As deep learning has been widely used for computer aided-diagnosis, we wished to know whether attribution maps obtained using gradient back-propagation could correctly highlight the patterns of disease subtypes discovered by a deep learning classifier. As the correctness of attribution maps is difficult to evaluate directly on medical images, we used synthetic data mimicking the difference between brain MRI of controls and demented patients to design more reliable evaluation criteria of attribution maps. We demonstrated that attribution maps may mix the regions associated with different subtypes for small data sets while they could accurately characterize both subtypes using a large data set. We then proposed simple data augmentation techniques and showed that they could improve the coherence of the explanations for a small data set.