PET Images Atlas-Based Segmentation Performed in Native and in Template Space: A Radiomics Repeatability Study in Mouse Models

Image segmentation is a task of the utmost importance in computer vision, especially in the biomedical field where accurate delineation of an organ or lesion can make a difference in patient's survival. Although there are several approaches, the Atlas-based co-registration method is most appropriate for low-contrast functional images, where organ boundaries are not easily recognizable. This technique is strongly dependent on the template choice and can even lead to inaccurate results if unproperly tuned; however, two different pipelines were similarly adopted in literature either warping the Atlas to the target image or warping the target image to the Atlas. This, unless proved to be equivalent, may result ambiguous, hence in this study we investigated the two algorithms equivalence employing a preclinical dataset of mice undergoing micro-PET/CT scans after chelator injection. We focused on seven selected organs (namely heart, bladder, stomach, spleen, liver, kidneys and lungs), and for each of them we computed the percentage of PET radiomics features with significant variations between the two algorithms. Our results showed that the two approaches considerably differed. Specifically, a mean significant difference of about 40% was found in the radiomics features extracted following the two different pipelines, posing the need to distinguish between the two registration output spaces.