Diagnostic accuracy of CT-based radiomics and deep learning for predicting lymph node metastasis in esophageal cancer

Abstract

Background: Esophageal cancer remains a global challenge due to late diagnoses and limited treatments. Lymph node metastasis (LNM) is crucial for prognosis, yet traditional diagnostics fall short. Integrating radiomics and deep learning (DL) with CT imaging for LNM diagnosis could revolutionize prognostic assessment and treatment planning.

Methods: A systematic review and meta-analysis were conducted by searching PubMed, Scopus, Web of Science, and Embase up to October 1, 2023. The focus was on studies developing CT-based radiomics and/or DL models for preoperative LNM detection in esophageal cancer. Methodological quality was assessed using the METhodological RadiomICs Score (METRICS).

Results: Twelve studies were reviewed, and seven were included in the meta-analysis, most showing excellent methodological quality. Training sets revealed a pooled AUC of 87 % (95 % CI: 78 %-90 %), and internal validation sets showed an AUC of 85 % (95 % CI: 76 %-89 %), with no significant difference (p = 0.39). Sensitivity and specificity for training sets were 78.7 % and 81.8 %, respectively, with validation sets at 81.2 % and 76.2 %. DL models in training sets showed better diagnostic accuracy than radiomics (p = 0.054), significant after removing outliers (p < 0.01). Incorporating clinical data improved sensitivity in validation sets (p = 0.029). No significant difference was found between models based on CE or non-CE imaging (p = 0.281) or arterial or venous phase imaging (p = 0.927).

Conclusion: Integrating CT-based radiomics and DL improves LNM detection in esophageal cancer. Including clinical data could enhance model performance. Future research should focus on multicenter studies with independent validations to confirm these findings and promote broader clinical adoption.

Keywords: Deep learning; Esophageal neoplasms; Lymphatic metastasis; Radiomics; Tomography; X-ray computed.