Unsupervised learning for concept detection in medical images: a comparative analysis

TL;DR

This paper evaluates various unsupervised learning methods for medical image concept detection, showing deep learning approaches outperform traditional methods, with GANs being promising but challenging on diverse datasets.

Contribution

It provides a comparative analysis of six unsupervised feature learning methods for biomedical images, highlighting the effectiveness of modern deep learning techniques.

Findings

Deep learning methods outperform traditional visual words.

GANs yield good results but are data-sensitive.

Modern approaches offer more powerful representations.

Abstract

As digital medical imaging becomes more prevalent and archives increase in size, representation learning exposes an interesting opportunity for enhanced medical decision support systems. On the other hand, medical imaging data is often scarce and short on annotations. In this paper, we present an assessment of unsupervised feature learning approaches for images in the biomedical literature, which can be applied to automatic biomedical concept detection. Six unsupervised representation learning methods were built, including traditional bags of visual words, autoencoders, and generative adversarial networks. Each model was trained, and their respective feature space evaluated using images from the ImageCLEF 2017 concept detection task. We conclude that it is possible to obtain more powerful representations with modern deep learning approaches, in contrast with previously popular computer vision methods. Although generative adversarial networks can provide good results, they are harder to succeed in highly varied data sets. The possibility of semi-supervised learning, as well as their use in medical information retrieval problems, are the next steps to be strongly considered.