Region-of-interest (ROI) segmentation is an important critical step and challenging task in the evolution of computer-aided detection (CAD) system for breast cancer. The discovery of breast cancer in early stages can save many women lives. However, most of the early detection systems are costly in terms of complexity, price and processing time; that make it unsuited for developing countries. The digital mammography is proven to be one of the most important diagnostic techniques for breast cancer tumors. Therefore, this work proposes a CAD system for breast cancer detection from digital mammography based on Gaussian Mixture Model (GMM) followed by Support Vector Machine (SVM). The best contribution of our proposed system is the usage of GMM for the first time in the literature for mammogram images segmentation into ROI areas. Besides, the discrimination between the three classes of tissues as normal, benign or malignant, is used without previous knowledge of mammogram images' type. Moreover, the proposed system is fully automated in all of its stages with reduced computation compared with recent used methods. Hence, it offers a suitable early detection system to our country regarding moneywise, timewise, and reduced complexity. A non-linear multi-class SVM is used for classifying the ROI into three classes: normal, benign or malignant tissue. The experiments show overall average classification accuracy of 90% for detecting normal, malignant or benign on randomly chosen 90 cases from the benchmark mini-MIAS dataset. On the other hand, the proposed method achieves 92.5% accuracy when classifying the benign from malignant cases.