# The value of a combined model based on ultra-radiomics and multi-modal ultrasound in the benign-malignant differentiation of C-TIRADS 4A thyroid nodules: a prospective multicenter study

## TL;DR

A new model combining ultra-radiomics and multi-modal ultrasound improves accuracy in distinguishing benign from malignant thyroid nodules classified as C-TIRADS 4A.

## Contribution

A combined diagnostic model using ultra-radiomics and multi-modal ultrasound features for thyroid nodule classification is proposed and validated.

## Key findings

- The combined model achieved an AUC of 0.956 in the training set and 0.863 in the test set.
- The model outperformed the multimodal ultrasound model in both training and test sets.
- Junior physicians using the model showed improved diagnostic performance compared to senior physicians.

## Abstract

To establish a combined model based on ultrasound radiomics combined with multimodal ultrasound and evaluate its value in diagnosing benign and malignant nodules classified as Chinese-Thyroid Imaging Report and Data System (C-TIRADS) 4A.

Prospective collection of data from 446 patients with thyroid nodules classified as C-TIRADS 4A between December 2023 and August 2024. Based on the enrollment timeline, patients were divided into a training set (n=312) and a test set (n=134) in a 7:3 ratio. Using clinical information, multimodal ultrasound features, and radiomics features, a radiomics model was constructed using the Random Forest (RF) machine learning algorithm. Logistic regression was employed to develop the multimodal ultrasound model and the combined model. The predictive efficiency and accuracy of these models were evaluated using Receiver Operating Characteristic (ROC) curves, calibration curves, and Decision Curve Analysis (DCA). The diagnostic efficacy of junior physicians assisted by the ultrasound radiomics model was compared with that of senior physicians. DeLong’s test was performed to compare the diagnostic performance of the models.

Multivariate analysis revealed that age (≤51 years), Sound Touch Elastography mean stiffness (STE Mean), orientation (vertical), margin (blurred), and margin (irregular) were independent risk factors for papillary thyroid carcinoma, and the multimodal ultrasound model was established. Based on 17 ultrasound radiomics features, a radiomics model was constructed using the RF machine learning algorithm. The combined model was developed by combining the two aforementioned models. In the training set, the areas under the curve (AUC) of the multimodal ultrasound model, ultrasound radiomics model, and combined model were 0.852, 0.940 and 0.956, respectively. In the test set, the AUC were 0.804, 0.832 and 0.863, respectively. DeLong’s test showed that the combined model performed best in the training set, and in the test set, the combined model outperformed the multimodal ultrasound model but showed no significant difference compared to the radiomics model. DCA indicated that the combined model achieved higher net benefits within a specific threshold probability range (0.15-0.90).

The combined model exhibits robust diagnostic capability in distinguishing benign from malignant thyroid nodules classified as C-TIRADS 4A.