2025/09/24 by Julie Auriac, Christophe Nioche, Narinée Hovhannisyan‐Baghdasarian +8 · 1 citation
Medicine · #Medical Imaging Techniques and Applications #Radiomics and Machine Learning in Medical Imaging #Lung Cancer Diagnosis and Treatment
paper · pdf · doi:10.1002/mp.70025
BACKGROUND: Manual segmentation of organs from PET/CT images is a time-consuming and highly operator-dependent task. Open software solutions are now available to automatically segment all major anatomical structures in CT images. PURPOSE: We compared the volumes and standardized uptake values (SUV) extracted from [18F]FDG-PET/CT patient scans for 33 anatomical structures segmented using two deep learning (DL) algorithms to determine if they are interchangeable. METHODS: Baseline [18F]FDG-PET/CT images were collected retrospectively for 315 women with metastatic breast cancer. A total of 33 anatomical volumes of interest (VOI) were segmented from the whole-body CT scans using both MOOSE v.3.0.14 and TotalSegmentator v.2.0.5 and copied onto the corresponding PET images. For each VOI, the volume from the CT image and SUVmax, SUVpeak and SUVmean from the PET image were extracted. The resulting values were compared using the relative difference for each feature. RESULTS: Following DL segmentation, resulting organ volumes differed by less than 10% for 19/33 organs in more than 80% (252/315) of patients. Four organs were segmented with volume differences greater than 20% in 1/5th of patients: bladder (48%, p < 0.0001), portal and splenic veins (34%, p < 0.0001), thyroid (16%, p < 0.0001), adrenal glands (15%, p < 0.0001). SUVmax and SUVpeak were affected by the choice of DL algorithms, with values differing by less than 10% in more than 80% of patients for only 16 and 19 out of 33 organs, respectively. In contrast, SUVmean was less affected with differences of less than 10% in more than 80% of patients for all anatomical structures, except the bladder, lungs and skull. CONCLUSIONS: The two software tools produce similar results in volume estimates for most anatomical structures. SUVmean is less dependent on the segmentation algorithm than SUVmax and SUVpeak and shows excellent reproducibility for all anatomical structures studied except for the bladder, the lungs and the skull.