Research on a Deep Learning-Based Model for Predicting Malignancy Risk of Pan-Endocrine System across Thyroid and Breast Nodules

Objective To construct a cross-organ AI model based on multi-task deep learning (DL) to achieve unified prediction of malignancy risks of thyroid and breast nodules.

Methods Clinical data of patients with thyroid nodules (n=2386) and breast nodules (n=2753) from three hospitals were collected. A multi-task deep learning model was constructed based on the Transformer architecture, and a design that combines feature sharing layers with organ-specific layers was adopted. Performance was evaluated through five-fold cross-validation, and tested on an independent external validation set (n=835), and SHAP analysis was used to explain model decisions.

Results The constructed AI model for pan-endocrine nodules achieved an AUC of 0.932 (95%CI: 0.914−0.951) in the prediction of thyroid nodules, with a sensitivity of 86.5% and a specificity of 89.2%. The predictive AUC of breast nodules was 0.917 (95%CI: 0.896−0.938), with a sensitivity of 84.3% and a specificity of 88.7%. Compared with the single-organ model, the pan-model performed better on small sample datasets (P<0.01), and remained stable in external validation (AUC>0.90). SHAP analysis revealed that edge irregularity, calcification type, and internal echo were common important features, while blood flow signals and TI-RADS/BI-RADS classification were organ-specific features.

Conclusion A high-performance cross-organ malignancy risk prediction model for thyroid-breast nodules was successfully constructed, confirming that pan-endocrine nodules can be precisely stratified for risk through a unified deep learning architecture, providing a new paradigm for AI-assisted diagnosis of endocrine tumors.