The need for diagnostic accuracy in breast cancer screening has motivated the use of computational schemes known as computer-aided diagnosis (CADe/Dx). However standardization of the evaluation criteria regariding these schemes is still a challenge, since they depend on the access to large and diversified image databases representing both benign and malignant cases. With this feature in view this work aimed at the development of a structured breast phantom able to generate many images patterns achieved not only by the variation of size, shape, contrast but mainly the distribution of simulated lesions. To assure the realism of images, phantom non-uniformity was an essential requirement in order to mimic breast structures, differing from the phantoms commonly developed for quality control purposes. Thereforethis phantom was made in layers composed by PVC film submerged in paraffin gel with a non-uniform distribution. Such distribution allowed simulating more or less dense regions, according to the material concentration. The percent density estimative of the simulated breast was determined with the LIBRA® software, resulting in the four categories of BIRADS® density classification. Nodules were simulated using two 3D printed models, one for circumscribed and another for spiculated lesions. For simulating microcalcifications, granulated hydroxyapatite was used distributed in four clusters. Phantom characteristics were studied comparing attenuation coefficients of the materials used and breast tissues. The texture and noise of the phantom images were also compared in relation to the clinical ones. Results showed that the developed phantom allowed generating an image pattern similar to that obtained in real mammograms and suitable for supplying an image database to be used as a ground truth for CAD and CADx schemes validation/evaluation.