Abstract / Summary
To develop and validate a two-stage two-path deep learning network (TTN) for automated nodule segmentation and prediction of central cervical lymph node metastasis (CLNM) in patients with small papillary thyroid carcinoma (PTC) eligible for active surveillance (AS) using ultrasound (US) images. We collected US images from PTC patients meeting AS criteria at our hospital (2032 images, 1016 patients) and an independent external test set from another hospital (200 images, 100 patients). CLNM status was confirmed by pathological examination. The TTN architecture consists of a first-stage U²-Net for nodule segmentation and a second-stage parallel ResNet for CLNM prediction, incorporating both transverse and longitudinal views together with clinical/US features. The model was trained on 90% of our dataset and tested on the remaining 10% (internal test set) and the external test set. Two multi-task learning networks were constructed for comparison. Model performance was evaluated using the area under the curve (AUC), sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and F1 score. The TTN model achieved superior segmentation and CLNM prediction performance in both internal and external validation sets, significantly outperforming the multi-task networks (all p < 0.05). In the internal validation cohort, the TTN model achieved an AUC of 0.849 (95% CI: 0.833, 0.864), with sensitivity of 90.0% (95% CI: 88%, 93%), specificity of 65% (95% CI: 63%, 68%), PPV of 81% (95% CI: 78%, 84%), NPV of 77% (95% CI: 75%, 80%), and F1 score of 0.86 (95% CI: 0.84, 0.89). In the external validation cohort, the model yielded an AUC of 0.823 (95% CI: 0.805, 0.840), sensitivity of 90.0% (95% CI: 87%, 93%), specificity of 54% (95% CI: 51%, 57%), PPV of 77% (95% CI: 74%, 80%), NPV of 77% (95% CI: 75%, 80%), and F1 score of 0.83 (95% CI: 0.80, 0.86). The TTN model provides an automated, accurate, and generalizable tool for simultaneous nodule segmentation and CLNM prediction in AS-eligible patients with small PTC. By non-invasively identifying those at risk of CLNM, it can assist clinicians in personalizing treatment decisions-supporting timely surgery for high-risk patients while avoiding unnecessary intervention in low-risk individuals.