Abstract
This study retrospectively analyzed the multimodal ultrasound features and clinical characteristics of 586 patients with unexplained cervical lymphadenopathy who were treated at three hospitals between October 2019 and December 2022. Statistically significant differences were found in the clinical and ultrasound features of all patients, including location, shape, margin, and color Doppler flow imaging (CDFI) (p<0.05). Deep learning models, particularly convolutional neural networks (CNNs), demonstrated great potential in classifying cervical lymph node pathologies using multimodal ultrasound images, including 2D imaging, color Doppler flow imaging (CDFI), and elastography. First, we pre-trained four convolutional neural networks using a public medical image dataset. Then, we fine-tuned the models for three-class classification of lymph nodes into metastatic, lymphoma, and benign using 2D, CDFI, and elastography images from the patients' lymph nodes. The pre-trained ResNet model performed excellently, with an elastography AUC of 0.925, outperforming other models. Elastography became the most reliable feature extraction dataset, significantly enhancing the model's accuracy in distinguishing between benign, lymphoma, and metastatic lymph nodes. Ablation experiments showed that pre-training significantly improved accuracy compared to non-pre-trained models. Grad-CAM visualization provided valuable interpretability, revealing how the model focuses on specific areas corresponding to each pathology. Based on this model, we developed a user-friendly server, CV4LymphNode (https://hwwlab.com/webserver/cv4lymphnode). This study highlights the potential of deep learning in accurately classifying cervical lymph node pathologies.