Prognostic Evidence for Artificial Intelligence-Derived PET Tumour-Burden Biomarkers in Cancer: A Systematic Review

Background/Objectives: Artificial intelligence (AI) can automate tumour-burden measurement on positron emission tomography (PET), including metabolic tumour volume (MTV) and total metabolic tumour volume (TMTV). This review evaluated prognostic evidence for AI-derived PET tumour-burden biomarkers, particularly comparisons with expert measurements. Methods: PubMed and Embase were searched on 29 August 2026; supplementary searching closed on 9 September. Revision follow-up used an export dated 13 October 2026 while retaining the original temporal eligibility boundary. English-language original reports evaluating AI-derived PET tumour burden and clinical outcomes were eligible. Core comparative and Contextual evidence were synthesized narratively by cancer, tracer, measurement, outcome and validation setting, accounting for shared cohorts. AI-assisted extraction and appraisal underwent human verification, without fully independent duplicate extraction or appraisal. Search coverage was incomplete, and no formal certainty-of-evidence grading was performed. The review was registered in PROSPERO (CRD420261497171). Results: We included 35 reports (13 Core and 22 Contextual) from 31 catalogue families; these families do not represent counts of independent validation cohorts. Selected studies reported associations between higher AI-derived MTV or TMTV and poorer outcomes, with confidence limited by confounding, threshold selection and validation design. Several reports found prognostic estimates similar to expert measurements without establishing equivalence. Heterogeneity in populations, outcomes, predictor scales and thresholds precluded pooled hazard ratios; incomplete paired uncertainty prevented pooling between-method differences. Of 62 prognostic factor assessment units, 37 had high risk of bias for confounding and 50 for statistical analysis and reporting. External validation of fixed prognostic thresholds remained limited. Conclusions: AI-derived MTV and TMTV showed prognostic associations in selected cancer settings, with limited confidence in their independence and transportability. Evidence does not establish interchangeability with expert measurements or improved outcomes from biomarker-guided treatment. Independent validation is needed before clinical use.