From pixels to prognosis: A QUADAS-2–Guided systematic review and meta-analysis of deep learning segmentation for DLBCL in PET and PET/CT

• First systematic review and meta-analysis on AI-based segmentation of DLBCL in PET/CT imaging. • Pooled Dice Similarity Coefficient (DSC) of 0.809 across 11 high-quality studies, indicating robust segmentation accuracy. • Comprehensive QUADAS-2 assessment revealed overall low risk of bias but noted methodological variability. • Identified gaps include limited multi-center data, lack of transformer-based models, and retrospective design dominance. • Future directions emphasize standardized datasets, architectural innovation (e.g., transformers), and prospective validation. This systematic review and meta-analysis evaluated the performance and methodological quality of deep learning models for automated segmentation of Diffuse Large B-Cell Lymphoma (DLBCL) on PET/CT imaging. A comprehensive literature search identified 15 eligible studies that were published up to July 2025. Of these, 11 studies were included in the quantitative synthesis, while 4 were assessed qualitatively. Using a random-effects model, the pooled mean DSC was 0.809 (95 % CI: 0.791–0.827), indicating strong overall segmentation performance. The reported DSC values across the individual studies ranged from 0.65 to 0.886. Single-center studies generally showed slightly higher median DSC values (≈0.82) than multi-center studies (≈0.78), although pooled subgroup analyses revealed comparable averages (0.77 vs. 0.73). Methodological quality, assessed using the QUADAS-2 tool, showed that most studies (approximately 67–73%) were at low risk of bias, with the remainder classified as moderate or unclear. Despite the variability in algorithms, study designs, and datasets, DL-based methods have consistently achieved reliable segmentation accuracy. Overall, DL models demonstrated promising potential for automated DLBCL segmentation in PET/CT imaging. Nevertheless, future studies should focus on larger and more diverse cohorts, improved reporting standards, and transparent handling of methodological limitations to enhance generalizability and clinical applicability.