Application of Artificial Intelligence-Based Transfer Learning Models to the Bethesda System for Thyroid Cytopathology
Clinical Snapshot
PICO Framework
| P — Population | 94 thyroid fine needle aspiration cytology (FNAC) cases with 949 representative images from patients undergoing thyroid evaluation |
| I — Intervention | Ensemble soft voting transfer learning model combining six deep learning architectures (Xception, ResNet50V2, DenseNet121, MobileNetV2, InceptionV3, EfficientNetB3) |
| C — Comparator | Individual transfer learning models and traditional cytopathological assessment according to TBSRTC |
| O — Outcomes | Diagnostic accuracy metrics including sensitivity, specificity, precision (PPV), negative predictive value (NPV), F1 score, and area under ROC curve for TBSRTC categorisation |
Bottom Line
This preliminary study represents the first application of ensemble transfer learning to thyroid cytopathology classification using the Bethesda system. While the reported diagnostic accuracy appears promising (sensitivity 91%, specificity 98%), the study's validity is significantly limited by its small sample size (94 cases), single-centre design, and lack of external validation. The absence of confidence intervals and insufficient methodological detail further compromise the reliability of these findings. Before clinical implementation, larger multicentre studies with external validation, comparison to expert cytopathologists, and assessment of clinical impact are essential. Australian pathologists should await more robust evidence before considering adoption, particularly given the regulatory requirements for AI-based medical devices and the need for integration with existing laboratory information systems.
Key Findings
P Value: Not reported
Effect Size: DenseNet121 sensitivity 0.91, specificity 0.98; ensemble model F1 score 0.867
Primary Outcome: Diagnostic accuracy of ensemble transfer learning model for TBSRTC categorisation
Nnt Or Sensitivity: Sensitivity 0.91, specificity 0.98 for best performing single model (DenseNet121)
Confidence Interval: Not reported
Clinical Application
Requires significant computational infrastructure and technical expertise; integration into existing cytopathology workflows challenging Could support cytopathologists in rural/remote areas where specialist expertise limited; would require TGA approval as medical device software; cost-effectiveness analysis needed for PBS consideration Patients undergoing thyroid FNAC requiring TBSRTC categorisation, particularly in resource-limited settings
Abstract
BACKGROUND: The Bethesda System for Reporting Thyroid Cytopathology (TBSRTC) provides a standardised framework for thyroid fine needle aspiration cytology (FNAC). Deep learning approaches, particularly transfer learning, have shown potential for cytology but are rarely applied to TBSRTC categorisation. AIMS AND OBJECTIVES: To evaluate the performance of an ensemble soft voting transfer learning model in categorising thyroid cytology smears according to TBSRTC. MATERIALS AND METHODS: A retrospective study of 94 thyroid FNAC cases with 949 representative images was conducted. Six transfer learning models (Xception, ResNet50V2, DenseNet121, MobileNetV2, InceptionV3, EfficientNetB3) were combined using ensemble soft voting with a weighted average. Model performance was assessed using sensitivity, specificity, precision (PPV), negative predictive value (NPV), F1 score and area under the receiver operating characteristic curve (AUCROC). RESULTS: DenseNet121 achieved the highest overall sensitivity (0.91) and specificity (0.98) among all single models. The weighted ensemble model achieved an F1 score of 0.867, marginally below DenseNet121 but with superior precision (0.882) and NPV (0.968). AUCROC was highest in DenseNet 121 (0.99) followed by the weighted ensemble (0.98). CONCLUSION: This is the first study to apply ensemble transfer learning to TBSRTC categorisation. The approach demonstrated strong predictive accuracy, particularly for benign and malignant categories. Larger datasets, ideally with whole-slide imaging, are needed to further validate these findings.
References
- 1.Dey, P., Savala, R., & Saikia, U. N. (2026). Application of Artificial Intelligence-Based Transfer Learning Models to the Bethesda System for Thyroid Cytopathology. Cytopathology, Article cyt.70045. https://doi.org/10.1111/cyt.70045
Related Research
The American journal of pathology
Deep Learning-Enhanced Biomarker Interpretation on Cytology Cell Blocks: Foundations and Emerging Opportunities in Spatial Pathobiology
2 July 2026
Cytopathology : official journal of the British Society for Clinical Cytology
Comparative Evaluation of Transfer Learning Models for Detecting Malignant Cells in Urinary Cytology
6 May 2026
This content is for educational purposes for healthcare professionals only and does not constitute clinical advice. Clinical decisions should be based on individual patient assessment, current guidelines, and appropriate specialist consultation. Editorial Standards · Privacy Policy · Terms of Service