# Application of artificial intelligence in head and neck tumor segmentation: a comparative systematic review and meta-analysis between PET and PET/CT modalities

## TL;DR

This study compares AI-based tumor segmentation in head and neck cancers using PET and PET/CT imaging, finding PET/CT to be more effective.

## Contribution

The study provides a systematic review and meta-analysis showing PET/CT outperforms PET-only in AI-assisted tumor segmentation.

## Key findings

- PET/CT showed higher Dice Similarity Coefficient, sensitivity, and precision compared to PET-only.
- Hausdorff Distance was reduced by approximately 3 mm with PET/CT.
- Subgroup analyses confirmed consistent results across different segmentation tasks.

## Abstract

For the effective treatment planning of head and neck cancers, precise tumor segmentation is vital. The combination of artificial intelligence (AI) technology with imaging systems like positron emission tomography (PET) and PET/ computed tomography (PET/CT) has made attempts to automate these processes. Despite these attempts, the usefulness of AI segmentation with PET imaging compared to PET/CT still lacks clarity.

A comprehensive search was performed on Scopus, Embase, PubMed, Cochrane, Web of Science, and Google Scholar for studies published before Dec 2024, with an update in March 2025. Included studies utilized AI algorithms to segment head and neck tumors via PET or PET/CT and provided quantitative performance measures. Pooled estimates of Dice Similarity Coefficient (DSC) sensitivity, precision, and Hausdorff Distance (HD95) were calculated using a random-effects model. Also, sensitivity analyses were performed to find the potential source of heterogeneity. Additionally, subgroup analyses were conducted for overall and primary tumor segmentation. Publication bias was assessed using weighted Egger’s test, followed by presentation of funnel plots for different metrics. Risk of bias (RoB) was evaluated using the QUADAS-C tool. Also, CLAIM was used to assess methodological quality and robustness of the included studies.

Eleven studies were included. All included studies were rated as having a low risk of bias. Also, CLAIM scores showed a high methodological quality in the studies. There was a significant difference between PET/CT and PET-only modalities. Pooled effectiveness metrics showed improvement in their respective DSC of 0.05 (95% CI 0.03–0.07), sensitivity by 0.04, and precision by 0.05, and HD95 decreased by approximately 3 mm. There was low heterogeneity for most metrics except HD95, which showed a high heterogeneity (I2 = 75%) and sensitivity, which showed a moderate heterogeneity (60.79%). Sensitivity analyses showed that leaving out the study by Dong et al. made the mean difference in HD95 smaller (from − 3.22 to − 1.82), but the result was still not statistically significant. When we did more sensitivity analysis by excluding SD-imputed studies, we found that the pooled effect sizes across all performance metrics did not change in direction or significance. We did subgroup analyses based on task type (overall vs. primary tumor segmentation) and modality comparison, and we found that all of the key metrics (Dice, Hausdorff Distance, Precision, Sensitivity) showed the same results, with no significant differences between the subgroups.

The performance of AI-assisted segmentation using PET/CT is greater than that of PET-only in neck and head tumors. These results justify the clinical use of AI-based PET/CT imaging beyond contouring due to its automation potential and highlight the importance of unified datasets alongside distributed learning systems that improve the applicability and consistency of clinical workflows.

The study protocol was registered at PROSPERO [CRD42024614436].

The online version contains supplementary material available at 10.1186/s12885-025-14881-8.