Random Forest-Based Detection of Metastases in Clinically Scanned Lymph Nodes Using Quantitative Ultrasound Imaging.

Abstract

Objective: Quantitative ultrasound (QUS) imaging has been used to characterize the microstructural properties of tissue using information contained in the backscattered radiofrequency (RF) echo signals. QUS methods were previously applied to detect metastases in excised human lymph nodes (LNs) that were raster scanned using a 30 MHz single-element transducer ex vivo. In the current study, a QUS-based method to detect in vivo LN metastases using a clinical scanner was developed.

Methods: Parallel RF frames were captured from 46 cervical and axillary LNs in 45 patients and two backscatter coefficient-based and two envelope statistics-based QUS parameters were computed and averaged for each frame. Different combinations of these four QUS parameters, along with the LN's short-axis and short-to-long axis ratio, were used to train random forest models to classify metastatic LNs.

Results: The average QUS parameters and radiomics features were significantly different between metastatic and benign LNs (p≤10^-4), except for effective scatterer diameter (p = 0.70). The best-performing random forest model, trained using a combination of QUS and radiomics features, identified metastatic LNs with an area under the receiver-operating characteristic curve of 0.91 and 67% specificity at 100% sensitivity.

Conclusion: These results demonstrate the potential of QUS imaging using a clinical scanner for identifying metastatic LNs in vivo to help clinicians perform a more selective LN biopsy or excision.