Abstract
OBJECTIVE: To extract and analyze the image features of two-dimensional ultrasound images and elastic images of four thyroid nodules by radiomics, and then further convolution processing to construct a prediction model for thyroid cancer. The purpose of this study was to explore the diagnostic efficacy of the model. METHODS: In this study, 199 cases of thyroid nodules were collected from August 2023 to July 2024, and all thyroid nodules had B-ultrasound-guided fine needle aspiration biopsy (FNA) pathological results/postoperative pathological results, including 79 cases of benign nodules and 120 cases of malignant nodules. In this study, four thyroid cancer prediction models were constructed and compared, including convolutional neural network (CNN), gradient boosting (GB), logistic regression (LR), and ultrasound and clinical feature models. In addition, the clinical feature model was constructed by using the clinical information of patients and ultrasound image features, and the predictive performance of four thyroid cancer models was evaluated and compared. The area under the receiver operating characteristic curve (AUC), accuracy, specificity, and sensitivity were used to validate the predictive power of the model. Finally, we used the Delong test to compare whether there was a significant difference in AUC between the four models. RESULTS: The CNN model performed well in the Area Under the Curve (AUC) and ACC (Accuracy) indicators, reaching 0.853 and 0.85, respectively, which were significantly better than the Gradient Boosting, Logistics regression and clinical characteristics models. The AUC, ACC, SPE, and SEN of the Gradient Boosting model were 0.653, 0.67, 0.709, and 0.63, respectively, the Logistics regression model was 0.701, 0.71, 0.6, and 0.714, and the clinical characteristic model was 0.663, 0.69, 0.708, and 0.57, respectively. The outstanding performance of CNN highlights its potential in the field of image recognition. SUMMARY: CNN model has shown strong predictive ability in ultrasound image analysis of suspicious thyroid nodules, which not only provides a powerful auxiliary diagnostic tool for clinicians, but also provides new directions and possibilities for future medical image analysis research.