Baselines for Automatic Medical Image Reporting

Despite the high number of machine learning models presented in the last few years for automatically annotating medical images with deep learning models, clear baselines to compare methods upon are still missing. We present an initial set of experimentations of a standard encoder-decoder architecture with the Indiana University Chest X-ray dataset. The experiments include different convolutional architectures and decoding strategies for the recurrent decoder module. The results here presented could potentially benefit those tackling the same task in languages with fewer linguistic resources than those available in English.