{"status":"ok","message-type":"work","message-version":"1.0.0","message":{"indexed":{"date-parts":[[2026,8,27]],"date-time":"2026-08-27T04:25:53Z","timestamp":1787804753990,"version":"build-2784847793"},"reference-count":35,"publisher":"Frontiers Media SA","license":[{"start":{"date-parts":[[2026,4,10]],"date-time":"2026-04-10T00:00:00Z","timestamp":1775779200000},"content-version":"vor","delay-in-days":0,"URL":"https:\/\/creativecommons.org\/licenses\/by\/4.0\/"}],"content-domain":{"domain":["frontiersin.org"],"crossmark-restriction":true},"short-container-title":["Front. Endocrinol."],"abstract":"<jats:sec>\n                    <jats:title>Background<\/jats:title>\n                    <jats:p>Fine-needle aspiration biopsy (FNAB) cytology is central to thyroid nodule evaluation, yet reliable differentiation across Bethesda categories remains challenging, particularly for the indeterminate Bethesda V (Suspicious for Malignancy) class. While transfer learning with ImageNet-pretrained models is a standard approach, it remains unclear whether emerging domain-specific medical foundation models offer superior performance compared to general purpose baselines in this specialized domain.<\/jats:p>\n                  <\/jats:sec>\n                  <jats:sec>\n                    <jats:title>Methods<\/jats:title>\n                    <jats:p>We benchmarked four frozen visual encoders\u2014ResNet50, EfficientNet-B0, ViT-Base (ImageNet pretrained), and MedSigLIP (medical image\u2013text pretrained)\u2014on the ThyroidEffi 1.0 dataset (N\u00a0=\u00a01,804), comprising Bethesda II (Benign), Bethesda V (Suspicious), and Bethesda VI (Malignant) cases. A unified evaluation protocol was employed using five-fold stratified cross-validation with a lightweight multilayer perceptron head. Performance was assessed using macro-F1, balanced accuracy, Expected Calibration Error (ECE), and McNemar\u2019s test for statistical significance.<\/jats:p>\n                  <\/jats:sec>\n                  <jats:sec>\n                    <jats:title>Results<\/jats:title>\n                    <jats:p>EfficientNet achieved the highest macro-F1 (0.845 \u00b1 0.021), followed closely by MedSigLIP (0.836 \u00b1 0.019), ResNet50 (0.829 \u00b1 0.015), and ViT (0.817 \u00b1 0.020). Pairwise statistical testing revealed that while EfficientNet significantly outperformed ViT (p &amp;lt; 0.05), the difference between EfficientNet and MedSigLIP was not statistically significant after multiple comparison correction. Notably, MedSigLIP demonstrated superior reliability attributes, achieving the highest recall for the challenging Suspicious class (0.808) and the best model calibration score (ECE\u00a0=\u00a00.025) compared to the general-purpose encoders (ECE: 0.044\u20130.082).<\/jats:p>\n                  <\/jats:sec>\n                  <jats:sec>\n                    <jats:title>Conclusions<\/jats:title>\n                    <jats:p>While domain-specific medical pretraining (MedSigLIP) did not yield a statistically significant advantage in aggregate classification accuracy compared to the best ImageNet-based model (EfficientNet), it provided superior calibration and sensitivity for borderline cases. These findings suggest that in thyroid cytology clinical workflow support, encoder selection should be guided by a joint view of discrimination and safety\u2014particularly calibration and Bethesda V sensitivity\u2014rather than aggregate accuracy alone, enabling threshold-based triage and selective expert review. In particular, well-calibrated models such as MedSigLIP suggest a potential benefit in reducing overconfident misclassification in borderline Bethesda V cases, pending prospective validation in real-world triage workflows.<\/jats:p>\n                  <\/jats:sec>","DOI":"10.3389\/fendo.2026.1800630","type":"journal-article","created":{"date-parts":[[2026,4,10]],"date-time":"2026-04-10T04:28:10Z","timestamp":1775795290000},"update-policy":"https:\/\/doi.org\/10.3389\/crossmark-policy","source":"Crossref","is-referenced-by-count":1,"title":["An early evaluation of MedSigLIP in thyroid cytology: a comparative frozen-encoder benchmark against ImageNet-pretrained encoders"],"prefix":"10.3389","volume":"17","author":[{"given":"Mehmet","family":"Poyrazer","sequence":"first","affiliation":[{"name":"Division of Endocrinology and Metabolism, University of Health Sciences Ankara Training and Research Hospital","place":["Ankara, T\u00fcrkiye"]}],"role":[{"vocabulary":"crossref","role":"author"}]},{"given":"R\u0131dvan","family":"Erten","sequence":"additional","affiliation":[{"name":"Division of Geriatrics, Department of Internal Medicine, Faculty of Medicine, Kocaeli University","place":["Kocaeli, T\u00fcrkiye"]}],"role":[{"vocabulary":"crossref","role":"author"}]}],"member":"1965","published-online":{"date-parts":[[2026,4,10]]},"reference":[{"key":"B1","doi-asserted-by":"publisher","first-page":"1","DOI":"10.1089\/thy.2015.0020","article-title":"2015 American Thyroid Association management guidelines for adult patients with thyroid nodules and differentiated thyroid cancer","volume":"26","author":"Haugen","year":"2016","journal-title":"Thyroid"},{"key":"B2","doi-asserted-by":"publisher","DOI":"10.1016\/j.surg.2025.109710","article-title":"Sonographic and pathologic features of Malignant hot thyroid nodules: a multi-institutional study","volume":"189","author":"Koelliker","year":"2026","journal-title":"Surgery"},{"key":"B3","doi-asserted-by":"publisher","DOI":"10.1089\/thy.2017.0221","article-title":"Cancer risk stratification of indeterminate thyroid nodules: a cytological approach","volume":"27","author":"Valderrabano","year":"2017","journal-title":"Thyroid"},{"key":"B4","doi-asserted-by":"publisher","DOI":"10.1001\/jama.2017.2719","article-title":"Trends in thyroid cancer incidence and mortality in the United States, 1974\u20132013","volume":"317","author":"Lim","year":"2017","journal-title":"JAMA"},{"key":"B5","doi-asserted-by":"publisher","DOI":"10.1089\/thy.2017.0500","article-title":"The 2017 bethesda system for reporting thyroid cytopathology","volume":"27","author":"Cibas","year":"2017","journal-title":"Thyroid"},{"key":"B6","doi-asserted-by":"publisher","DOI":"10.1159\/000339959","article-title":"The bethesda system for reporting thyroid cytopathology: a meta-analysis","volume":"56","author":"Bongiovanni","year":"2012","journal-title":"Acta Cytol"},{"key":"B7","doi-asserted-by":"publisher","DOI":"10.1089\/thy.2018.0779","article-title":"Molecular testing versus diagnostic lobectomy in Bethesda III\/IV thyroid nodules: a cost-effectiveness analysis","volume":"29","author":"Nicholson","year":"2019","journal-title":"Thyroid"},{"key":"B8","doi-asserted-by":"publisher","DOI":"10.1002\/cncy.22756","article-title":"Evaluating local thyroid cytopathology practices by molecular quality metrics: a multi-institutional study on 4651 FNAs with a focus on the role of the interventional cytopathologist","volume":"131","author":"Nacchio","year":"2023","journal-title":"Cancer Cytopathol"},{"key":"B9","doi-asserted-by":"publisher","DOI":"10.1007\/s12022-025-09856-1","article-title":"Subclassifying \u201cAtypia of Undetermined Significance (AUS)\u201d category in the 2023 Bethesda System for Thyroid Cytopathology: analyzing K-TIRADS, BRAF V600E mutation, and risk of Malignancy","volume":"36","author":"Bang","year":"2025","journal-title":"Endocr Pathol"},{"key":"B10","doi-asserted-by":"publisher","DOI":"10.3390\/cancers17213525","article-title":"Artificial intelligence in thyroid cytopathology: diagnostic and technical insights","volume":"17","author":"Negrelli","year":"2025","journal-title":"Cancers"},{"key":"B11","doi-asserted-by":"publisher","first-page":"17843","DOI":"10.1038\/s41598-019-54434-1","article-title":"Diagnosis of thyroid nodules: performance of a deep learning convolutional neural network model vs. radiologists","volume":"9","author":"Park","year":"2019","journal-title":"Sci Rep"},{"key":"B12","doi-asserted-by":"publisher","DOI":"10.3803\/EnM.2024.2058","article-title":"Deep learning technology for classification of thyroid nodules using multi-view ultrasound images: potential benefits and challenges in clinical application","volume":"40","author":"Kim","year":"2025","journal-title":"Endocrinol Metab (Seoul)"},{"key":"B13","doi-asserted-by":"publisher","DOI":"10.4103\/jpi.jpi_43_18","article-title":"Artificial intelligence in cytopathology: a neural network to identify papillary carcinoma on thyroid fine-needle aspiration cytology smears","volume":"9","author":"Sanyal","year":"2018","journal-title":"J Pathol Inform"},{"key":"B14","doi-asserted-by":"publisher","DOI":"10.1002\/cncy.22238","article-title":"Application of a machine learning algorithm to predict Malignancy in thyroid cytopathology","volume":"128","author":"Range","year":"2020","journal-title":"Cancer Cytopathol"},{"key":"B15","article-title":"Transfusion: understanding transfer learning for medical imaging","volume-title":"Advances in Neural Information Processing Systems","author":"Raghu","year":"2019"},{"key":"B16","doi-asserted-by":"crossref","DOI":"10.1109\/CVPR.2009.5206848","article-title":"ImageNet: a large-scale hierarchical image database","volume-title":"Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)","author":"Deng","year":"2009"},{"key":"B17","article-title":"ImageNet-21K pretraining for the masses","volume-title":"Advances in Neural Information Processing Systems (NeurIPS) 2021, Datasets and Benchmarks Track","author":"Ridnik","year":"2021"},{"key":"B18","doi-asserted-by":"crossref","DOI":"10.1109\/CVPR.2019.00277","article-title":"Do better ImageNet models transfer better","author":"Kornblith","year":"2019"},{"key":"B19","unstructured":"MedSigLIP model card. Health AI developer foundations"},{"key":"B20","doi-asserted-by":"publisher","first-page":"2507.05201","DOI":"10.48550\/arXiv.2507.05201","article-title":"MedGemma technical report","author":"Sellergren","year":"2025","journal-title":"arXiv"},{"key":"B21","doi-asserted-by":"crossref","DOI":"10.1109\/ICCV51070.2023.01100","article-title":"Sigmoid loss for language image pre-training","author":"Zhai","year":"2023"},{"key":"B22","doi-asserted-by":"crossref","DOI":"10.1109\/ICCV48922.2021.00346","article-title":"Big self-supervised models advance medical image classification","author":"Azizi","year":"2021"},{"key":"B23","doi-asserted-by":"publisher","DOI":"10.1016\/j.compbiomed.2025.111419","article-title":"Reliable leukemia detection via transfer-enhanced Bayesian CNNs","volume":"202","author":"Hita","year":"2026","journal-title":"Comput Biol Med"},{"key":"B24","article-title":"On calibration of modern neural networks","author":"Guo","year":"2017"},{"key":"B25","doi-asserted-by":"publisher","DOI":"10.1002\/nbm.70227","article-title":"A comprehensive framework for uncertainty quantification of voxel-wise supervised deep learning models in IVIM MRI","volume":"39","author":"Casali","year":"2026","journal-title":"NMR BioMed"},{"key":"B26","doi-asserted-by":"publisher","DOI":"10.1002\/mp.17995","article-title":"The impact of uncertainty estimation on radiomic segmentation reproducibility and scan-rescan repeatability in kidney MRI","volume":"52","author":"Damiano","year":"2025","journal-title":"Med Phys"},{"key":"B27","doi-asserted-by":"crossref","DOI":"10.1109\/CVPR.2016.90","article-title":"Deep residual learning for image recognition","author":"He","year":"2016"},{"key":"B28","article-title":"EfficientNet: rethinking model scaling for convolutional neural networks","author":"Tan","year":"2019"},{"key":"B29","article-title":"An image is worth 16\u00d716 words: transformers for image recognition at scale","author":"Dosovitskiy","year":"2021"},{"key":"B30","doi-asserted-by":"publisher","first-page":"2504.14139","DOI":"10.48550\/arXiv.2504.14139","article-title":"ThyroidEffi 1.0: a cost-effective system for high-performance multi-class thyroid carcinoma classification","author":"Pham-Ngoc","year":"2025","journal-title":"arXiv"},{"key":"B31","doi-asserted-by":"crossref","DOI":"10.1007\/978-1-4899-4541-9","volume-title":"An introduction to the bootstrap","author":"Efron","year":"1993"},{"key":"B32","doi-asserted-by":"publisher","DOI":"10.1007\/BF02295996","article-title":"Note on the sampling error of the difference between correlated proportions or percentages","volume":"12","author":"McNemar","year":"1947","journal-title":"Psychometrika"},{"key":"B33","doi-asserted-by":"publisher","first-page":"289","DOI":"10.1111\/j.2517-6161.1995.tb02031.x","article-title":"Controlling the false discovery rate: a practical and powerful approach to multiple testing","volume":"57","author":"Benjamini","year":"1995","journal-title":"J R Stat Soc Ser B (Methodol)"},{"key":"B34","article-title":"Benchmarking neural network robustness to common corruptions and perturbations","author":"Hendrycks","year":"2019"},{"key":"B35","doi-asserted-by":"publisher","DOI":"10.22581\/muet1982.0035","article-title":"Deep learning-based dual optimization framework for accurate thyroid disease diagnosis using CNN architectures","volume":"44","author":"Haider","year":"2025","journal-title":"Mehran Univ Res J Eng Technol"}],"container-title":["Frontiers in Endocrinology"],"original-title":[],"link":[{"URL":"https:\/\/www.frontiersin.org\/articles\/10.3389\/fendo.2026.1800630\/full","content-type":"unspecified","content-version":"vor","intended-application":"similarity-checking"}],"deposited":{"date-parts":[[2026,4,10]],"date-time":"2026-04-10T04:28:13Z","timestamp":1775795293000},"score":1,"resource":{"primary":{"URL":"https:\/\/www.frontiersin.org\/articles\/10.3389\/fendo.2026.1800630\/full"}},"subtitle":[],"short-title":[],"issued":{"date-parts":[[2026,4,10]]},"references-count":35,"alternative-id":["10.3389\/fendo.2026.1800630"],"URL":"https:\/\/doi.org\/10.3389\/fendo.2026.1800630","relation":{},"ISSN":["1664-2392"],"issn-type":[{"value":"1664-2392","type":"electronic"}],"subject":[],"published":{"date-parts":[[2026,4,10]]},"article-number":"1800630"}}