Intelligent Diagnosis of Follicular Carcinoma Thyroid Cancer with a Novel Deep Learning Model

Distinguishing follicular thyroid carcinoma (FTC) from follicular adenoma (FTA) preoperatively remains a significant challenge in thyroid nodule management. This study aims to develop and validate a novel, interpretable deep learning model that explicitly leverages the critical diagnostic feature-tumor margins-to address this issue. A total of 577 patients, 435 females and 142 males (mean age, 51.05 $\pm$ 8.31), were collected from two different centers. A total of 4358 thyroid US images were prospectively collected, with 3140 images from one center randomly divided into the training set and the validation set with a ratio of 4:1 for training the deep learning (DL) model, while 1218 images from the other center were viewed as a test dataset for the evaluation. We propose an end-to-end graph convolutional network that constructs a graph representation from ultrasound image patches, explicitly modeling the structural relationships between features, particularly at the tumor boundary. The model is optimized using a maximum code rate reduction (MCR²) loss to enhance feature discrimination. The overall prediction accuracy and AUC of the independent test set and validation dataset were 90.13%, 82.10%, 92.35%, and 87.36%, respectively, achieving significant and consistent improvement compared to other deep learning baselines. Our proposed model could diagnose FTC with good performance. By successfully incorporating domain knowledge and validating on multicenter data, this study represents a significant step toward reliable AI-assisted diagnosis of thyroid cancer.