Application of multimodal ultrasound radiomics in the diagnosis of superficial lymph node tuberculosis.
Authors
Affiliations (4)
Affiliations (4)
- Department of Ultrasound, Hangzhou Red Cross Hospital, Hangzhou, China.
- Department of Ultrasound, Affiliated Hangzhou First People's Hospital, School of Medicine, Westlake University, Hangzhou, China.
- Department of Ultrasound, Affiliated Hangzhou First People's Hospital, School of Medicine, Westlake University, Hangzhou, China. [email protected].
- Affiliated Hangzhou First People's Hospital, School of Medicine, Westlake University, 50 Jingxin Street, Gongshu District, Hangzhou City, Zhejiang Province, 310000, China. [email protected].
Abstract
To develop preoperative diagnostic models for superficial lymph node tuberculosis (LNTB) using radiomic features extracted from multimodal ultrasound imaging, including gray-scale ultrasound (US), ultrasound elastography (UE), and contrast-enhanced ultrasound (CEUS), in conjunction with various machine learning algorithms. A retrospective study was conducted on 222 patients with lymphadenopathy. The patients were randomly divided into a training group (n = 156) and a validation group (n = 66) in a 7:3 ratio. 837 radiomics features were extracted from images of each modality (US, UE and CEUS). After initial screening by hypothesis testing, the least absolute shrinkage and selection operator (LASSO) regression with five-fold cross-validation was used for feature dimensionality reduction and selection. After feature selection, five machine learning models-logistic regression, decision tree, random forest, support vector machine, and AdaBoost-were used to construct radiomic-based models. Receiver operating characteristic (ROC) curves were generated, and the area under the curve (AUC) was computed to assess the performance of each model in predicting superficial LNTB. Clinical decision curve (DCA) is used to measure the net benefits under various probability thresholds. The diagnostic performance of ultrasound physicians was also compared with that of the best-performing machine learning model. Among the models generated by different algorithms, the decision tree model exhibited the best performance, achieving an AUC of 0.909 (95% CI, 0.789-0.949) in the training cohort and 0.866 (95% CI, 0.774-0.958) in the validation cohort. The AUC, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) for ultrasound physicians were 0.693 (95% CI, 0.568-0.818), 0.698, 0.689, 0.722, and 0.664, respectively. The DCA shows that the decision tree model has the best net income in the range of clinical relevance threshold of 0.6-0.8. Delong test showed that decision tree model was superior to ultrasonic doctor's diagnosis (Z = 2.98, p<0.0029). The radiomic model constructed from US, UE, and CEUS demonstrated robust diagnostic performance for superficial LNTB, with the decision tree model yielding the best results.