CHARACTERIZATION OF SONOGRAPHICALLY INDETERMINATE OVARIAN-TUMORS WITHMR-IMAGING - A LOGISTIC-REGRESSION ANALYSIS

Purpose: The goal of this study was to maximize the discrimination bet ween benign and malignant masses in patients with sonographically inde terminate ovarian lesions by means of unenhanced and contrast-enhanced MR imaging, and to develop a computer-assisted diagnosis system. Mate rial and Methods: Findings in precontrast and Gd-DTPA contrast-enhance d MR images of 104 patients with 115 sonogtaphically indeterminate ova rian masses were analyzed, and the results were correlated with histop athological findings. Of 115 lesions, 65 were benign (23 cystadenomas, 13 complex cysts, 11 teratomas, 6 fibrothecomas, 12 others) and 50 we re malignant (32 ovarian carcinomas, 7 metastatic tumors of Che ovary. 4 carcinomas of the fallopian tubes, 7 others). A logistic regression analysis was performed to discriminate between benign and malignant l esions, and a model of a computer-assisted diagnosis was developed. Th is model was prospectively tested in 75 cases of ovarian tumors found at other institutions. Results: From the univariate analysis, the foll owing parameters were selected as significant for predicting malignanc y (p less than or equal to 0.05): a solid or cystic mass with a large solid component or wall thickness greater than 3 mm; complex internal architecture; ascites; and bilaterality. Based on these parameters, a model of a computer-assisted diagnosis system was developed with the l ogistic regression analysis. To distinguish benign from malignant lesi ons, the maximum cut-off point was obtained between 0.47 and 0.51. in a prospective application of this model, 87% of the lesions were accur ately identified as benign or malignant. Conclusion: Benign and malign ant ovarian lesions can be distinguished in most sonographically indet erminate lesions by means of parameters obtained from contrast-enhance d MR imaging.