CT Radiomics-Based Machine Learning Model for Predicting Capsular and Neural Invasion in Thyroid Carcinoma: Diagnostic Accuracy Study.
Thyroid carcinoma is the most prevalent endocrine malignancy, with a worldwide increasing incidence. Capsular invasion and neural invasion (NI) are pivotal prognostic factors for recurrence and survival; however, their preoperative noninvasive assessment remains challenging.
We aimed to identify computed tomography (CT) radiomic biomarkers associated with capsular invasion in thyroid carcinoma, construct machine learning models integrating radiomic and clinical data, and evaluate their utility for NI risk stratification.
In this retrospective cohort, 111 patients with thyroid carcinoma were divided into capsular invasion-positive (n=63) and capsular invasion-negative (n=48) groups, with 37 (33.3%) cases presenting concurrent NI. Radiomic features were extracted from arterial and venous phase CT images at original resolution, including 111 gray-level co-occurrence matrix features. Nine key radiomic features (A1-A9) were selected via least absolute shrinkage and selection operator regression (λ=0.017). To preserve the physical meaning of texture features (eg, spatial correlation and contrast reflecting tumor microstructural heterogeneity), no resampling or scaling was performed on the regions of interest during radiomic feature extraction. Nomogram models and random forest (RF) models were constructed based on clinical indicators (galectin-3, etc) and radiomic features, respectively. Additionally, a neural network (NN) model integrating multimodal data was developed. Model stability was verified using 5-fold cross-validation and 1000-time bootstrap resampling, while performance was evaluated via receiver operating characteristic curves, calibration curves, and decision curve analysis.
Model performance analysis revealed that among the nomogram models, the clinical indicator-based nomogram achieved an internally estimated area under the curve (AUC) of 0.9418 (95% CI 0.892-0.976) in the capsular invasion prediction task. The radiomic-based nomogram had an internally estimated AUC of 0.9334 (95% CI 0.881-0.968) in the capsular invasion prediction task and 0.8001 (95% CI 0.663-0.898) in the cross-label association analysis task. In RF models, clinical indicator-based and radiomic-based RFs exhibited an AUC of 0.7646 (95% CI 0.651-0.857) and 0.8102 (95% CI 0.703-0.892) in the cross-label association analysis task, respectively. The NN model performed promisingly, with an AUC of 0.775 (95% CI 0.621-0.903) in the cross-label association analysis task and a mean absolute error of <0.05 on the calibration curve.
Capsular invasion is a strong predictor of NI risk in thyroid carcinoma. Radiomic models based solely on preoperative CT images show potential for preoperative NI risk stratification. Models incorporating clinical parameters (obtained from postoperative tissue), including the integrated multimodal model, are more accurately characterized as postoperative risk stratification tools. The NN model, which integrated raw CT images with clinical data, achieved an AUC of 0.775 (95% CI 0.621-0.903), underscoring the potential of such multimodal analysis to capture complex relationships between imaging phenotypes and tissue-level biomarkers for enhanced postoperative assessment. This framework's radiomic component points toward purely image-based, preoperative evaluation tools' development.
We aimed to identify computed tomography (CT) radiomic biomarkers associated with capsular invasion in thyroid carcinoma, construct machine learning models integrating radiomic and clinical data, and evaluate their utility for NI risk stratification.
In this retrospective cohort, 111 patients with thyroid carcinoma were divided into capsular invasion-positive (n=63) and capsular invasion-negative (n=48) groups, with 37 (33.3%) cases presenting concurrent NI. Radiomic features were extracted from arterial and venous phase CT images at original resolution, including 111 gray-level co-occurrence matrix features. Nine key radiomic features (A1-A9) were selected via least absolute shrinkage and selection operator regression (λ=0.017). To preserve the physical meaning of texture features (eg, spatial correlation and contrast reflecting tumor microstructural heterogeneity), no resampling or scaling was performed on the regions of interest during radiomic feature extraction. Nomogram models and random forest (RF) models were constructed based on clinical indicators (galectin-3, etc) and radiomic features, respectively. Additionally, a neural network (NN) model integrating multimodal data was developed. Model stability was verified using 5-fold cross-validation and 1000-time bootstrap resampling, while performance was evaluated via receiver operating characteristic curves, calibration curves, and decision curve analysis.
Model performance analysis revealed that among the nomogram models, the clinical indicator-based nomogram achieved an internally estimated area under the curve (AUC) of 0.9418 (95% CI 0.892-0.976) in the capsular invasion prediction task. The radiomic-based nomogram had an internally estimated AUC of 0.9334 (95% CI 0.881-0.968) in the capsular invasion prediction task and 0.8001 (95% CI 0.663-0.898) in the cross-label association analysis task. In RF models, clinical indicator-based and radiomic-based RFs exhibited an AUC of 0.7646 (95% CI 0.651-0.857) and 0.8102 (95% CI 0.703-0.892) in the cross-label association analysis task, respectively. The NN model performed promisingly, with an AUC of 0.775 (95% CI 0.621-0.903) in the cross-label association analysis task and a mean absolute error of <0.05 on the calibration curve.
Capsular invasion is a strong predictor of NI risk in thyroid carcinoma. Radiomic models based solely on preoperative CT images show potential for preoperative NI risk stratification. Models incorporating clinical parameters (obtained from postoperative tissue), including the integrated multimodal model, are more accurately characterized as postoperative risk stratification tools. The NN model, which integrated raw CT images with clinical data, achieved an AUC of 0.775 (95% CI 0.621-0.903), underscoring the potential of such multimodal analysis to capture complex relationships between imaging phenotypes and tissue-level biomarkers for enhanced postoperative assessment. This framework's radiomic component points toward purely image-based, preoperative evaluation tools' development.