Noninvasive in vivo detection of prognostic indicators for high-risk uveal melanoma: ultrasound parameter imaging.

PURPOSE Primary malignant melanoma of the choroid and ciliary body has traditionally been treated without histologic staging, using purely clinical indicators. The presence of extravascular matrix patterns (EMP) in histologic sections of uveal melanoma has been shown to be an independent indicator of metastatic risk. These patterns are of a dimension and physical composition that are likely to be detected with ultrasound backscatter analysis. Our aim was to determine whether ultrasound parameter imaging could detect the presence of EMP at a diagnostically significant level for treatment staging and for planning investigational studies of therapeutic modalities. DESIGN Prospective, masked ultrasound-pathologic correlative study. PARTICIPANTS One hundred seventeen patients diagnosed with previously untreated choroidal melanoma were scanned within 2 weeks before enucleation. METHODS Tumors were evaluated histologically and divided into high-risk and low-risk groups on the basis of the presence of 2% or more histologic cross-sectional area composed of EMP patterns. Digital ultrasound data were processed to generate parameter images representing the size and concentration of ultrasound scatterers. Histologic and ultrasound images and data were correlated, and linear and nonlinear statistical methods were used to create multivariate models for noninvasive differentiation of high-risk and low-risk tumors. MAIN OUTCOME MEASURES Presence or absence of high-risk EMP and associated ultrasound parameter classification models. RESULTS Of the 117 tumors, 69 were classified as low risk, and 48 were classified as high-risk with histologic analysis. A classification that used ultrasound parameter image features with linear discriminant analysis could correctly identify 79.5% of cases retrospectively and 75.2% of cases by use of cross-validation, an estimate of prospective classification ability. By use of a more powerful classification technique (support vector machine), 93.1% of cases were correctly classified retrospectively. With a cross-validation procedure, 80.10% of cases were correctly classified. CONCLUSIONS Ultrasound can be used noninvasively to classify tumors into high-risk and low-risk groups by detecting the presence of EMP patterns. By the use of previous studies that compared the histologic presence of EMP patterns with patient survival, estimates of hazard rates associated with ultrasound risk groups can be made. The noninvasive ultrasound classification is potentially useful as a prognostic variable and as a tool for stratification of patient populations for tumor treatment evaluation.