11:54 a.m., June 29, 2022
“We should stop training radiologists now. It is quite obvious that in five years, ‘deep learning’ will do better than radiologists. » Delivered in 2016 by computer scientist Geoffrey Hinton, co-recipient of the prestigious 2018 Turing Prize for his work on neural networks, the award strongly marked the collective imagination. Fueled by industrial promises and relayed by certain doctors in the media, the theme has spread to the press and social media where we have seen the image of a coming revolution in medicine take hold due to the development of artificial intelligence (AI).
Yet, six years later, the statement is more false prophecy than visionary anticipation… The “great replacement” of doctors with AI has not happened. Conversely, the evolution of medical demography and the increase in the number of imaging examinations show more of a lack of radiologists than their technical unemployment.
Software is however available and put on the market, most often by start-ups such as AZmed, Gleamer, Incepto, Pixyl or Therapixel. Radiology thus ranks first among the fields of application of AI software validated by the American regulatory agency FDA (more than 200 software) or benefiting from European Conformity (CE) marking (around 200). Scientific publications on the subject are also flourishing, with more than 8,800 articles published between 2000 and 2018.
Read also – Do you think artificial intelligence will revolutionize your work?
Faced with promises backed by real progress, how then can we explain that the actual use of AI remains relatively timid (about 30% of American radiologists claiming to have already used it)? The key to reading these new uses must here take professional logic into account. Our recent survey, published as part of a dossier in the journal Networks dedicated to AI control, provides an initial analysis of these ongoing transformations.
Professional reappropriation and regulation
Following media discourse on its possible disappearance, the profession initially scrutinized the development of the technology with suspicion. In a study published in 2021, 38% of European practitioners questioned feared that AI was a threat to their activity.
Far from remaining passive, the profession and its representative bodies then mobilized to deal with the irruption of AI, anticipate possible uses and defend their territory.
Radiologists have engaged in normative work aimed at reclaiming these tools and promoting their use in accordance with their professional expectations.
This definition of “good practices” takes several forms: identification of tasks where AI would be beneficial, proposal of guidelines for software development (technical conditions and relevance of training databases, etc.), recommendations concerning their evaluation or their use, etc.
Above all, radiologists remind us that their job is not simply to read images.
Radiologists are not alone in their hesitation of using AI in practice with a common fear of job loss or insecurities. However, there are many important tasks outside of what AI provides that will continue to be needed for radiologists to perform. https://t.co/qVbOwGIS5n
— Therapixel (@therapixel) June 3, 2021
Even if AI would do as well as them on an interpretation task, which is currently debatable, it would still be very difficult to substitute software for a practitioner. Rather than being replaced, the professional sector therefore favors the narrative of the radiologist working with AI in order to improve patient care.
But the criticisms also relate to the very promises of the AI compared to its real performance. It appears that these tools need to be better evaluated: in 2021, more than 60% of software with CE marking had not been the subject of any scientific publication. In the absence of robust clinical trials, the question arises of their effectiveness in real conditions.
Read also – Artificial intelligence: how Europe must accelerate its development
These demands are not purely formal protests: they act in return on the industrialists of the sector. Indeed, several of them are now taking up the argument of the non-replacement of radiologists, or are trying to show their credentials by publishing articles evaluating the performance of their software.
This compliance with doctors’ expectations can of course be read as an adaptation of the sales pitch to the sales target. But it is also a question for these companies of succeeding in initiating and stabilizing collaborations… Indeed, the participation of radiologists is necessary for the evaluation of the software and the access to the image databases allowing to train the algorithms .
State regulation in retreat?
Health professions therefore find themselves on the front line to deal with the arrival in their practice of these new devices involving AI. Conversely, administrative and sectoral regulations are still under construction. They no longer concern only the case of radiology, but the more general one of the deployment of AI in medicine.
Admittedly, the regulatory framework is changing to take these new technologies into account – as evidenced by the recent bioethics law or the application of the European regulation relating to medical devices in May 2021. But, although the latter tightens the conditions for obtaining CE marking, it still appears to be less demanding than its American equivalent (FDA validation). And it pales in comparison to the standards set for the marketing of medicines.
Even though their price is one of the central issues to come, the fact that they are medical devices intended for professionals also raises questions with regard to their reimbursement procedure. While the Haute Autorité de Santé has evaluated applications intended for patients, for example for monitoring diabetes or in oncology, and has published an evaluation grid on this subject, professional software in radiology and elsewhere remains to this day less scrutinized by health authorities.
Their use intended for practitioners perhaps explains why they have passed under the radar of citizen debate. The issues related to health data and their protection were nevertheless the subject of lively exchanges, in particular concerning the development of the Health Data Hub (public interest grouping bringing together the CNAM, the CNRS, the High Authority for Health, etc., working on the “major strategic orientations relating to the National Health Data System (SNDS) set by the State and in particular the Ministry of Solidarity and Health”).
Read also – Artificial intelligence: why it is exaggerated to say that France has the best researchers
In fact, the questions abound:
- Issue of patient consent to the analysis of their examinations by AI,
- Critique of the commercialization of imagery data,
- Questions raised by the recent discovery of the ability of algorithms to assign an ethnoracial category to patients, etc.
If current developments in radiology invite us to remain cautious in the face of the promises of AI and not to give in to a form of technological solutionism, transformations are underway and call for political and ethical vigilance.
We would like to give thanks to the writer of this article for this amazing content
Why Artificial Intelligence Hasn’t Replaced Doctors Yet
Explore our social media accounts as well as other pages related to themhttps://www.ai-magazine.com/related-pages/