Multiparametric MRI-Based Habitat Radiomics Combined with Deep Transfer Learning for Predicting Extrathyroidal...

The objective of the study is to develop and validate a multiparametric MRI (mpMRI)-based model that integrated with habitat-based radiomics, deep transfer learning (DTL), and quantitative parameters for the preoperative prediction of extrathyroidal extension (ETE) in papillary thyroid carcinoma (PTC). This retrospective study collected clinical, pathological, and mpMRI data from patients with confirmed PTC who had a thyroid mpMRI scan within 2 weeks before surgery and a PTC lesion of ≥ 10 mm. A total of 140 patients with 140 PTC lesions were included. Data were split into a training cohort (1.5 T MRI) and a validation cohort (3.0 T MRI). DTL models using a ResNet152 were developed to predict ETE from T2-weighted imaging (T2WI), diffusion-weighted imaging (DWI), delayed contrast-enhanced images, and apparent diffusion coefficient (ADC) maps. K-means clustering identified habitat subregions in these images, from which radiomics features were extracted and combined with DTL features. The least absolute shrinkage and selection operator (LASSO) was used for feature selection to build habitat-based deep learning radiomics (DLR) models for each mpMRI modality. A quantitative MRI parameter model was developed by analyzing mpMRI features through logistic regression. We combined optimal DTL and habitat-based DLR models from four MRI modalities to identify independent risk factors for predicting ETE using multivariate logistic regression and then developed a nomogram for ETE prediction in PTC. Model performance was assessed with area under the curve (AUC) and validated via tenfold cross-validation, while the DeLong test compared AUC differences. Protrusion value and ADC-best ratio were key predictors of ETE. Combined with T2WI_DTL signature and DWI_Habitat1_DLR signature, the Habitat-DTL nomogram was formed. The model had an AUC of 0.963 in the training cohort and 0.884 in the validation cohort, with no significant difference in ETE prediction between cohorts (P = 0.168). The Habitat-DTL nomogram, integrating MRI quantitative parameters, DTL, and habitat-based DLR, demonstrated promising performance for the preoperative prediction of ETE in PTC, serving as a potential noninvasive tool to facilitate clinical decision-making and personalized treatment.