Artificial intelligence-based PET/CT analysis in lymphoma: segmentation, differential diagnosis, and prognostic stratification

Abstract Lymphoma is biologically and prognostically heterogeneous, making accurate imaging-based assessment essential for individualized management. 18 F-fluorodeoxyglucose positron emission tomography/computed tomography ( 18 F-FDG PET/CT) is central to staging, response assessment, and follow-up, but manual lesion delineation and visual interpretation remain time-consuming and variable. Artificial intelligence (AI) may improve the reproducibility and clinical value of PET/CT by enabling automated quantification and feature analysis. This review summarizes the current evidence and clinical potential of AI in lymphoma PET/CT across three major domains: lesion segmentation, differential diagnosis, and prognostic stratification. We review deep learning (DL) approaches, including three-dimensional (3D) U-Net-based architectures, for automated lesion segmentation and the extraction of metabolic tumor burden biomarkers such as total metabolic tumor volume (TMTV) and maximum tumor dissemination distance (Dmax). We highlight the transition from radiomics workflows that depend on manual or semi-automated delineation to DL systems that enable end-to-end automation. We also discuss machine learning (ML) models for differential diagnosis, including prediction of bone marrow involvement, histologic subtype classification, aggressiveness assessment, and differentiation from sarcoidosis, metastatic lymph nodes, and selected non-lymphomatous malignancies. For prognostic evaluation, we summarize convolutional neural networks (CNNs), graph neural networks (GNNs), automated ML, and multimodal radiomics models for predicting progression-free survival (PFS), overall survival (OS), relapse risk, and treatment response. AI-based PET/CT analysis is moving from manual feature engineering toward automated, interpretable, and multimodal decision support. With prospective validation, standardized imaging protocols, and integration with clinical and molecular data, these methods may support more precise and individualized lymphoma management.