Deep-learning-based automated delineation of non-Hodgkin lymphoma in [18F]FDG PET/CT

Ziel/Aim: Total metabolic tumor volume (MTV) is a significant prognostic factor in patients with aggressive non-Hodgkin lymphoma. Extraction of image-based biomarkers, however, requires delineation of all tumor lesions, which is non-trivial, time-consuming, and error-prone with the currently available semi-automated methods. The goal of this study was the development and evaluation of a deep-learning-based method for automated delineation of lymphoma in [ 18 F]FDG PET/CT. Methodik/Methods: Automated delineation was performed with a 3D U-Net convolutional neural network (CNN) developed from scratch using nnUNet framework. 641 [ 18 F]FDG PET/CT scans of non-Hodgkin lymphoma patients from the PETAL [ 1 ] trial were used for network training and testing using a 5-fold cross-validation scheme. The ground truth delineations were developed iteratively by an experienced observer with assistance of intermediate CNNs models trained on smaller subsets of the data. Ergebnisse/Results: The derived CNN models are capable of accurate delineation, achieving a Dice similarity coefficient of 0.944. Sensitivity of lesion detection was 0.754 and positive predictive value was 0.881. The mean absolute error of total MTV determination was 10.4 ml with the average total MTV of 127.7 ml. Schlussfolgerungen/Conclusions: In this work, we present a CNN able to perform delineation of tumor lesions in lymphoma with only minimal manual corrections possibly required. It thus is able to accelerate study data evaluation in quantitative PET and does also have potential for supervised clinical application. Publication History Article published online: 12 March 2025 © 2025. Thieme. All rights reserved. Georg Thieme Verlag KG Oswald-Hesse-Straße 50, 70469 Stuttgart, Germany