Abstract. In this thesis work, the problem of applying active learning for a label efficient training of deep learning models is addressed. Firstly, in chapter one, the problem is introduced as well as the objectives and results of this thesis work. In chapter 2, a state of the art of active learning and deep learning models is presented with a particular emphasis on medical scenarios. In chapter three in active learning approach based on the expected gradient length is introduced for deep convolutional neural networks for applying in medical problems where data is scarce and train deep models could be unfeasible due to the lack of annotated samples. In chapter four an implemented framework for interactively training of deep learning models based on the previous discused algorithms is presented, where the active learning techniques improve the random selection strategy to classify between healthy eyes patches and patches that contain an early stage of diabetic retinopathy. Finally, in the last chapter, the conclusions of this thesis work are discussed as well as some promising lines of work for further research.