Diagnostic accuracy of radiomics and artificial intelligence models in diagnosing lymph node metastasis in head and neck cancers: a systematic review and meta-analysis

Abstract

Introduction: Head and neck cancers are the seventh most common globally, with lymph node metastasis (LNM) being a critical prognostic factor, significantly reducing survival rates. Traditional imaging methods have limitations in accurately diagnosing LNM. This meta-analysis aims to estimate the diagnostic accuracy of Artificial Intelligence (AI) models in detecting LNM in head and neck cancers.

Methods: A systematic search was performed on four databases, looking for studies reporting the diagnostic accuracy of AI models in detecting LNM in head and neck cancers. Methodological quality was assessed using the METRICS tool and meta-analysis was performed using bivariate model in R environment.

Results: 23 articles met the inclusion criteria. Due to the absence of external validation in most studies, all analyses were confined to internal validation sets. The meta-analysis revealed a pooled AUC of 91% for CT-based radiomics, 84% for MRI-based radiomics, and 92% for PET/CT-based radiomics. Sensitivity and specificity were highest for PET/CT-based models. The pooled AUC was 92% for deep learning models and 91% for hand-crafted radiomics models. Models based on lymph node features had a pooled AUC of 92%, while those based on primary tumor features had an AUC of 89%. No significant differences were found between deep learning and hand-crafted radiomics models or between lymph node and primary tumor feature-based models.

Conclusion: Radiomics and deep learning models exhibit promising accuracy in diagnosing LNM in head and neck cancers, particularly with PET/CT. Future research should prioritize multicenter studies with external validation to confirm these results and enhance clinical applicability.

Keywords: Deep learning; Head and neck cancer; Lymph node metastasis; PET/CT imaging; Radiomics.

Fig. 1

PRISMA flow diagram showing the review process, PRISMA: Preferred Reporting Items for Systematic Reviews and Meta-Analyses

Fig. 2

Summary Receiver Operating Curves (SROCs) for subgroup meta-analysis comparing models based on different modalities, The between-group difference is derived from the bivariate model, AUC: Area under the curve. DL: Deep learning

Fig. 3

Paired Forest plots for the subgroup meta-analysis comparing models based on different modalities, The between-group difference is derived from the bivariate model, ADC: apparent diffusion coefficient. ag: attention-guided. CE: contrast-enhanced. CECT: contrast-enhanced CT. CI: Confidence interval. CNN: Convolutional neural network. DCE: dynamic contrast-enhanced. DECT: dual-energy CT. DLR: Deep learning radiomics. DWI: diffusion-weighted imaging. GBM: Gradient Boosting Machine. HCR: hand-crafted radiomics. LN: lymph node. LR: logistic regression. NB: naïve Bayes. NN: neural network. RF: Random forest. SVM: Support vector machine

Fig. 4

Summary Receiver Operating Curves (SROCs) for subgroup meta-analysis comparing models based on deep learning vs. hand-crafted radiomics, The between-group difference is derived from the bivariate model, AUC: Area under the curve. DLR: Deep learning radiomics. HCR: hand-crafted radiomics

Fig. 5

Paired forest plots for the subgroup meta-analysis comparing models based on deep learning vs. hand-crafted radiomics, The between-group difference is derived from the bivariate model, ADC: apparent diffusion coefficient. ag: attention-guided. CE: contrast-enhanced. CECT: contrast-enhanced CT. CI: Confidence interval. CNN: Convolutional neural network. DCE: dynamic contrast-enhanced. DECT: dual-energy CT. DLR: Deep learning radiomics. DWI: diffusion-weighted imaging. GBM: Gradient Boosting Machine. HCR: hand-crafted radiomics. LN: lymph node. LR: logistic regression. NB: naïve Bayes. NN: neural network. RF: Random forest. SVM: Support vector machine

Fig. 6

Summary Receiver Operating Curves (SROCs) for subgroup meta-analysis comparing models based on radiomics features extracted from lymph nodes vs. primary tumor, The between-group difference is derived from the bivariate model, AUC: Area under the curve. LN: lymph node

Fig. 7

Paired forest plots for the subgroup meta-analysis comparing models based on radiomics features extracted from lymph nodes vs. primary tumor. ADC: apparent diffusion coefficient. ag: attention-guided, CE: contrast-enhanced. CECT: contrast-enhanced CT. CI: Confidence interval. CNN: Convolutional neural network. DCE: dynamic contrast-enhanced. DECT: dual-energy CT. DLR: Deep learning radiomics. DWI: diffusion-weighted imaging. GBM: Gradient Boosting Machine. HCR: hand-crafted radiomics. LN: lymph node. LR: logistic regression. NB: naïve Bayes. NN: neural network. RF: Random forest. SVM: Support vector machine. The between-group difference is derived from the bivariate model

Fig. 8

Paired funnel plots are used to assess potential publication bias/small study effect among reported values for diagnostic accuracy of the main models of each study, FPR: False positive rate, Se: Sensitivity