A robust deep-learning model for fully automatic segmentation of lymphoma lesions on whole-body [18F]FDG PET/CT images

Segmentation of lymphoma lesions on [¹⁸F]FDG PET/CT images is being increasingly used in clinical practice for more objective disease quantification, especially for the computation of the total metabolic tumor volume. Efforts in the development of fully automatic segmentation through the use of deep learning (DL) have been employed and published. Yet, proper external validation causes a decrease in performance, limiting the translation into multicentric and multi-scanner clinical settings. This study proposes a DL model based on a 3D U-Net network trained and tested with multi-scanner-based images and multi-ground truth generation segmentations. A two-channel 3D U-Net having as input PET and low-dose CT was used. This allows the use of anatomical and metabolic information of the patient in the DL model. Evaluation performance with voxel overlap metrics, lesion detection metrics, and quantification of patients’ tumor burden was performed. The developed model proved to be robust across scanners and reproduces physician-based manual and semiautomatic segmentations. Plus, good or excellent agreement with the different ground truth generation segmentations in quantifying lymphoma patients’ tumor burden was observed. This brings trustworthiness to the application of the proposed DL-based model in multi-scanner/multicenter clinical settings. In addition, we made the network publicly available to be tested in other clinical settings.