Learning to improve area-under-FROC for imbalanced medical data classification using an ensemble method