Thyroid Cancer Central Lymph Node Metastasis Risk Stratification Based on Homogeneous Positioning Deep Learning

Due to the absence of definitive diagnostic criteria, there remains a lack of consensus regarding the risk assessment of central lymph node metastasis (CLNM) and the necessity for prophylactic lymph node surgery in ultrasound-diagnosed thyroid cancer. The localization of thyroid nodules is a recognized predictor of CLNM; however, quantifying this relationship is challenging due to variable measurements. In this study, we developed a differential isomorphism-based alignment method combined with a graph transformer to accurately extract localization and morphological information of thyroid nodules, thereby predicting CLNM. We collected 88,796 ultrasound images from 48,969 patients who underwent central lymph node (CLN) surgery and utilized these images to train our predictive model, ACE-Net. Furthermore, we employed an interpretable methodology to explore the factors influencing CLNM and generated a risk heatmap to visually represent the distribution of CLNM risk across different thyroid regions. ACE-Net demonstrated superior performance in 6 external multicenter tests (AUC = 0.826), surpassing the predictive accuracy of human experts (accuracy = 0.561). The risk heatmap enabled the identification of high-risk areas for CLNM, likely correlating with lymphatic metastatic pathways. Additionally, it was observed that the likelihood of metastasis exceeded 80% when the nodal margin's minimum distance from the thyroid capsule was less than 1.25 mm. ACE-Net's capacity to effectively predict CLNM and provide interpretable disease-related insights can importantly reduce unnecessary lymph node dissections by 37.9%, without missing positive cases, thus offering a valuable tool for clinical decision-making.