Prediction of gross extrathyroidal extension in papillary thyroid carcinoma using a deep learning radiomics model...

Gross extrathyroidal extension (ETE) affects surgical management and prognosis in papillary thyroid carcinoma (PTC), yet noninvasive assessment is difficult.

In this retrospective exploratory feasibility study, 101 patients with pathologically confirmed PTC who underwent multi-phase contrast-enhanced MRI (MCE MRI) were analyzed. From the second and seventh contrast phases, four derived image types (ADD, ADD/SUB, Max, Min) were generated. Deep features from six image modalities were extracted using fine-tuned ResNet101, and radiomics classifiers were trained to predict gross ETE. Logistic regression identified independent clinical and MRI predictors, which, together with the deep learning radiomic signatures (DLR_signature) derived from the best-performing derived image type and algorithm, were integrated into a nomogram. Model performance was assessed by ROC, calibration, decision curve analysis, and ten-fold cross-validation.

Gross ETE was present in 37/101 cases (36.6%). The Min-modality deep transfer learning model achieved AUCs of 0.852 (training) and 0.853 (validation); the corresponding radiomics classifier had AUCs of 0.870 and 0.805. Multivariate analysis identified mean apparent diffusion coefficient (ADC_mean) (P = 0.001) and protrusion_value (P < 0.001) as independent predictors. The integrated nomogram combining Min-modality DLR_signature, protrusion_value, and ADC_mean achieved AUCs of 0.956 (training) and 0.895 (validation), with average AUCs of 0.875 and 0.846 in cross-validation. Calibration and decision-curve analyses showed favorable performance in the current dataset.

The proposed nomogram integrating a Min-modality DLR_signature with quantitative MRI parameters showed promising potential for the noninvasive prediction of gross ETE in PTC. These findings support the feasibility of using MCE MRI post-processing for preoperative ETE prediction.