A Stacked Deep Autoencoder Model for Biomedical Figure Classification

Figures in biomedical research papers are a significant source of vital information surrounding the experimental settings and results. One of the most important steps towards mining biomedical figures and reducing the vast search space is figure classification. A challenge that presents itself when classifying biomedical figures is the huge imbalance between the devised classes. This paper presents a novel method to overcome this imbalance in the available biomedical figures, by improving the prediction accuracy on the small classes, hence improving the overall accuracy. A stacked deep autoencoder model is trained to reconstruct the figures, then trimmed, by removing its decoder, and used to automatically extract visual features. The remaining encoder is fine-tuned to classify each class against the rest, with each encoder generating features that feed into a one-vs-all support vector machine (SVM). The ensemble of SVMs determines the class of the figure by taking the highest probability. The obtained results show that the proposed method improves not only the accuracy on the smaller classes in the taxonomy, but also the overall accuracy.