Abstract
A key component to the success of deep learning is the availability ofmassive amounts of training data. Building and annotating large datasets forsolving medical image classification problems is today a bottleneck for manyapplications. Recently, capsule networks were proposed to deal withshortcomings of Convolutional Neural Networks (ConvNets). In this work, wecompare the behavior of capsule networks against ConvNets under typicaldatasets constraints of medical image analysis, namely, small amounts ofannotated data and class-imbalance. We evaluate our experiments on MNIST,Fashion-MNIST and medical (histological and retina images) publicly availabledatasets. Our results suggest that capsule networks can be trained with lessamount of data for the same or better performance and are more robust to animbalanced class distribution, which makes our approach very promising for themedical imaging community.