A Meta-Analysis of Deep-Learning-Based Fusion of Positron Emission Tomography and Computed Tomography in Oncologic...

The objective of the study is to assess whether deep-learning fusion of positron emission tomography (PET) and computed tomography (CT) improves performance compared with matched unimodal PET and CT and whether different fusion architectures are associated with model performance. Five databases were searched (January 2018-8 May 2025) following Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) 2020. Studies were classified by clinical task, fusion level (early, intermediate, late, or hybrid), and fusion mechanism (fixed or learnable). Incremental and absolute performance were synthesised using robust variance estimation with publication-level clustering. Of 892 records, 103 studies (133 analytical units) were included; 30 units entered incremental synthesis, and 61 entered absolute synthesis (60 weighted). PET/CT fusion showed positive effects for Dice similarity coefficient (mean difference (MD) = 0.044; 95% confidence interval (CI), 0.027-0.060), area under the receiver operating characteristic curve (AUC; MD = 0.090; 95% CI, 0.036-0.144), and concordance index (C-index; MD = 0.038; 95% CI, 0.037-0.038); Dice and C-index were exploratory. Incremental I² was 0% for Dice and C-index and 94.6% for AUC. Absolute performance was 0.758, 0.820, and 0.728, respectively, with substantial heterogeneity. Intermediate and hybrid fusion had higher absolute Dice than early fusion. Learnable fusion operation did not significantly outperform fixed fusion operation. Deep-learning PET/CT fusion shows potential value in oncologic imaging, but heterogeneous evidence and limited external validation constrain its interpretation. A clearer distinction between fusion level and mechanism, supported by more standardised and externally validated studies, is needed to identify clinically meaningful fusion strategies.