In this paper we propose a computerized breast cancer detection and breast masses classification system utilizing mammograms. The motivation of the proposed method is to detect breast cancer tumors in early stages with more accuracy and less negative false cases. Our proposed method utilizes clustering of different features by segmenting the breast mammogram and then extracts deep features using the presented Convolution Neural Network (CNN). The extracted features are then combined with subjective features such as shape, texture and density. The combined features are then utilized by the Extreme Learning Machine Clustering (ELMC) algorithm to combine segments together to identify the breast mass Region of Interest (ROI). We present a detection method utilizing the ELMC clustering technique. Building a multi-feature set, the ELMC classifier is utilized to perform classification of normal, benign and cancer breast masses. Feature fusion is performed on the extracted shape, texture and density features forming a fusion feature set. In the automated detection phase, we utilize the fusion feature sets for classification. Extensive experimentation has been carried out to validate the ability of our proposed method. We utilized a dataset of 600 female mammograms. The experiments measure the accuracy of our proposed detection and classification method. The CNN coupled with the Extreme Learning Machine Clustering algorithm achieves the highest accuracy, sensitivity, specificity and ROC measures when combined with a multi-feature set. The model achieves 98.53% cancer detection accuracy, 95.6% benign detection accuracy and 95% for normal cases.