To improve the accuracy of breast cancer diagnosis, many computer-aided diagnostic methods have been proposed. However, data imbalance and the limited size of medical datasets pose significant challenges. Although generative models have been introduced to address these issues, directly applying these models remains difficult due to the limited size of the original dataset. This is particularly challenging for mammograms because cancerous growths are uncommon and often very small, which makes creating these images harder. In this study, we use a denoising diffusion probabilistic model to generate the breast region in mammograms. Then, by utilizing the loss of local high-frequency information and comparing it with the original image, we distinguish between the breast and glandular tissue. Finally, we combine linear interpolation and prior knowledge to generate synthetic images. Our approach significantly enhances performance on the INbreast and CBIS-DDSM datasets.