Breast cancer is the most commonly diagnosed form of cancer in woman. While mammography is recognised as the standard method of diagnosing breast cancer, it has been shown that thermography based cancer diagnosis is able to identify cancer patients earlier than mammography. Clearly, correctly diagnosis cancer patients is of higher importance than correctly recognising benign ones. Therefore in this paper, we present a cost-sensitive classification approach to breast cancer analysis from thermograms. We first extract a series of statistical image feature from the thermal images and utilise these in the classification process. Importantly, we use a multiple classifier system where the ensemble, consisting of cost-sensitive tree classifiers, is tuned using a genetic algorithm in such a way that misclassifying malignant cases carries a higher penalty/cost than misclassifying benign cases, resulting in a classification with significantly improved sensitivity. Our experimental results on a large dataset of breast thermograms with known ground truth demonstrate that our proposed approach gives excellent classification performance.