查看文章

Representation learning is a field that has rapidly evolved during the last decade, with much of this progress being driven by the latest breakthroughs in deep learning. Digital medical imaging is a particularly interesting application since representation learning may enable better medical decision support systems. ImageCLEFcaption focuses on automatic information extraction from biomedical images. This paper describes two representation learning approaches for the concept detection sub-task. The first approach consists of k-means clustering to create bags of words from SIFT descriptors. The second approach is based on a custom deep denoising convolutional autoencoder. A set of perceptron classifiers were trained and evaluated for each representation type. Test results showed a mean F1 score of 0.0488 and 0.0414 for the best run using bags of words and the autoencoder, respectively.