Comparison of multi-organ CT image segmentation tools for whole-body [<sup>18</sup>F]FDG-PET/CT clinical imaging.
Authors
Affiliations (6)
Affiliations (6)
- School of Informatics, University of Edinburgh, 10 Crichton St, Edinburgh, EH8 9AB, Scotland, UK. [email protected].
- Centre for Cardiovascular Science, University of Edinburgh, 47 Little France Cres, Edinburgh, EH16 4TJ, Scotland, UK. [email protected].
- Edinburgh Imaging, University of Edinburgh, 47 Little France Cres, Edinburgh, EH16 4TJ, Scotland, UK. [email protected].
- Centre for Cardiovascular Science, University of Edinburgh, 47 Little France Cres, Edinburgh, EH16 4TJ, Scotland, UK.
- School of Informatics, University of Edinburgh, 10 Crichton St, Edinburgh, EH8 9AB, Scotland, UK.
- Edinburgh Imaging, University of Edinburgh, 47 Little France Cres, Edinburgh, EH16 4TJ, Scotland, UK.
Abstract
Automated multi-organ segmentation looks to assist clinicians and researchers working with Positron Emission Tomography/Computed Tomography (PET/CT) imaging in streamlining the time-consuming, operator-dependent task of manual delineation. This study aimed to compare two state-of-the-art automated multi-organ CT segmentation tools with typically perceived "gold-standard" manual delineation. We compare Multiple-Organ Objective Segmentation (MOOSE) and TotalSegmentator against manual labels of six tissues on a dataset of 24 patients of lung cancer. We evaluated CT segmentation performance using the Dice-Sørensen Coefficient (DSC), Hausdorff Distance (HD), Average Symmetric Surface Distance (ASSD) and pixel-based metrics Precision and Recall. Alongside technical analysis, we perform evaluation using clinically relevant metrics including organ volume, mean standardised uptake value (SUV<sub>mean</sub>), maximum standardised uptake value (SUV<sub>max</sub>), and Hounsfield units. Both MOOSE and TotalSegmentator produce overall comparable DSC results. Conversely, MOOSE and TotalSegmentator segmentation results in significantly different volumes and SUVs compared with manual delineation for the lungs, brain, and kidneys. Data presented here highlights the need to assess multi-organ segmentation tools performance using multi-pronged metrics beyond Dice-Sørensen scores.