In the real world, especially in the medical imaging context, data scarcity and limited labeled data are recurrent and frequent problems. This is very often a bottleneck to high-performance of recent Deep Learning approaches that are very data-hungry. In this work, we show that active learning could be very effective in data scarcity situations, where obtaining labeled data is expensive. We compare several acquisition functions (AF) such as BALD, MeanSTD, and MaxEntropy on the ISIC 2016 Melanoma detection dataset, explore the impact of selecting either the most or least uncertain samples, and leverage the effect of acquired pool sizes on the performance of the model. Our results on the Melanoma detection test set, demonstrate that uncertainty is useful to the Melanoma detection task and that it is more beneficial to select the most uncertain pool samples. These results suggest that active learning could be very useful for medical imaging tasks (in particular) and more generally in low-resource settings.