"Used incorrectly, algorithms and AI can waste time"

Daniel Sage, scientific advisor at the EPFL Center for Imaging holds regular Image Analysis Breakfasts © 2021 Alain Herzog

Daniel Sage, scientific advisor at the EPFL Center for Imaging holds regular Image Analysis Breakfasts © 2021 Alain Herzog

Algorithms and machine learning let scientists compare and analyze images quickly and precisely, but using them often requires advanced skills in programming. To give EPFL scientists advice on employing these tools, Daniel Sage from the EPFL Center for Imaging holds regular Image Analysis Breakfasts, bringing together researchers from all disciplines.

Scientific imaging may be the only thing that fields as diverse as exoplanets, microcracks in materials, fruit flies and remote sensing all have in common. While the clues that scientists look for in the images can vary widely, algorithms and machine learning can help researchers across the board transform visual data into information. However, these methods come with a common set of challenges for all researchers. To help EPFL scientists overcome the stumbling blocks often associated with advanced image processing techniques, Daniel Sage – a scientific advisor at the EPFL Center for Imaging – along with a panel of experts gives twice-monthly Image Analysis Breakfasts. These events are designed to provide tips on the latest imaging methods for researchers at all EPFL schools and colleges.

Coaching sessions for new ideas on imaging

Recent advancements in image capture technology enable scientists to obtain vast amounts of information rapidly, opening up new avenues of research in just about every field, from materials science and astronomy to biology, physics and the environment. But these advancements have given rise to fresh problems with storing and quantifying the images and automating the analysis methods. Untrained scientists who use these methods incorrectly could find themselves with unreliable results or even see their research efforts fail. “Imaging algorithms and machine learning can be powerful tools, and are even essential for some kinds of research,” says Sage. “But using them for research applications requires a specific set of skills. Otherwise, scientists can get off on the wrong foot, and that could cost them months of time. They could believe they’ve detected something which actually isn’t there, for example, causing them to draw false conclusions from their experiments. And once the research is completed, it’s hard for them to see that there’s been an error.” With his Image Analysis Breakfasts, Sage aims to give scientists new ideas on how to use imaging more effectively, but not necessarily to solve their specific problems on the spot. The hope is that the advice from outside experts can help scientists adjust their approaches and carry out their research productively.

From the cosmos to the infinitely small, images have several things in common

Every research field has pretty much developed its own image analysis program, although the differences between them relate more to how they’re used and what’s taught rather than to any fundamental discrepancies. “For example, ImageJ is widely used in the life sciences, but there’s no reason why environmental scientists can’t use it too,” says Sage. One of the most universal challenges in image analysis is how to deal with the “noise” that invariably appears when images are taken. This is an issue that comes up in all types of scientific imaging, since the signals that researchers look for are usually weak relative to the noise around them. Sage explains: “The same methods are used to resolve this problem regardless of the kind of image. These include getting rid of the background or identifying points of reference, like the contours of an object.” Other such issues that cut across all disciplines frequently come up during his Breakfasts. One example is image reconstruction, which is often done from partial or indirect measurements. It can be hard for scientists to train deep learning programs if they don’t have a large enough image database.

The turning point when a machine beat a human

“Computer-based image analysis methods have been around for years, but the real turning point came in 2015 when a deep learning program won all the image-recognition competitions, beating out all the scientists,” says Sage. Since then, these complicated programs have become so popular that they’re sometimes used when they shouldn’t be, such as when conventional methods would work just as well and engender fewer risks.

To set up the Breakfasts, the organizers ask scientists to send in their image-related problems a few days beforehand. The organizers review the problems and select a handful to discuss at the event, and ask the scientists to present their images and describe the problem they face. Scientists can also choose to simply sit in on the events. For now the Breakfasts are being held entirely online. “One professor even said that his PhD students should also attend regularly,” says Sage, “so that they can learn about the many different ways to resolve imaging problems and open their minds to new methods.”