SUV scale, by itself, does not to establish thresholds between benign and malignant uptake in high-level analyses, such as pattern recognition. The objective of this work is to investigate in PET image volume with high-uptake regions, two additional descriptors, besides the SUV measurements: the amount of information given by the Hartley function (I_Hartley) and its expected value, the Shannon entropy (H). To estimate these descriptors, two models of the probability distribution were obtained from a high-uptake region of interest (ROI): (i) the normalized grayscale histogram from SUV intensity levels (P_i), which provides global I_HG and H_G; and (ii) the normalized gray level co-occurrence matrix (GLCM) of these graylevels (P_g,k) at the same range, which provides local I_HL and H_L. The beginning results have shown that for the ROI (12x12 pixels) and for mean SUV ranging of 6.6213±0. 5196 g/ml, with SUV_Max = 14,7372 g/ml, the global entropy (2,3778±0,0364) has a higher average uncertainty that local entropy (2,2069±0,0758), with a confidence interval of 99.95% (p_value < 0,05%). This can be explained by analysing the sample from the amount of information, I_Hartley, noting that on average local P_g,k provides up to 90,55±9,18% more information when compared to the amount of information given by global P_i. Therefore, these initial results suggest that, for build algorithms for PET image segmentations using threshold based in entropy measures, it is more appropriate to use a distribution functions estimator which considers the local information of the pixels intensities. The main application of this approach will be for, among other things, to construct pathological phantoms from PET images for dosimetry applications.