Ultrasound‐Based Transfer Learning Model to Assist Partially Cystic Thyroid Nodule Diagnosis

ABSTRACT Purpose A transfer learning model based on ultrasound was established to predict the malignant probability of partially cystic thyroid nodule (PCTN) preoperatively, providing clinicians with a non‐invasive primary screening method. Methods 258 PCTNs of 258 patients from January 2020 to January 2024 were analyzed retrospectively. The dataset was randomly divided into a training set and a test set in a ratio of 8:2. Five different pre‐trained models were chosen for transfer learning, including EfficientNet , Inception_v3 , Mobilenet_v3 , SqueezeNet , and VGG19 . The area under the curve (AUC), accuracy, sensitivity, and specificity of the training and test cohorts were calculated. Grad‐Class Activation Map (Grad‐CAM) was drawn to interpret the results visually. All the ultrasound images were reviewed by two radiologists; multivariate logistic analyses explored the independent risk factors for malignant PCTN. The diagnostic effectiveness of transfer learning models and radiologists was compared. Results Inception_v3 model achieved the highest AUC of 0.9243 (95% CI: 0.8849–0.9439) in predicting the malignancy of PCTN in the training cohort, with an accuracy of 85.19%, sensitivity of 85.26%, and specificity of 85.00%. The diagnostic efficiency of the Inception_v3 model was better than that obtained by multivariate logistic regression analysis with AUC of 0.8247 (95% CI: 0.7579–0.8915) in the training cohort, with an accuracy of 83.33%, a sensitivity of 68.00%, and a specificity of 71.80%. Red or warm‐colored regions in Grad‐CAM represented that these features were more important to model decisions, while blue or cool‐colored regions represented those features that were less important. Conclusion Ultrasound‐based transfer learning model could predict the malignant probability of PCTN noninvasively before surgery, especially the Inception_v3 model, to assist clinical decision‐making.