Quantitative-ultrasound detection of cancer in human lymph nodes based on support vector machines

Histological assessment of lymph nodes excised from cancer patients suffers from an unsatisfactory rate of false-negative determinations. We are evaluating high-frequency quantitative ultrasound (QUS) to detect metastatic regions in lymph nodes freshly excised from cancer patients. Three-dimensional (3D) RF data were acquired from 289 lymph nodes of 82 colorectal-, 15 gastric-, and 70 breast-cancer patients with a custom scanner using a 26-MHz, single-element transducer. Following data acquisition, individual nodes underwent step-sectioning at 50-µm to assure that no clinically significant cancer foci were missed. RF datasets were analyzed using 3D regions-of-interest that were processed to yield 13 QUS estimates including spectral-based and envelope-statistics-based parameters. QUS estimates are associated with tissue microstructure and are hypothesized to provide contrast between non-cancerous and cancerous regions. Leave-one-out classifications, ROC curves, and areas under the ROC (AUC) were used to compare the performance of support vector machines (SVMs) and step-wise linear discriminant analyses (LDA). Results showed that SVM performance (AUC = 0.87) was superior to LDA performance (AUC = 0.78). These results suggest that QUS methods may provide an effective tool to guide pathologists towards suspicious regions and also indicate that classification accuracy can be improved using sophisticated and robust classification tools. [Supported in part by NIH grant CA100183.]