JMIR Publications

ABSTRACT

Background:

Thyroid carcinoma represents the most prevalent malignancy of the endocrine system, with its incidence exhibiting a significant global upward trend in recent years

Objective:

This study aims to identify CT radiomic signatures associated with capsular invasion (CI) in thyroid carcinoma, screen potential radiogenomic biomarkers, and evaluate their risk stratification utility for neural invasion (NI).

Methods:

A retrospective cohort of 111 thyroid carcinoma patients was categorized into control (non-CI, n = 48) and study (CI, n = 63) groups, with 37 cases exhibiting concurrent NI. Gray-level features were extracted from CT images to generate gray-level co-occurrence matrices (GLCMs), and least absolute shrinkage and selection operator (LASSO) was used for feature dimensionality reduction. Nomogram and random forest (RF) models were constructed based on clinical indicators and gray-level features, respectively, while a neural network (NN) integrating both was developed. Model performance was evaluated using ROC, calibration, and decision curves.

Results:

A total of 111 GLCMs and 50,176 pixel gray-level features were extracted, with 9 features selected via LASSO regression. Binary logistic regression showed CI was a risk factor for NI (OR = 25.25). The nomogram and RF models based on both clinical indicators and radiomic features demonstrated comparable diagnostic efficacy (AUC > 0.7). The integrated NN achieved superior performance (AUC = 0.775).

Conclusions:

CI exhibits significant predictive value for NI risk in thyroid carcinoma. The multi-modal integration framework combining radiomic signatures and clinical parameters demonstrates enhanced risk stratification capability, providing a novel preoperative assessment paradigm for personalized surgical planning and neural preservation strategies.