Abstract:

Objective To explore the predictive value of multimodal ultrasound radiomics features for lymphovascular invasion （LVI） in papillary thyroid cancer （PTC）. Methods A total of 131 patients with PTC who underwent multimodal ultrasound examination and were confirmed by surgical pathology at People's Hospital of Xinjiang Uygur Autonomous Region from June 2024 to March 2025 were retrospectively collected. Patients were randomly divided into a training set （n=92） and a validation set （n=39） at a ratio of 7∶3. The training set was further divided into an LVI group （n=23） and a non-LVI group （n=69） according to whether the patients had LVI. Multivariate logistic regression analysis was used to identify factors influencing LVI and construct a clinical model. The least absolute shrinkage and selection operator （LASSO） algorithm was applied to screen radiomics features from different ultrasound modalities and develop a multimodal ultrasound radiomics score （Radscore） model. Four deep learning models were constructed using four machine learning algorithms with the selected radiomics features. A fusion model was developed by integrating logistic regression analysis. Model performance was evaluated using receiver operator characteristic （ROC） curves， calibration curves， and clinical decision curve. Results Statistically significant differences were observed between LVI and non-LVI groups in maximum tumor diameter （t=3.30， P=0.001）， tumor multifocality （χ²=17.97， P<0.001）， enhancement rate （χ²=5.69， P=0.017）， and wash-in perfusion index （WiPI） （t=2.69， P=0.009）. The results of the multivariate analysis showed that， maximum tumor diameter （OR=1.26， 95%CI：1.09-1.47， P=0.002）， WiPI （OR=0.87， 95%CI：0.78-0.97， P=0.009） and tumor multifocality （OR=6.05， 95%CI：1.97-18.60， P=0.002） were all independent influencing factors for the occurrence of LVI in patients with PTC. After dimensionality reduction via LASSO regression， 6 radiomic features for conventional ultrasound and 11 radiomic features for contrast-enhanced ultrasound were obtained. A multimodal ultrasound model was constructed based on the aforementioned features：logit（P）=-0.27+0.17×conventional ultrasound+0.30×contrast-enhanced ultrasound. The multimodal ultrasound Radscores of the LVI group patients in the training set and validation set （3.38±1.20， 4.02±1.45） were significantly higher than those of the non-LVI patients （1.76±0.83， 1.40±0.54）， with statistically significant differences （t=7.20， P<0.001； t=8.28， P<0.001）. Among the 4 constructed deep learning models， the support vector machine model exhibited the best performance and could be used for subsequent studies. ROC curve analysis showed that the fusion model， which was constructed by combining the clinical model， multimodal ultrasound radiomics model， and deep learning model， had an area under the curve of 0.90 （95%CI：0.81-0.95） for diagnosing LVI， representing the optimal diagnostic performance. Calibration curve analysis indicated that the predictive probabilities of the fusion model showed good agreement with the actual probabilities and closely approached the ideal curve. Decision curve analysis showed that the fusion model had the highest net benefit. Conclusions Multimodal ultrasound radiomics models exhibit good predictive performance for LVI in PTC patients. The fusion model combining multimodal ultrasound radiomics， clinical features， and deep learning can improve diagnostic efficiency， demonstrating high clinical application value. This approach shows certain calibration abilities and is expected to provide an objective basis for precise and personalized clinical treatment.

Key words: Thyroid cancer， papillary, Lymphovascular invasion, Multimodal ultrasound radiomics