Grayscale ultrasound radiomics for characterizing subpleural pulmonary lesions: a multicenter prospective study

Objectives To develop and validate a radiomics model based on grayscale ultrasound (GSUS) images for characterizing subpleural pulmonary lesions (SPLs). Materials and methods In this prospective, multicenter study, 738 patients with CT-confirmed SPLs were enrolled from three institutions and assigned to training ( n = 407), internal validation ( n = 146), and external validation ( n = 185) cohorts. A total of 1320 radiomics features were extracted from both lesion and perilesional regions on GSUS images. Feature selection was performed through intra- and inter-class correlation coefficients (ICCs) analyses, Pearson correlation analyses, and least absolute shrinkage and selection operator (LASSO) regression. Clinical–radiomics fusion models were subsequently constructed by integrating selected radiomics features with key clinical variables using multivariate logistic regression. Model performance was evaluated comprehensively using the area under the receiver-operating characteristic curve (AUC), sensitivity, specificity, F1-score, and additional diagnostic metrics. Results Five predictive models were constructed based on clinical, radiologic, and radiomics features. Among them, the integrated model combining lesion-based radiomics with clinical variables achieved the best diagnostic performance, with AUCs of 0.884 (95% CI: 0.828–0.940) in the internal validation cohort and 0.848 (95% CI: 0.791–0.904) in the external validation cohort. Calibration and decision curve analyses demonstrated good model calibration and favorable clinical utility. The diagnostic accuracy of the model was comparable to that of experienced lung ultrasound radiologists. Conclusions The GSUS-based radiomics model effectively differentiates between benign and malignant SPLs, demonstrating strong diagnostic performance and promising clinical applicability. Critical relevance statement The proposed ultrasound-based radiomics model provides a reproducible, noninvasive decision-support tool for characterizing subpleural pulmonary lesions, offering particular value in patients for whom invasive procedures are unsuitable or in settings where CT or biopsy is not readily available. Key Points Accurate characterization of subpleural pulmonary lesions remains challenging using conventional imaging techniques. The grayscale ultrasound radiomics model achieved accuracy comparable to expert radiologists. This model provides a noninvasive and accessible tool when CT or biopsy is limited.