This study evaluates a robust parametric modeling approach for computer-aided detection (CAD) of vertebrae column metastases in whole-body MRI. Our method involves constructing a model based on geometric primitives from purely anatomical knowledge of organ shapes and rough variability limits. The basic intensity range of primary 'simple' objects in our models is derived from expert knowledge of image formation and appearance for certain tissue types. We formulated the classification problem as a multiple instance learning problem for which a novel algorithm is designed based on Fisher's linear discriminant analysis. Evaluation of metastases detection algorithm is done on a separate test set as well as on the training set via leave-one-patient-out approach.