A Multimodal Ultrasound Radiomics Model Integrating Contrast-Enhanced Phases: Retrospective Temporal Test Cohort Evaluation and Comparison with Sonographers for Cervical Lymph Node Characterization.
Authors
Affiliations (2)
Affiliations (2)
- Department of Ultrasound, The Fourth Affiliated Hospital of Harbin Medical University, Harbin 150001, Heilongjiang, China (M.G., Q.L., H.H., N.Z., Y.G., K.M.).
- Department of Ultrasound, The Fourth Affiliated Hospital of Harbin Medical University, Harbin 150001, Heilongjiang, China (M.G., Q.L., H.H., N.Z., Y.G., K.M.). Electronic address: [email protected].
Abstract
To develop a multimodal ultrasound radiomics model based on conventional ultrasound and three-phase contrast-enhanced ultrasound (CEUS) for distinguishing benign from malignant cervical lymph nodes, and to evaluate its performance in a patient-level temporal test cohort against sonographers with different experience levels. This retrospective study included 245 patients in the development cohort (August 2022-December 2024) and 60 subsequent patients in the temporal test cohort (January-December 2025). One target lymph node was analyzed per patient, and the two cohorts were strictly non-overlapping at the patient level. For each case, matched two-dimensional ultrasound (2D-US), color Doppler flow imaging (CDFI), and arterial-, parenchymal-, and clearance-phase CEUS key frames were collected. Deep features were extracted from manually cropped ROIs using a frozen pre-trained ResNet50. Feature selection, hyperparameter tuning, and decision-threshold locking were completed only within the development cohort; the temporal test cohort was used once for final reporting. Sonographers reviewed the same static key-frame image set in randomized order under blinded conditions. The CEUS-integrated SVM model (Model 2-SVM) showed the best overall performance among the tested models. In the temporal test cohort, it achieved an area under the curve (AUC) of 0.851 (95% CI: 0.762-0.940), sensitivity of 84.8% (95% CI: 69.1-93.3), specificity of 66.7% (95% CI: 47.8-81.4), accuracy of 76.7% (95% CI: 64.6-85.6), PPV of 75.7% (95% CI: 59.9-86.6), and NPV of 78.3% (95% CI: 58.1-90.3). The senior and junior sonographers achieved AUCs of 0.826 and 0.783, respectively. At the validation-locked threshold, the model showed numerically higher sensitivity than both sonographers but lower specificity than the senior sonographer. Threshold-free and paired statistical comparisons are reported in the corresponding tables. A multimodal ultrasound radiomics model incorporating CEUS key-frame information demonstrated feasible diagnostic performance in a retrospective temporal test cohort and may serve as a hypothesis-generating adjunct for cervical lymph node assessment. This retrospective single-center design, manual ROI delineation, and the absence of prospective external validation require further multi-center evaluation before clinical application.