AI Systems for Clinical Care in Radiology – Sophisticated Assistants

Keywords

technology assessment - medicolegal Issues - health policy and practice

Draft of instructions for use

Radiomics, which is the mathematical analysis of radiological data and the correlation of quantitative image features with clinical information, has arrived in diagnostic imaging. The new systems that make AI available for diagnostic imaging have the potential to gradually replace the current means of evaluating medical image documents, to cause a fundamental change in interpretation techniques, and thus to develop from a radiologist's assistant to a veritable competitor. Anyone who is ready to recognize the new AI systems and accept their help will certainly save a significant amount of time but might ultimately jeopardize at least a portion of their interpretation independence and responsibility. Although radiology has always been a truly technology-based discipline that is dependent on the further development of photographic methods, the introduction of new imaging equipment has always given users the opportunity to develop innovative examination methods, i.e., has provided an opportunity for creativity. However, the effect of dedicated AI will be profound and far-reaching if it becomes established like a multipurpose prosthesis in the technical heart of diagnostic radiology. Nonetheless, there should be fundamental readiness to recognize the new applications. Continued debate and resistance are not useful. Preparation for interacting with AI partners already shows that collaborating with AI can only be successful if a series of conditions are met on an ongoing basis. These conditions must be defined and met, and they require some significant concessions and sacrifices on the part of both providers and users.

It is desirable and necessary for the new programs to be tested in clinical studies. It would be useful for departments with varying degrees of independence to be set up at academic radiology facilities to test and further develop innovative AI systems. However, it is even more important for users to have a solid foundation of theoretical knowledge regarding a new AI system prior to working with it. The idea of working with AI without knowing its potential and limitations should be unthinkable and as preposterous as the thought of using a new imaging method without first studying its physical principles and mode of operation in detail. In other words: Over time, radiologists' level of familiarity with the theoretical foundations and the modules of relevant AI systems should reach their level of knowledge of radiological imaging techniques. Anyone who wants to fully embrace these new systems must have detailed knowledge of the relevant mathematics, physics, and philosophy and the mode of operation of the modules. Suitable learning materials can be identified on an individual basis with the spectrum ranging from independent learning to weekend seminars.

However, the manufacturers of AI applications also need to actively participate in this process. They must be required without exception to include documentation with the modules providing users with general information about the company and its employees and about contact people. They are also required to provide the sources of the clinical data used for comparison with the images. The participating clinics and institutions, the time period of the reports, the organization of collaboration, and the quality criteria used for the selection must be described. Moreover, the clinical data used for correlation must be updated on a continuous basis. Information with an unknown expiration date or electronic junk files must not be used for the creation of current associations and correlations of the information extracted from the images.

Once the necessary level of familiarity with the new technology in theory as well as practice has been achieved and the AI system is to be implemented in the clinical routine, three questions must first be answered. Question one: Is the AI module to be used regularly or only occasionally? This question is so fundamental that the answer must be defined on a permanent basis regardless of whether the decision is made by a single person or the members of an institution. Exceptions or solo actions should be avoided if possible. The expected long-term consequences of allowing co-interpretation play a decisive role in the decision. Question two is equally important: Will the AI module be consulted before or after the radiologist's own assessment of the image material? In other words: Should the AI finding be viewed first with the radiologist's own interpretation then being based on this information or should the radiologist's report be completed first and then compared to the AI finding? Regardless of the radiologist's fundamental responsibility for the diagnostic report, this decision can be made based on the level of specialist training as well as external factors like current workload. Question three is rhetorical but nevertheless important: How should the AI findings be documented in the diagnostic report? There can only be one answer: AI findings must always be documented completely without exception, i.e., word-for-word and with a full listing of all sources. Any deviation from this would severely damage the credibility of the AI system as well as the user. Moreover, the diagnostic report must include intrinsic deficiencies in the AI interpretation like bias in the selection of samples, differences in the technical and methodological examination standards, and deficiencies in the clinical datasets.

The use of AI modules can never be used as the reason or an excuse for even partially not performing one's own independent evaluation of the image material – not even in routine cases. Generative AI does not exist to replace the radiologist's own expertise. The developers of these revolutionary modules need to understand that they are expected to provide trial operation for a period of multiple months. However, when testing the performance and feasibility of the method any type of willingness to make concessions is unacceptable even after this trial period, not only because the AI portfolio does not include an assumption of responsibility for its own results.

Medical imaging post-processing has become so commonplace over the years that the tools used for this purpose are no longer even recognized as intelligent products but only as digital instruments. However, with the new versions, AI has added three additional areas to its portfolio and successfully combined these with the goal of increasing the quality of its conclusions regarding diagnosis and differential diagnosis: Multimodal imaging (e.g., PET-CT), multimodal parameterization (e.g., with the help of perfusion, diffusion, and spectroscopy), and multimodal evaluation of images (e.g., via fingerprinting). From this powerful position, AI considers itself qualified and authorized to correlate the various parameters with one another in such a superior manner that it is able not only to make diagnoses and differential diagnoses but also to identify histological subtypes (e.g., of hepatocellular carcinoma) and create molecular typing (e.g., of non-small cell lung cancer) with varying degrees of comprehensiveness. However, the transfer of this data to individual cases is associated with significant uncertainty. Therefore, the corresponding dedicated examinations must always be performed even in light of the relatively high prediction probability (80–90%) of diagnostic information generated by AI shown in studies because these values are only the sum of acquired data and are not reliable statements in the individual case.

The therapeutic recommendations and prognostic statements provided by innovative AI are associated with more and graver unknowns than a diagnostic bulletin. Conventional radiology does not have similar concerns regarding the results of image-based theranostics and states that AI's claim to exclusivity in this regard should be respected. Admittedly, this information is additional information outside the area of expertise that is not determined but only acquired by AI. For obvious reasons, the contribution to personalized precision medicine here is as questionable as AI's claim to exclusivity. It may be obvious, for example, to correlate the clinical benefit of radiological interventions involving the liver (e.g., TACE, TARE) with tumor features visible on images (e.g., vascularization, heterogeneity) and to include this information in the individual differential therapeutic decision. However, the general condition of the patient, possible comorbidities, individual tolerance of effective therapeutic substances, and the experience and competence of the radiologist performing the intervention and his team are significantly more important when evaluating feasibility and the chances of success. A conclusion about the indication for and chances of success of an individual measure as well as a decision about a sequential multimodal therapy made based on image data should be viewed with great skepticism since a series of additional factors like availability, application limitations, tolerance, and interactions play a significant role in the individual case. It is unrealistic to predict the further development of a symptom or even the course of a disease on the basis of image-based data acquired at a certain point in time no matter how diverse the data is. Therefore, the recent attempt to predict the mortality of a chronic disease like liver cirrhosis based on computer-assisted analysis of body composition can only be described as absurd. Such ridiculous claims by AI must be publicly criticized. Of course, it is important to avoid if possible damaging the trust some patients have begun to place in the results provided by specialized AI systems.

Artificial intelligence promises a transition from semiautomatic to fully automatic image interpretation. Although AI currently only primarily delivers supplementary information, powerful constructs including classifications and development profiles, i.e., an all-encompassing result, are supposed to soon be created. For example, in the case of coronary heart disease, the module provides information not only about the morphology but also the functional relevance of stenoses and occlusions and provides supplementary information about treatment alternatives and their prognosis. In the case of colorectal cancer, not only the primary tumor and the distant metastases are described and evaluated but also the type and nature of nodular lesions in the adjacent mesenteric adipose tissue. Moreover, AI provides the probability of success for surgical interventions and adjuvant treatment measures and the duration of progression-free survival or the probability of local recurrence. Prognostic information is most commonly used in oncology. However, similar information packets are now also offered for a series of other diseases with characteristic radiological findings, e.g., interstitial pneumopathies.

The pioneers of AI will continue to conquer new areas of application and develop increasingly comprehensive programs. This demands progressive thinking and the pressure of competition. However, the moment at which AI information is no longer only available to the representatives of academic disciplines but is also available to the general public, i.e., the point at which patients and their families can consult an AI opinion as a personalized second or third opinion, will be much more consequential for the communication and broad effect of results. Anyone with image material stored on a CD or DVD or computer and who is willing to pay a certain, perhaps sliding, fee can become an independent and responsible analyst of their own image material. Comparative use of multiple modules also seems possible. In certain circles and situations, multiple supplementary reports could even become routine. The more personal data beyond image material the patient is willing to provide to the programs, the higher the quality of the evaluation promises to be because the information can be compared to that of subcohorts of the collective used by the AI system. The foreseeable consequences for radiologists are serious. Contact with patients and their families will be at risk of becoming an unpleasant diplomatic mission and a mere caricature of the familiar confidential dialog and counseling that have long been cultivated and valued as established parts of the traditional collection and communication of findings.

Patients will be able to check the primary radiology report and the subsequent AI findings in detail and in the case of diagnosis-, treatment-, and course-relevant inconsistencies or contradictions, they will ask the responsible examiner for clarification. An AI consultation performed completely independently of the initial examiners will allow patients to perform quality control of the practical work of an entire medical discipline. Even patients who are critical of AI will still push radiologists for a detailed explanation of the opinions and conclusions. Formulating replies, particularly with regard to the reason for possible deviations from the AI opinion, can be time-consuming. Accompanying verbal explanations are similarly challenging. They should be neither apologetic nor combative. Some patients will recognize over the course of the conversation that their post-procedural research using an AI system that does not bear any responsibility for its findings has put the radiologist in an awkward position. However, physicians will be reminded each time of the risk posed by AI opinions and of how quickly they can be reduced from an independent expert to a second reviewer, data processor, and information broker for controversial opinions. A loss of trust must almost always be expected even if the AI system is ultimately incorrect. However, if the subsequent AI consultation uncovers an objectively false interpretation by the radiologist, legal consequences are also conceivable.

Conflict of Interest

The authors declare that they have no conflict of interest.

Publication History

Received: 15 November 2024

Accepted after revision: 03 February 2025

Article published online:
10 March 2025