Active Learning Strategies and Convolutional Neural Networks forMammogram Classification

Deep learning has been used successfully in a variety of applications due to the large data availability and the growth in computing power. However, some domains present a shortage of both samples and labels, for instance, the medical area. In this work, we propose machine learning approaches that include traditional supervised classifiers and active learning methods for the breast lesion domain, in order to aid breast cancer diagnosis. We propose the introduction of active learning strategies in this process, to sort out the most informative samples in the dataset. The active learning process reduces the burden of the dataset annotation, while also improving the robustness of our models. Hence, we achieved considerable gains with fewer labeled training images, minimizing the specialist’s annotation effort. The validation of our proposed methodology is done on a public breast lesion-related dataset and our results show considerable accuracy gains over the traditional supervised learning approach and reductions of up to \(68\%\) in the labeled training sets.