Deep Learning for All

Biomedical research labs are producing image data at an ever-increasing rate. How can scientists make the most of this volume of data without spending excessive time manually counting and classifying elements in the images, such as individual cell types in diseased tissue? Although machine learning-assisted image analysis has existed for over a decade, a major problem is that its accessibility has been limited to research laboratories who work in the area of computer science. To address this limitation, a team of Freiburg scientists from the Department of Computer Science and the Signalling Research Excellence Clusters BIOSS and CIBSS have developed a deep learning plug-in for a widely used image analysis software. They have described their new software package in the current issue of “Nature Methods”.

Deep learning is a current machine learning method in which complex features, for example in images, are learned automatically on the basis of data. This makes it very easy to train the system for new tasks. The performance of a deep learning process increases with the amount of data provided for learning. This has led to the assumption that successful training of artificial neural networks requires large data sets – in the case of imaging data, thousands of annotated training images.  For many specialised applications, it is not feasible to generate this volume of data for training.

In 2015, Prof. Dr. Olaf Ronneberger and Prof. Dr. Thomas Brox of the Department of Computer Science developed U-Net, a deep learning network for analysis of biomedical images. The major advantages of U-Net are its ability to efficiently learn from relatively few training images and to achieve very high accuracy.

Now the team has developed a U-Net plug-in for a widely used biomedical image analysis software called ImageJ. The plug-in software is designed for 2D and 3D imaging data, and it can train on new data sets and extract features defined by the user. This avoids the need to develop a new software package for each research lab focusing on different biological samples and features within imaging data. This flexibility makes the plug-in broadly accessible to biomedical research labs. “The U-Net plug-in democratises deep learning-based image analysis”, says Brox. Running on a consumer-grade computer, the software needs only a few hours to adapt to new user-defined image analysis.

“More and more, the top journals are demanding rigorous quantification and statistics for imaging data.”, says collaborating researcher on the study Prof. Dr. Marco Prinz from the Institute of Neuropathology at the University Medical Center. “The U-Net plug-in can be easily incorporated into our existing workflows. It saves us time and allows us to extract valuable information from our images.”

Original Publication:
Falk, T., Mai, D., Bensch, R., et int., Prinz, M., Palme, K., Simons, M., Diester, I., Brox, T., Ronneberger, O. (2018): U-Net – Deep Learning for Cell Counting, Detection, and Morphometry. In Nature Methods. DOI: 10.1038/s41592-018-0261-2