Multimodal Prediction Model Integrating Clinical Parameters, Peritumoral and Habitat-Based Radiomics for Preoperative Differentiation of Follicular Thyroid Carcinoma from Adenoma: A Two-Center Study.

RATIONALE AND

OBJECTIVES

Preoperative differentiation of follicular thyroid carcinoma (FTC) from follicular thyroid adenoma (FTA) remains challenging. This study aimed to develop and validate a multimodal prediction model integrating clinical parameters, immunological markers, peritumoral radiomics, and habitat-based features for accurate preoperative FTC diagnosis.

MATERIALS AND

METHODS

This retrospective two-center study included 775 patients with pathologically confirmed follicular thyroid neoplasms (training set: n = 353; internal validation: n = 151; external validation: n = 271). Radiomics features were extracted from intratumoral regions, peritumoral extensions (1 mm, 3 mm, 5 mm), and habitat subregions identified through K-means clustering. Four machine learning algorithms (logistic regression, support vector machine, random forest, gradient boosting machine) were evaluated. Multivariate logistic regression identified independent clinical and immunological predictors. An integrated model combining optimal radiomics signatures with clinical predictors was constructed and compared with six established thyroid imaging reporting and data systems (TIRADS). Model interpretability was enhanced through SHAP analysis and nomogram visualization.

RESULTS

Multivariate analysis identified four independent predictors: thyroglobulin (TG), solid composition, hypoechoic echogenicity, and lymphocyte-to-monocyte ratio (LMR). The Peri-3mm gradient boosting machine (GBM) model achieved area under the curves (AUCs) of 0.881, 0.865, and 0.868 across cohorts. The combined Habitat GBM model demonstrated AUCs of 0.900, 0.852, and 0.860. The integrated model incorporating clinical predictors with Peri-3mm and Habitat radiomics achieved superior performance with AUCs of 0.950, 0.918, and 0.924 in training, internal validation, and external validation sets, respectively, substantially outperforming K-TIRADS (AUCs: 0.639, 0.648, 0.612). Decision curve analysis confirmed superior clinical utility of the integrated model across all threshold probabilities.

CONCLUSION

The multimodal prediction model integrating clinical, peritumoral, and habitat-based features demonstrates excellent performance for preoperative FTC differentiation, significantly surpassing conventional TIRADS systems and offering enhanced clinical decision-making support.