tEPIS project facilitates Grand Challenge in digital pathology

Ever wondered why digital radiology is now routine, whereas digital pathology (the acquisition, storage and sharing of histology-slide images) is still in its infancy?

Even with today's high-speed broadband connections, sharing terabyte files over the internet remains a problem. It's why the widespread sharing of digital pathology images (high-resolution images of tissue sections) is nowhere near as advanced as the sharing of digital radiology images. Typical radiology images are megabyte files. A typical high-resolution pathology slide image is 2 to 3 Gbytes in size. Aggressive image compression is not the answer, because it has the potential to introduce image artefacts that impair diagnostic quality.

The tEPIS (TraIT Enhanced Pathology Image Sharing) platform, developed as a result of collaboration between the Life Science Health FES 2009 program and CTMM TraIT, offers a solution for research groups that want to share whole-slide pathology images among their researchers. Integrated into the TraIT tool set, it allows fully annotated pathology images to be appended to OpenClinica clinical data files. As Jeroen van der Laak, Assistant Professor at the Department of Pathology at Radboud University Medical Center, explains, the tEPIS platform overcomes the image size problem using techniques borrowed from Google Earth.

"In Google Earth you get a low resolution image of the entire world and you only get higher resolution data when you zoom in on the region you're interested in. The tEPIS digital pathology platform does a similar thing. You get a low resolution image of the entire slide, sufficient for a pathologist to identify areas of interest, and the system then delivers those regions in high resolution. This reduces the download requirement to the point where anyone capable of using a web browser can work with the images."

Although tEPIS is designed as a research tool, the ultimate aim of digital pathology is to increase the efficiency of a pathologist's routine workload.

"For something like a sentinel lymph node for a breast cancer patient, the pathologist has to review ten to twenty pathology slides to see if any of them contain tumour cells," says Jeroen. "We identified this part of the pathology workflow as a strong opportunity for computer-aided diagnosis, because on the one hand it's tedious work, yet on the other hand the consequences for the patient are large. Even simply using computer-aided diagnosis to re-order the slides so that the pathologist looks at the slides most likely to contain cancer cells first could save a lot of time."

However, recognizing potentially cancerous cells in whole-slide digital pathology images requires the use of very sophisticated pattern recognition algorithms. It's something that Jeroen and his team at Radboud UMC, together with colleagues at Utrecht UMC and Technical University Eindhoven, are working on but in parallel they are also throwing down the gauntlet to the algorithm development community by organizing the CAMELYON16 Grand Challenge via the website.

"We have already found that pattern recognition in whole-slide images is a lot more difficult than doing it on manually extracted regions of interest, because whole-slide images contain a lot more noise. But it's also challenging because of the sheer amount of data," explains Jeroen. "In the CAMELYON16 Grand Challenge, research groups will be able to use the tEPIS platform to access the sentinel lymph node histopathology slides of around 300 known-outcome breast cancer patients, which they can use as the seed data for deep-learning techniques such as neural networks to develop new pattern recognition algorithms. We will then give them another 150 outcome-blind cases on which they can run their algorithms and send us back their results, which we'll compare to the ground truth. The winners will be announced at the 2016 International Symposium on Biomedical Imaging (ISBI 2016) in Prague in April, where we will invite the top performing research groups to present their work. The assessment is completely objective, because right from the outset we let them know exactly how we will analyze the outcome of their algorithms."

The important thing is that Jeroen and his team don't just want the biomedical research community to participate.

"Most of today's deep-learning techniques are considered problem agnostic, which means they can be applied to many different fields of research," he says."The advantage of accessing the images via the tEPIS platform is that it makes it possible for people who know very little about pathology images to take part in CAMELYON16."

More information:

The FES2009  Life Science Health funded tEPIS project at



Originally published in the CTMM Newsletter #22, December 2015