Deep learning to assist composition classification and thyroid solid nodule diagnosis: a multicenter diagnostic study.

OBJECTIVES: This study aimed to propose a deep learning (DL)-based framework for identifying the composition of thyroid nodules and assessing their malignancy risk. METHODS: We conducted a retrospective multicenter study using ultrasound images from four hospitals. Convolutional neural network (CNN) models were constructed to classify ultrasound images of thyroid nodules into solid and non-solid, as well as benign and malignant. A total of 11,201 images of 6784 nodules were used for training, validation, and testing. The area under the receiver-operating characteristic curve (AUC) was employed as the primary evaluation index. RESULTS: The models had AUCs higher than 0.91 in the benign and malignant grading of solid thyroid nodules, with the Inception-ResNet AUC being the highest at 0.94. In the test set, the best algorithm for identifying benign and malignant thyroid nodules had a sensitivity of 0.88, and a specificity of 0.86. In the human vs. DL test set, the best algorithm had a sensitivity of 0.93, and a specificity of 0.86. The Inception-ResNet model performed better than the senior physicians (p < 0.001). The sensitivity and specificity of the optimal model based on the external test set were 0.90 and 0.75, respectively. CONCLUSIONS: This research demonstrates that CNNs can assist thyroid nodule diagnosis and reduce the rate of unnecessary fine-needle aspiration (FNA). CLINICAL RELEVANCE STATEMENT: High-resolution ultrasound has led to increased detection of thyroid nodules. This results in unnecessary fine-needle aspiration and anxiety for patients whose nodules are benign. Deep learning can solve these problems to some extent. KEY POINTS: • Thyroid solid nodules have a high probability of malignancy. • Our models can improve the differentiation between benign and malignant solid thyroid nodules. • The differential performance of one model was superior to that of senior radiologists. Applying this could reduce the rate of unnecessary fine-needle aspiration of solid thyroid nodules.

AI diagnosis of Bethesda category IV thyroid nodules.

Thyroid nodules are a common disease, and fine needle aspiration cytology (FNAC) is the primary method to assess their malignancy. For the diagnosis of follicular thyroid nodules, however, FNAC has limitations. FNAC can classify them only as Bethesda IV nodules, leaving their exact malignant status and pathological type undetermined. This imprecise diagnosis creates difficulties in selecting the follow-up treatment. In this retrospective study, we collected ultrasound (US) image data of Bethesda IV thyroid nodules from 2006 to 2022 from five hospitals. Then, US image-based artificial intelligence (AI) models were trained to identify the specific category of Bethesda IV thyroid nodules. We tested the models using two independent datasets, and the best AI model achieved an area under the curve (AUC) between 0.90 and 0.95, demonstrating its potential value for clinical application. Our research findings indicate that AI could change the diagnosis and management process of Bethesda IV thyroid nodules.

Value of Artificial Intelligence in Improving the Accuracy of Diagnosing TI-RADS Category 4 Nodules.

OBJECTIVE: Considerable heterogeneity is observed in the malignancy rates of thyroid nodules classified as category 4 according to the Thyroid Imaging Reporting and Data System (TI-RADS). This study was aimed at comparing the diagnostic performance of artificial intelligence algorithms and radiologists with different experience levels in distinguishing benign and malignant TI-RADS 4 (TR4) nodules. METHODS: Between January 2019 and September 2022, 1117 TR4 nodules with well-defined pathological findings were collected for this retrospective study. An independent external data set of 125 TR4 nodules was incorporated for testing purposes. Traditional feature-based machine learning (ML) models, deep convolutional neural networks (DCNN) models and a fusion model that integrated the prediction outcomes from all models were used to classify benign and malignant TR4 nodules. A fivefold cross-validation approach was employed, and the diagnostic performance of each model and radiologists was compared. RESULTS: In the external test data set, the area under the receiver operating characteristic curve (AUROC) of the three DCNN-based secondary transfer learning models-InceptionV3, DenseNet121 and ResNet50-were 0.852, 0.837 and 0.856, respectively. These values were higher than those of the three traditional ML models-logistic regression, multilayer perceptron and random forest-at 0.782, 0.790, and 0.767, respectively, and higher than that of an experienced radiologist (0.815). The fusion diagnostic model we developed, with an AUROC of 0.880, was found to outperform the experienced radiologist in diagnosing TR4 nodules. CONCLUSION: The integration of artificial intelligence algorithms into medical imaging studies could improve the accuracy of identifying high-risk TR4 nodules pre-operatively and have significant clinical application potential.

Comparison of machine learning models based on multi-parametric magnetic resonance imaging and ultrasound videos for the prediction of prostate cancer.

Objective: To establish machine learning (ML) prediction models for prostate cancer (PCa) using transrectal ultrasound videos and multi-parametric magnetic resonance imaging (mpMRI) and compare their diagnostic performance. Materials and methods: We systematically collated the data of 383 patients, including 187 with PCa and 196 with benign lesions. Of them, 307 patients (150 with PCa and 157 with benign lesions) were randomly selected to train and validate the ML models, 76 patients were used as test set. B-Ultrasound videos (BUS), mpMRI T2 sequence (T2), and ADC sequence (ADC) were obtained from all patients. We extracted 851 features of each patient in the BUS, T2, and ADC groups and used a t-test, the Mann-Whitney U test, and LASSO regression to screen the features. Support vector machine (SVM), random forest (RF), adaptive boosting (ADB), and gradient boosting machine (GBM) models were used to establish radiomics models. In addition, we fused the features screened via LASSO regression from three groups as new features and rebuilt ML models. The performance of the ML models in diagnosing PCa in the BUS, T2, ADC, and fusion groups was compared using the area under the ROC curve (AUC), sensitivity, specificity, and accuracy. Results: In the test cohort, the AUC of each model in the ADC group was higher than that of in the.BUS and T2 groups. Among the models, the RF model had the best diagnostic performance, with an AUC of 0.85, sensitivity of 0.78 (0.61-0.89), specificity of 0.84 (0.69-0.94), and accuracy of 0.83 (0.66-0.93). The SVM model in both the BUS and T2 groups performed best. Based on the features screened in the BUS, T2, and ADC groups fused to construct the models, the SVM model was found to perform best, with an AUC of 0.87, sensitivity of 0.73 (0.56-0.86), specificity of 0.79 (0.63-0.90), and accuracy of 0.77 (0.59-0.89). The difference in the results was statistically significant (p<0.05). Conclusion: The ML prediction models had a good diagnostic ability for PCa. Among them, the SVM model in the fusion group showed the best performance in diagnosing PCa. These prediction models can help radiologists make better diagnoses.

Differentiating non-lactating mastitis and malignant breast tumors by deep-learning based AI automatic classification system: A preliminary study.

Objective: To explore the application values of deep-learning based artificial intelligence (AI) automatic classification system, on the differential diagnosis of non-lactating mastitis (NLM) and malignant breast tumors, via its comparation with traditional ultrasound interpretations and the following interpretation conclusions made by the sonographers with various seniorities. Methods: A total of 707 patients suffering from breast lesions (475 malignant breast tumors and 232 NLM), were selected from the following three medical centers, including Zhejiang Cancer Hospital, Hebei Province Hospital of Traditional Chinese Medicine, and Yantai Affiliated Hospital of Binzhou Medical University, and the time period was set from April 2020 to September 2021. All selected cases firstly accepted the routine breast ultrasound diagnosis, followed by the interpretations from a senior sonographer with more than 15 years of work experience, and an intermediate-aged sonographer with more than 5 years of work experience, independently. Meanwhile, a third physician also interpreted the same ultrasound images by deep learning-based AI automatic classification system, independent of the interpretation results from the previous two physicians. The kappa test was performed to evaluate the consistency between the conventional ultrasound interpretation results and pathological results interpreted from physicians with different working experiences. Results: In total, 475 cases of malignant breast tumors (512 nodules) and 232 cases of NLM (255 nodules) were pathologically diagnosed. The accuracy, sensitivity, and specificity of conventional ultrasound interpretations vary from different sonographers with different working experiences. The accuracy, sensitivity, and specificity for intermediate-aged sonographers and senior sonographers were 76.92% (590/767), 84.71% (216/255), and 73.95% (374/512) and 87.35% (670/767), 86.27% (220/255), and 87.89% (450/512), respectively (P<0.001). In contrast, if the threshold was set as 0.5, the accuracy, sensitivity, and specificity from deep learning-based AI automatic classification system were 83.00%, 87.20%, and 85.33%, separately, and the area under the curve was 92.6. The results of the kappa consistency test indicated that the diagnosis results from the image interpretations by senior physicians and deep-learning based AI automatic classification system showed high consistency with postoperative pathological diagnosis results, and the kappa values are 0.72 and 0.71, respectively, with the P-value of less than 0.001. In contrast, the consistency between the image interpretation results from intermediate-aged physicians with less working experience, and postoperative pathological diagnosis results, seemed to be relatively lower, with a kappa value of only 0.53 and P-value of less than 0.001. Conclusions: The deep learning-based AI automatic classification system is expected to become a reliable auxiliary way to distinguish NLM and malignant breast tumors due to its high sensitivity, accuracy, and specificity.

Machine learning prediction of prostate cancer from transrectal ultrasound video clips.

Objective: To build a machine learning (ML) prediction model for prostate cancer (PCa) from transrectal ultrasound video clips of the whole prostate gland, diagnostic performance was compared with magnetic resonance imaging (MRI). Methods: We systematically collated data from 501 patients-276 with prostate cancer and 225 with benign lesions. From a final selection of 231 patients (118 with prostate cancer and 113 with benign lesions), we randomly chose 170 for the purpose of training and validating a machine learning model, while using the remaining 61 to test a derived model. We extracted 851 features from ultrasound video clips. After dimensionality reduction with the least absolute shrinkage and selection operator (LASSO) regression, 14 features were finally selected and the support vector machine (SVM) and random forest (RF) algorithms were used to establish radiomics models based on those features. In addition, we creatively proposed a machine learning models aided diagnosis algorithm (MLAD) composed of SVM, RF, and radiologists' diagnosis based on MRI to evaluate the performance of ML models in computer-aided diagnosis (CAD). We evaluated the area under the curve (AUC) as well as the sensitivity, specificity, and precision of the ML models and radiologists' diagnosis based on MRI by employing receiver operator characteristic curve (ROC) analysis. Results: The AUC, sensitivity, specificity, and precision of the SVM in the diagnosis of PCa in the validation set and the test set were 0.78, 63%, 80%; 0.75, 65%, and 67%, respectively. Additionally, the SVM model was found to be superior to senior radiologists' (SR, more than 10 years of experience) diagnosis based on MRI (AUC, 0.78 vs. 0.75 in the validation set and 0.75 vs. 0.72 in the test set), and the difference was statistically significant (p< 0.05). Conclusion: The prediction model constructed by the ML algorithm has good diagnostic efficiency for prostate cancer. The SVM model's diagnostic efficiency is superior to that of MRI, as it has a more focused application value. Overall, these prediction models can aid radiologists in making better diagnoses.

Value of deep learning models based on ultrasonic dynamic videos for distinguishing thyroid nodules.

Objective: This study was designed to distinguish benign and malignant thyroid nodules by using deep learning(DL) models based on ultrasound dynamic videos. Methods: Ultrasound dynamic videos of 1018 thyroid nodules were retrospectively collected from 657 patients in Zhejiang Cancer Hospital from January 2020 to December 2020 for the tests with 5 DL models. Results: In the internal test set, the area under the receiver operating characteristic curve (AUROC) was 0.929(95% CI: 0.888,0.970) for the best-performing model LSTM Two radiologists interpreted the dynamic video with AUROC values of 0.760 (95% CI: 0.653, 0.867) and 0.815 (95% CI: 0.778, 0.853). In the external test set, the best-performing DL model had AUROC values of 0.896(95% CI: 0.847,0.945), and two ultrasound radiologist had AUROC values of 0.754 (95% CI: 0.649,0.850) and 0.833 (95% CI: 0.797,0.869). Conclusion: This study demonstrates that the DL model based on ultrasound dynamic videos performs better than the ultrasound radiologists in distinguishing thyroid nodules.