Radiologists worldwide use mammography as a reliable tool for breast cancer screening. However, mammography assessment is challenging even for well-trained radiologists, leading to a pressing need for Computer Aided Diagnosis (CAD) systems. In this work, a novel technique for the detection and classification of breast Micro-Calcifications (MCs), which are diagnostically significant but difficult to detect findings, is presented. The proposed method is based on the subtraction of temporally sequential mammogram pairs, after pre-processing and image registration, followed by machine-learning. The classification was performed using several features extracted from the subtracted mammograms and selected during training to optimize the accuracy of the results. Six classifiers were tested in a leave-one-patient-out, 4, 5 and 10 fold cross-validation process. This technique was evaluated on a unique dataset, consisting of temporal sequences of mammograms from 80 patients taken between 1 to 6 years apart. The resulting 320 mammograms were reviewed by 2 radiologists who precisely marked each MC location. The accuracy of classifying MCs as benign or suspicious improved from 91.42% without temporal subtraction and an Ensemble of Decision Trees (EDT), to 99.55% with the use of sequential mammograms and Support Vector Machines (SVMs) with leave-one-patient-out validation. The improvement was statistically significant (p-value <; 0.005). These results verify the accuracy and the effectiveness of the proposed technique should to be further evaluated on a larger dataset.