Multimodal Deep Fusion of Ultrasound Images and Clinical Factors for Pre-operative Prediction of Lateral Cervical Lymph Node Metastasis in Papillary Thyroid Carcinoma.
Authors
Affiliations (3)
Affiliations (3)
- Department of Ultrasound, Guangdong Provincial People's Hospital (Guangdong Academy of Medical Sciences), Southern Medical University, Guangzhou, Guangdong, China.
- School of Computer Science & Information Security, Guilin University of Electronic Technology, Guilin, Guangxi, China.
- Department of Ultrasound, Guangdong Provincial People's Hospital (Guangdong Academy of Medical Sciences), Southern Medical University, Guangzhou, Guangdong, China. Electronic address: [email protected].
Abstract
Lateral cervical lymph node metastasis (LLNM) is missed in up to 30% of papillary thyroid carcinoma (PTC) patients pre-operatively, leading to incomplete surgery or unnecessary two-stage lateral neck dissection. Here we aimed to develop and externally validate a multimodal deep learning network (M-DLN) that fuses ultrasound images with clinical factors for the pre-operative prediction of LLNM. In this retrospective, multicentre study, 1019 consecutive PTC patients (2014-2024) from two tertiary centres were split into training (n = 710), internal validation (n = 87), internal test (n = 87) and external validation (n = 135) cohorts. Three ResNet-50 backbones extracted features from type B, real-time tissue elastography and monochrome superb microvascular imaging ultrasound; a three-layer, fully connected network processed 15 clinical variables. Decision-level soft-voting fused image and clinical probabilities. Model explainability was provided by gradient-weighted class activation mapping and Shapley additive explanations. Discrimination, calibration and clinical utility were assessed by area under the curve and decision curve analysis. In both the internal test cohort and external validation cohort, the M-DLN model-comprising a deep learning network based on multimodal ultrasound images integrated with clinical information-demonstrated a highly efficient and robust performance in predicting LLNM in PTC, with areas under the curve of 0.901 (95% confidence interval: 0.874-0.928) and 0.853 (95% confidence interval: 0.825-0.876), respectively. Decision curve analysis indicated that M-DLN provided a substantial net clinical benefit. In the external validation cohort, our M-DLN model yielded a high sensitivity of 0.946, a moderate specificity of 0.652 and an overall accuracy of 0.874. The high sensitivity effectively reduced the rate of missed LLNM diagnosis, which is essential to avoid under-treatment, while the moderate specificity was acceptable given the clinical priorities. The open-source M-DLN system, integrating routinely acquired multimodal ultrasound images with clinical data, provides accurate, interpretable and externally validated pre-operative identification of LLNM in PTC, and could guide initial thyroid surgery precision.