First place in an international artificial intelligence competition, a distinction that has been systematically achieved for the last five years, was once again won this year by a group of students and researchers from the Department of Informatics of the Athens University of Economics and Business.

It’s about the “ImageCLEFmed Caption” competitionin which systems of research groups take part, which, using artificial intelligence, attempt to automatically link medical images from X-rays, ultrasounds, CT and MRI scans, etc., with tags of medical concepts for the better management of large volumes of data, but also to produce medical texts diagnoses.

The Natural Language Processing research group of the Athens University of Economics and Business took first place in this year’s competition in the automatic assignment of medical labels (Concept Detection) and third place in the automatic description of medical images (Caption Prediction). The team is made up of Professor of Artificial Intelligence at the Department of Informatics of the Athens University of Economics and Business and collaborating researcher of the “Archimedes” Unit of the “Athena” Research Center, Ion Androutsopoulos, the researcher of the Department of Informatics of the Athens University of Economics and Business and the University of Stockholm, Yannis Pavlopoulos , the PhD candidate of the Department of Informatics of the OPA and doctoral scholar of the “Archimedes” Unit, Georgios Moschovi, the doctoral candidate of the Department of Informatics of the OPA, Phoivos Charalambakos, and the postgraduate student of the Master’s Program “Computer Science” of the Department of Informatics of the OPA , Panagiotis Kaliosis.

The OPA team entered the competition with systems based on deep learning (deep learning), a form of machine learning that uses deep artificial neural networks to encode and categorize images, as well as generate text, a technology also used by ChatGPT . The team developed new forms of neural networks specifically for the competition, but also modified existing forms by combining different machine learning methods.

The aim of the competition, in which the OPA team was distinguished, is for the participating systems to be able to support doctors in searching for medical images by means of keywords. Also, support them by automatically producing medical opinion drafts from medical exam images, which are then reviewed and improved by doctors.

“We are not trying to replace the human doctor, we are trying to help him. For example it may be faster and easier for a doctor to see an initial form of the diagnosis and correct or approve it himself than to write it from scratch. Or maybe if a doctor is not very experienced, if he sees a first form of the diagnosis from a system trained on the diagnoses of many other experienced doctors, he can get some ideas about medical findings that may have escaped him”, explains the APE-MPE Professor of Artificial Intelligence of OPA, Ion Androutsopoulos.

But how likely are such sophisticated artificial intelligence systems to replace doctors in the future, we ask the researchers. “Always, especially in medical applications, there must be the human doctor who will make the final decision and make the final diagnosis because he is the one who understands best and has the final responsibility. After all, a doctor does much more than the specific systems. He doesn’t just look at the pictures, he takes the patient’s history, studies previous diagnoses and readings from laboratory tests. He has much more information and experience. Medicine is an entire science”, Ion Androutsopoulos and Phoivos Charalambakos answer to APE-MPE.

The doctoral fellow, Giorgos Moschovis, also clarifies that “in the future, as the systems and the quality of the data improve, that is, if, for example, better medical machines are made that give us better medical images, we will also have better representations of the images, so possibly we can help the doctor more. Again, however, we will not replace him.”

It is not the first time that the team has distinguished itself in this competition, which is organized by the international non-profit association CLEF, a partnership of universities with the main participants being the Technical University of Bucharest in Romania, the University of Applied Sciences in Switzerland and the Technical University of Dormunk. The first place in the automatic assignment of medical labels has been consistently occupied by the Natural Language Processing group of the OPA Informatics Department since 2019, when it started participating, while in the ranking for the automatic description of medical images it has been consistently in the top three for the last few years.

The key to the team’s success, according to the graduate student of the OPA Department of Informatics, Panagiotis Kaliosi, is that “we rely on machine learning methods, especially deep learning, which in recent years have been at the forefront of artificial intelligence, and we enrich them with newer developments, such as those used in ChatGPT, but without getting carried away by the reputation that certain methods have acquired.”

“We have experience with more machine learning methods, even more traditional ones that don’t use neural networks, and we try and synthesize them all together. So we insist on exploring many methods and not just this one that may be in the headlines at this time. And this rewards us”, adds characteristically Mr. Androutsopoulos.

The group’s goal is to open future cooperation channels with companies that produce medical diagnostic equipment, so that the technology developed can be integrated into their systems.

It is noted that the system developed by the Natural Language Processing team was presented last week at the annual conference on information retrieval systems organized by the non-profit association CLEF in Thessaloniki.