Diagnosis of thyroid nodules using ultrasound images based on deep learning features: online dynamic nomogram and gradient-weighted class activation mapping.

Background: The high incidence of thyroid nodules (TNs) necessitates accurate and effective differentiation between benign and malignant cases to avoid overtreatment, which is crucial for both patients and doctors. The aim of this study was to develop and visualize an integrated model to enhance the diagnostic capability of young radiologists in evaluating TNs. Methods: A retrospective collection of 1,501 ultrasound (US) images of TNs were randomly divided into training and validation sets. An independent test set comprised 541 patients from The First Affiliated Hospital of Anhui Medical University and the Affiliated Hospital of Integration Chinese and Western Medicine with Nanjing University of Traditional Chinese Medicine. We fine-tuned five ImageNet-pretrained deep learning (DL) models via transfer learning (TL) on US images to generate prediction scores and construct the final model. Gradient-weighted class activation mapping (Grad-CAM) was employed to highlight sensitive areas of the US images that contribute to nodule classification. Additionally, a comprehensive model was established utilizing both US image features and DL features, and an online dynamic nomogram was subsequently created for practical application. Models were compared and evaluated for discrimination, calibration, and effectiveness, to ascertain whether the DL model can improve radiologists' diagnosis of TNs. Results: The DL model demonstrated superior performance compared to the US model, with area under the receiver operating characteristic (ROC) curve (AUC) values of 0.875 and 0.787 on the test set, respectively. When combined into a comprehensive diagnostic model, a significant improvement was observed (test set AUC: 0.907). The Net Reclassification Index (NRI) results indicate that the heat maps and scores output by the DL model help to improve radiologists' classification accuracy in distinguishing between benign and malignant nodules. Conclusions: The integrated model based on US image features and DL features demonstrates good diagnostic performance for distinguishing between benign and malignant TNs.