Performance of AI methods in PET-based imaging for outcome prediction in lymphoma: A systematic review and meta-analysis

Abstract

Objectives

To evaluate the predictive performance of artificial intelligence (AI) methods using pre-treatment PET-based imaging for outcome prediction in lymphoma through a systematic review and meta-analysis.

Methods

PubMed-MEDLINE, Scopus, and Web of Science were searched for original studies on AI prediction models using PET-based imaging in lymphoma up to October 2024. Eligible studies reported outcomes including progression-free survival (PFS), overall survival (OS), or treatment response. Meta-analyses, subgroup analyses, meta-regressions, sensitivity analysis, and publication bias analysis were conducted using Stata software.

Results

Seventy-five studies were included, predominantly focusing on non-Hodgkin lymphoma (NHL, n = 61). AI methods included deep learning (DL, n = 13), machine learning (ML, n = 2), combined ML/radiomics (n = 23), and radiomics (n = 37). Pooled analyses showed strong predictive performance for PFS (HR: 4.11 [3.20–5.29], AUC: 0.78 [0.68–0.86], C-index: 0.79 [0.76–0.83]) and OS (HR: 3.38 [2.29–4.99], AUC: 0.75 [0.66–0.83], C-index: 0.79 [0.76–0.81]) in the main groups with consistent results in the validation groups. For treatment response, pooled OR was 5.36 [1.53–18.78], and AUC was 0.85 [0.74–0.92]. DL outperformed other AI methods in PFS and treatment response prediction.

Conclusion

AI methods, particularly DL, show strong predictive performance for lymphoma outcomes using PET-based imaging, supporting their potential utility in precision medicine. Further prospective studies are needed for clinical integration.