Moderate evidence
More
Study finds ultrasound AI model predicts thyroid cancer spread with 87.9% accuracy
Study title: A Machine Learning Model Based on Ultrasound Radiomics and Multi-modal Clinical Features for Predicting Lymph Node Metastasis of Papillary Thyroid Carcinoma in the Setting of Hashimoto's Thyroiditis
- Evidence type
- Controlled or validated comparison
- Sample
- n = 377
- Sources
- 1 sourcePeer-reviewed
- Published
- 2 min read
- Truvace record
- TRV-2026-3002cite or check it
Study title: A Machine Learning Model Based on Ultrasound Radiomics and Multi-modal Clinical Features for Predicting Lymph Node Metastasis of Papillary Thyroid Carcinoma in the Setting of Hashimoto's Thyroiditis
Researchers built machine learning models to estimate the risk that papillary thyroid carcinoma has spread to lymph nodes in patients who also have Hashimoto's thyroiditis. The best model, XGBoost, reached an accuracy of 0.879 in an independent external cohort of 107 patients.
Read the study abstract
To investigate the value of multi-modal features (ultrasound characteristics, BRAF V600E mutation, and TPO-Ab) in predicting lymph node metastasis (LNM) of papillary thyroid carcinoma (PTC) in the context of Hashimoto's thyroiditis (HT), and to construct machine learning models for individualized risk assessment. A retrospective cohort of 377 patients with pathologically confirmed PTC and concurrent HT was enrolled and divided into LNM n=186 and non-LNM n=191 groups. Preoperative ultrasound features (e.g., Minim…
The gain
In a study of 377 patients with papillary thyroid cancer and Hashimoto's thyroiditis, an XGBoost model combining ultrasound radiomics and clinical markers reached 0.879 accuracy at predicting lymph node metastasis in an independent external cohort of 107 patients.
In brief
The models were trained on a retrospective cohort of 377 patients, 186 with lymph node metastasis and 191 without. They used preoperative ultrasound features, TPO-Ab blood levels and BRAF V600E mutation status. Radiomics features accounted for 81.5% of the model's predictive contribution.
The authors say the approach could support individualized risk stratification and clinical decision-making for these patients. The evidence comes from retrospective records rather than a prospective clinical deployment.
Main points
- The study enrolled 377 patients with pathologically confirmed papillary thyroid carcinoma and concurrent Hashimoto's thyroiditis: 186 with lymph node metastasis and 191 without.
- Five algorithms were compared: random forest, support vector machine, XGBoost, logistic regression and k-nearest neighbors.
- In univariate analysis, calcification type, nodule diameter, minimum tumor-to-capsule distance and aspect ratio were significantly associated with metastasis; nodule diameter contributed most.
- Radiomics features supplied 81.5% of predictive contribution, morphological features 12.7%, TPO-Ab 3.8% and BRAF V600E 2.0%.
- The best XGBoost model reached 0.879 accuracy in an independent external cohort of 107 patients.
The full story
The retrospective cohort split into 186 patients with lymph node metastasis and 191 without. Preoperative ultrasound features such as minimum tumor-to-capsule distance, calcification type and ratio, lesion location and nodule morphology were collected alongside serological and pathological indicators.
Five models (random forest, support vector machine, XGBoost, logistic regression and k-nearest neighbors) were built and assessed with ROC and calibration curves. Combining several feature types improved prediction accuracy, the authors report.
Feature importance analysis showed radiomics features contributing 81.5%, morphological features 12.7%, TPO-Ab 3.8% and BRAF V600E 2.0%. The authors conclude the models enable individualized risk stratification for lymph node metastasis in this patient group.
Sources
1 on file- Ultrasound in Medicine & Biologydoi:10.1016/j.ultrasmedbio.2026.10.008
Talk it over
The XGBoost model reached 0.879 accuracy in an external cohort of 107 patients, after being built on 377 retrospective records. What can a retrospective design tell us about how it would work with new patients in a clinic?
Study finds ultrasound AI model predicts thyroid cancer spread with 87.9% accuracyhttps://truvace.com/item/a-machine-learning-model-based-on-ultrasound-radiomics-and-multi-modal-clinical-
The XGBoost model reached 0.879 accuracy in an external cohort of 107 patients, after being built on 377 retrospective records. What can a retrospective design tell us about how it would work with new patients in a clinic?
Answer in the threadRadiomics features supplied 81.5% of the model's predictive contribution, and TPO-Ab and BRAF V600E supplied 3.8% and 2.0%. What finding would make you trust or doubt the model's use of these inputs?
Answer in the threadWho might benefit from individualized risk stratification for lymph node metastasis in patients with Hashimoto's thyroiditis, and which patients might be missing from the 377 and 107 patients studied?
Answer in the thread
If you have more time
If you were a patient with papillary thyroid carcinoma and Hashimoto's thyroiditis, how would you want a doctor to use the model's risk estimate alongside the ultrasound?
Answer in the threadWhat would a prospective test of this model need to measure to settle whether it changes treatment decisions?
Answer in the thread
ace
What readers say