The Predictive Value Of Multimodal Deep Learning Combining Ultrasound And CT For Central Cervical Lymph Node Metastasis In Papillary Thyroid Carcinoma

Objective The aim of this study was to investigate the predictive value of deep learning for central cervical lymph node metastasis(CLNM)in papillary thyroid carcinoma(PTC)by combining two-dimensional ultrasound images and neck-enhanced CT images.Methods Data were obtained from the PACS system of Zhejiang Cancer Hospital between August 2018 and August 2021.Patients with PTC,confirmed by postoperative routine pathology and with pathological information on regional lymph nodes,were enrolled.Clinical information(including age,gender,nodule location,nodule size,ultrasound characteristics,pathological results,etc.)and the corresponding preoperative ultrasound and neck-enhanced CT images were collected.The ResUNet was used as the backbone network.Since ultrasound offers higher resolution for thyroid nodule texture and CT better displays the relationship between nodules and surrounding tissues,we attempted to fuse the texture feature of nodules in ultrasound images and the relationship between nodules and surrounding tissues in CT images to predict central CLNM.Results The accuracy,sensitivity,and specificity of preoperative ultrasound diagnosis of central CLNM were 0.590(482/817),0.296(128/432),and 0.919(354/385),respectively.The area under the curve(AUC)of CT single modality was 0.795(95%CI:0.700-0.870),with an accuracy of 0.802,sensitivity of 0.917,specificity of 0.688,precision of 0.746,and F1 score of 0.822.The AUC of ultrasound single modality was 0.866(95%CI:0.782-0.927),with an accuracy of 0.781,sensitivity of 0.854,specificity of 0.708,precision of 0.745,and F1 score of 0.796.The AUC of ultrasound combined with CT multimodality was 0.870(95%CI:0.786-0.930),with an accuracy of 0.823,sensitivity of 0.833,specificity of 0.813,precision of 0.816,and F1 score of 0.823.Conclusion:In this study,a multimodal deep learning model combining ultrasound and CT was constructed by feature fusion of ultrasound and CT images of PTC primary tumors.Compared with ultrasound or CT single modality,the prediction efficiency of the multimodality model was improved.When compared to ultrasound doctors,it demonstrated higher sensitivity but slightly lower specificity.This multimodal model may be used for preliminary screening of central CLNM,assisting doctors in improving the accuracy of preoperative CLNM diagnosis,and contributing to reasonable preoperative risk stratification of PTC patients to reduce unnecessary surgery and lymph node dissection.However,external data sets must still be used for validation.