News

« back to home « previous news item | next news item »

Although renowned artificial intelligence (AI) experts predicted a decade ago that radiology as a profession would become irrelevant, but medical profession is more threatened by staff shortages than AI nowadays, says Professor Dr. Péter Bogner, radiologist, who plays a key role in the implementation of the technology in Hungary. We discussed the role of AI in imaging diagnostics and its limitations with the Professor of the University of Pécs on the occasion of International Day of Radiology.

Written by Miklós Stemler

Nobel Prize Winner Geoffrey Hinton, known as the godfather of AI, stated in 2016 that due to the new technology, the training of radiologists would be unnecessary, as within five years, AI would perform imaging diagnostics at a higher level than experts. Five years have already passed, but instead of the presence of unemployed radiologists, there is a worldwide labour shortage in this field, but AI-based solutions are only partially able to deal with this situation.

– I show my students a meme consisting of three photos. On the first one Elon Musk predicts in 2015 that he would only produce self-driving cars by 2018. The second shows Geoffrey Hinton's statement in 2016 about the disappearance of radiologists, and there is a traffic jam on the third picture happening these days depicting radiologists leaving for work every morning due to the lack of self-driving cars – Péter Bogner gives a picture of the predictions made by famous people.

The former Head of the Department of Medical Imaging of the University of Pécs and a board member of the Hungarian Society of Radiologists could hardly be accused of being against the use of AI, because he played a key role in the AI-based reform of stroke diagnostics and treatment in Hungary. The method developed by the Professor and his colleagues in 2018 combined the advantages of teleradiology with diagnostics supported by AI.

– Stroke is among those diseases, in which AI can be used with great efficiency. In order to give the right treatment for the patient as soon as possible, a head CT is essential, which can determine the affected brain area and decide whether there is a possibility of removing the blood vessel blockage with a catheter or not. This is the point, where we can use the AI algorithm trained by a huge number of CT scans, which can determine this very fast and with a great reliability – he explains.

However, all of this requires access to technology, and especially in Hungary significant financial and human resource is needed. The solution developed by the Department of Neurosurgery, Department of Neurology, Department of Medical Imaging and the Pécs Diagnostic Center became an AI-based diagnostic network in the entire Southern Transdanubia Region, and then the internationally unique initiative was expanded to a national level in 2022 with the help of the National Institute of Mental Health, Neurology and Neurosurgery. This is a success story, but besides its strengths, it also points out the limitations of diagnostics supported by AI.

– AI performs particularly well regarding certain diseases, such as stroke, lung cancer, pulmonary embolism, bone fractures, and mammography, but it is not suitable for creating comprehensive medical findings. Regarding head MRIs, many aspects must be considered, but AI can only answer one question at a time. Different softwares needed to determine different diseases, but these softwares can be purchased at high prices. This fact puts a significant burden also on the healthcare systems of richer countries and seriously hinders the spread of these methods, and many experts say that it is simply not worth using AI on a daily basis – the Professor highlights the practical problems.

Another fact also presents the challenges, because Bayer, one of the world's largest healthcare companies, announced in September that it would end radiology platform that integrated various AI diagnostic solutions. Based on significant language models, ChatGPT and its peers already provide complete human knowledge, or at least its illusion, one possible solution could be a universal, “thinking” model trained by a huge number of medical scanning images, but this seems difficult to be implemented now.

– Systems based on language models need incredible resources, even though they „only” analyse letters, words and sentences, not much more complex radiological images. The currently available generative models require great computing power and energy, which needs are difficult to meet, so I don’t really see how this could be achieved. There may be distortions that still affect chatbots, as well, so the origin of the images really matters when models are trained. Different countries often have different characteristic regarding diseases, so a programme developed in China may not be useful in Europe, and vice versa. And I haven't even mentioned rare diseases, when collecting enough images for the training is a serious challenge – Péter Bogner lists the tasks to be solved. Although AI is able perform better than radiologists in certain tasks, it also tends to make incomprehensible and even life-threatening mistakes.

– While I was preparing for a presentation, I tested algorithms to recognise our dog on a photo, so these ranked the breed of the dog in percentage terms. The dog was a bichon, and although the breed was always among the first three results, but in many cases according to the algorithm this breed didn’t „win”. This experiment was undoubtedly useful, because I was able to get to know many dog breeds, but a CT or MRI scan is not that simple. The patient has a medical referral and a medical history, so radiologists must analyse the whole picture. A simple image recognition programme is not enough for this, and although there are attempts to programme a complex decision-making process to create a diagnosis, but the breakthrough has not come yet – adds Péter Bogner.

So, more radiologists will be needed in the foreseeable future, because the number of tests – primarily CT scans – has increased dramatically worldwide in recent years, as they are important during monitoring therapies and diagnosis, as well. AI might also play a role in fighting the serious staff shortage and overload that has arisen as a result.

– For example, we use AI-based speech recognition programme to prepare radiological findings, and large language models are already able to put free-word evaluation into a structured form. Solutions already exist, which can predict the failure of a critical component in CT and MR devices, so they can be replaced on time, and one of the and important developments in recent years is that algorithms are able to shorten the typically long preparation time of MR images. These provide great help in our everyday work, but they cannot and will not replace radiologists – sums up Péter Bogner.

Geoffrey Hinton, Nobel Prize winner in Physics in 2024, sees this in a similar way, as in May he spoke about AI solutions that would complement the work of radiologists making it more efficient and accurate, instead of letting them disappear. In case this vision comes true, both patients and doctors would undoubtedly benefit from it.

Photos:

Dávid Verébi, Szabolcs Csortos