Diagnostic performance of machine learning and deep learning algorithms for thyroid cancer metastasis: a systematic
Mohammad Amouzadeh Lichahi1, Saeed Anvari2, Hossein Hemmati3
1Student Research Committee, School of Medicine, Anzali International Campus, Guilan University of Medical Sciences, Rasht, Iran.
Background:
Metastasis significantly influences prognosis in thyroid cancer, especially in papillary thyroid carcinoma. With the rise of artificial intelligence (AI) in medical diagnostics, machine learning (ML) and deep learning (DL) models are being increasingly explored for their ability to enhance the early detection of metastatic spread. This systematic review and meta-analysis aimed to evaluate the diagnostic performance of ML and DL algorithms in detecting metastasis in thyroid cancer.
Method:
We conducted a comprehensive search of scientific databases, including PubMed, IEEE, Scopus, and Web of Science, covering literature up to July 1st, 2025. This review included studies published in English that used diagnostic models for metastasis in adults with thyroid cancer. Key metrics analyzed were the area under the receiver operating characteristic curve (AUC-ROC) sensitivity, specificity, and the diagnostic odds ratio (DOR) with a 95% confidence interval (CI). Heterogeneity was quantified using I² statistics, and subgroup and moderator analyses were conducted to identify sources of variability. Risk of bias was assessed using the PROBAST tool. Bias risk and concerns were evaluated using the PROBAST checklist. This study was registered with PROSPERO (CRD42024622930).
Results:
Thirty-five studies encompassing 162 estimates were included. The pooled sensitivity was 0.747 (95% CI: 0.715-0.775) and specificity was 0.746 (95% CI: 0.706-0.783). The pooled DOR was 9.45 (95% CI: 7.27-12.28), indicating a strong association between AI predictions and actual metastatic status. The overall AUC-ROC was 0.818. Subgroup analysis demonstrated particularly high accuracy in models targeting distant metastasis. ML models showed slightly higher discriminative ability compared to DL models, and robust performance was observed across a variety of cancer subtypes and input data sources. Moderator analysis further confirmed the stability and adaptability of these models under different clinical and technical settings.
Conclusion:
ML and DL algorithms demonstrate favorable diagnostic performance in identifying metastasis in thyroid cancer and may serve as supportive tools in clinical decision-making. Their consistent results across different metastasis types and technical settings highlight their potential to complement existing diagnostic approaches. These findings encourage further exploration and refinement of AI-based methods for integration into routine oncologic practice.
