Multi-Stakeholder AI Ethics in Radiology

AI applications in clinical radiology

In clinical medicine, labor is divided meticulously and  efficiently, with sequential and parallel task allocation between and within groups of specialists, each of which has clearly defined responsibilities. The requirements for the use of AI support systems in this context relate, on the one hand, to the technical ability of the systems to collect and evaluate a large amount of information, draw conclusions from it, and communicate these. Given the nonlinear progression of diseases, such AI support systems  would have to harness new information to lead to a reassessment of old information. 

Such universal AI systems do not currently exist. AI solutions are only used for specific aspects of medical decision-making processes, such as the detection of radiological anomalies [1, 2, 3]. On the other hand, the use of AI systems in medicine is subject to particular ethical requirements regarding the transparency, explainability, and traceability of their output, all of which are essential for the interaction between medical professionals and AI as well as for the trust  between medical professionals and patients [4, 5, 6]. 

However, observing ethical criteria in AI development does not automatically lead to improvements in medical care. In addition, ethical considerations do not always take patients and medical staff equally into account, creating a tension between physician assistance and physician competition. As early as 2018, radiologists expressed concerns about the future decision-making authority of AI [7].

As part of the HUMAINE joint project on ergonomics, we investigated the added value of an integrative AI development approach for users and patients, i.e., for a human-centered and thus ethical introduction of AI in radiology workplaces. An exemplary use case was the AI-assisted detection of structural abnormalities (lesions) causing epilepsy as well as the associated therapy options in magnetic resonance imaging (MRI) of epilepsy patients (in short: AI-assisted Epi-MRI). This study is located at the interface between neuroradiology and neurology/epileptology and touches on numerous ethical aspects of medical AI/human co-creation.

Framework conditions for ethical AI in radiology—current status and shortcomings

The detection of lesions that cause epilepsy raises numerous ethical considerations, which can be illustrated using the influential four ethical principles according to Beauchamp and Childress [8] (beneficence, non-maleficence, autonomy, and justice). The principle of beneficence is central to lesion detection, as it has high therapeutic relevance for patients [9, 10]. However, detection rates by non-specialist radiologists have been suboptimal to date due to a lack of knowledge [11] and interdisciplinary interface issues (lack of a neurological focus hypothesis) [12]. 

The principle of non-maleficence is challenging, especially for radiologists with little experience, as the technical requirements for radiologists are high. The spectrum of lesions to be recognized is wide, and the lesions are sometimes very subtle. Knowledge of pathophysiology and the interdisciplinary clinical consequences of the findings is required [13, 14].

The principle of autonomy becomes relevant when AI has to be integrated into the established intra-radiological and interdisciplinary workflow [15–18], where it stands in direct competition with physicians [7]. The immediacy and wide availability of AI-generated findings, compared to significantly slower medical diagnosis and communication, contribute to this challenge for autonomy. 

The principle of fairness is affected, among other things, because the time required for a thorough evaluation is hardly compatible with the clinical workload [19]. AI support can therefore offer more capacity for fair access to sound diagnostics.

However, the integration of AI applications in this field adds further ethical requirements, as AI systems themselves must meet certain ethical standards. In fact, ethical aspects are an overarching principle goal when integrating them into work processes and patient care. The FUTURE-AI consensus criteria [6] (Fig. 1) were developed by scientists, physicians, ethicists, social scientists, and legal experts from numerous countries and outline requirement profiles for six core principles.

Based on these core ethical principles of the FUTURE-AI consensus criteria, key challenges for clinical radiology can be identified:

• Fairness: AI-assisted lesion detection requires a large quantity of high-quality data. This is often difficult to achieve in any given context, as high standards of anonymization and data protection are required.
• Universality: To date, no algorithm is capable of detecting all types of lesions equally well.
• Traceability: Quality control of AI applications has so far focused on the technical side [20, 21]. However, the influence of human decisions on the quality of AI-human co-creation must also be measured against ground truth [22]. Initial methodological groundwork has been laid for this [23], but no comprehensive quality measures are available yet.
• User-friendliness: Radiologists’ hopes (AI as assistant) and fears (AI as competition) associated with integrating AI in radiology workplaces [7, 24, 25] have not yet been investigated for the specific use case.
• Robustness: AI applications in radiology must be tolerant of poor or fluctuating conditions, which may include technical image artifacts, anatomical variants, age-related changes, and insufficient or incorrect preliminary medical information. The quality of the findings, e.g., anomaly detection and derived clinical recommendations, must not suffer as a result [16].
• Explainability: Diagnosis is made by clinical users. AI can provide support if it delivers comprehensible outputs that enable radiologists to develop well-founded diagnoses and treatment recommendations.

Research approach

In response to the challenges mentioned above, we have developed modules that support the FUTURE AI core principles and aim to help realize a comprehensive, prototypical model of human-centered AI-assisted radiology, including workflow design. The aim of this work is not to offer a comprehensive solution for all use cases but to stimulate future developments. Figure 2 shows the five modules created by an interdisciplinary board comprising ergonomists, occupational sociologists, physicians, computer scientists, statisticians, and providers of radiological communication systems (see list of authors).

Module 1: Optimization of the mathematical foundations of anomaly detection in 3D MRIs

Focus on FUTURE-AI core principles: fairness, universality

With the Spatially Context Aware Transformer (SpyCAT), we have developed an approach for detecting small, local anomalies [26] that is explicitly derived from formulated modeling assumptions and can explicitly characterize the normality of MRIs and identify anomalies as deviations from it. SpyCAT was trained using publicly available 3D data and 900 MRIs from the Interdisciplinary Epileptology-Neuroradiology Outpatient Clinic of Ruhr Epileptology. To comply with GDPR requirements, a new defacing method was developed and validated [27]. 

This approach not only yields better predictive accuracy compared to other methods but also widens  applicability in line with Future AI principles. To increase fairness, the annotated training will incorporate the international MELD database [28], while the short processing time of less than 30 seconds per MRI sequence will enhance user-friendliness. SpyCAT also follows other Future AI principles.

Module 2: Expansion of AI annotation to include diagnostic and therapeutic recommendations

Focus on FUTURE AI core principles: universality, robustness

The output of  clinical recommendations from the findings requires that these have already been entered at the time of AI training. In our use case, for example, we linked the categories of lesion etiology according to SNOMED classification with their potential epileptogenicity. One challenge is that the same lesion etiologies can have different clinical consequences; for example, the size and location of a specific lesion also determine its operability. 

To validly transfer such extended annotations onto practical use, building on the approach in Module 1 requires an additional layer of machine learning, including additional data. In order to obtain this data in the future, current developments in  multicenter consortium structures should be further pursued. This module particularly concerns the principles of universality, as therapeutic recommendations increase applicability, and robustness, as any influence of unfavorable external circumstances must be ruled out, especially in the case of clinical recommendations.

Module 3: Development of a metric for evaluating the quality of AI-human co-creation of findings

FUTURE-AI core principles in focus: traceability, explainability

Our quality control system for AI/human diagnosis co-creation [29] extends the  basic accuracy formula Quality Q = (ground truth – error) / ground truth with several optional components. The ground truth can be defined variably (entire examinations, individual lesion types, etc.), and differentially weighted diagnostic axes, including clinical recommendations derived from anomalies, can be incorporated into the end result. False negative and false positive findings can be evaluated as quality-reducing depending on the application. The effect of MRI training and critical or non-critical attitudes towards false positive and negative results can also be measured. 

A pilot study showed significantly better results with AI/human co-diagnosis than with human diagnosis alone. The quality control procedure takes into account the principle of traceability through verifiable metrics, whereas explainability is achieved through application-specific weighting and transparent metrics.

Module 4: Ergonomic analysis of positive and negative attitudes of radiologists toward AI assistance 

Focus on FUTURE AI core principles: user-friendliness, universality, robustness  

A standardized questionnaire (perception of human-AI interaction and cognitive, affective, and behavioral acceptance) in conjunction with expert interviews (interpretation using grounded theory [30]) was used to assess radiologists’ attitudes toward AI assistance. Radiologists recognize the (objectively given) diagnostic benefits and time savings of AI applications, but they are critical of the unclear reliability of AI solutions and their own possible uncertainties when making diagnoses using AI. The use of AI does not reduce their own contribution to diagnosis, provided that both AI and radiologists make interactive diagnoses (obligatory co-creation of findings). 

However, the majority view the “dictation of findings by AI” critically. While it was positively noted that AI affords more time for individual interactions with patients, negative attitudes were driven by the expectation that using AI could further increase workloads and thus induce more stress. Freedom in organizing their own work and relief from complex diagnostic decisions are among radiologists’ top priorities. 

Overall, radiologists do not see their role fundamentally challenged by AI assistance. These findings are closely linked to the FUTURE-AI principles of user-friendliness (in particular of human-centered AI, user interaction, and efficiency), universality (especially adaptability and applicability), and robustness (reducing fears about the unreliability of AI assistance).

Module 5: Radiological expectations for AI integration into typical workflows

FUTURE-AI core principles in focus: explainability, user-friendliness  

According to the results of the interview study (see above), radiologists’ expectations regarding the integration of AI into their own workflow depend on their workplace, role or position at the institute/clinic, and other personal criteria. Therefore, individual workplace considerations must always be taken into account. 

However, there are also more general criteria for user acceptance. These include, for example, an accompanied and interactive AI introduction at the workplace, expectations towards AI analysis speed, the ability to enter clinical information into the AI, the provision of derived recommendations by the AI, the AI’s indication of the probability of the correctness of results, the presentation of alternative findings, and the AI’s assistance in formulating findings.

The FUTURE-AI principle of explainability is crucial, as both the provision of recommendations and the desire to input clinical information require a transparent human-AI interface. In addition, the principle of user-friendliness is addressed in the form of interactive elements and efficient usability.

Prototypical HUMAINE model of human-centered, AI-assisted radiology

The findings from all modules are incorporated into a HUMAINE model for human-centered, AI-assisted radiology as integrated technology and work design (Fig. 2). The basic expectation is that the resulting model of human-centered radiological AI will meet all expectations if the module phase succeeds in  evaluating the interests of stakeholders and all relevant ethical criteria. Whether the application-specific goals, including a global ethical improvement, have been achieved remains open.

Outlook on the transfer potential for other use cases and contexts

The current use case, “AI-assisted Epi-MRT,” represents numerous requirements that are also placed on modern radiology and the people working in it. We therefore assume that most of the knowledge gained here can be transferred to other radiological use cases. 

The primary indication is that AI can benefit  both quality and quantity [31]. We believe that our modular approach to developing workplace-specific, human-centered AI applications in all areas of medicine can increase AI acceptance and contribute to better outcomes. In this context, the FUTURE-AI consensus criteria are suitable for reviewing numerous ethical expectations of human-centered clinical AI, both from the perspective of AI users and the patients “affected” by it. The expectation of added value should apply to all AI stakeholders.

If the circle of potential beneficiaries is expanded to include the healthcare system or stakeholders who were not the focus of the current analysis (such as manufacturers and operators of radiological hardware and software and communications infrastructure), it may be necessary to define modules other than, or in addition to, those invoked here. The reference to manufacturers and operators builds a bridge to the application of AI in industry and administration. Workflows in these areas can be just as complex as in medicine.

There, too, many different and sometimes conflicting stakeholder interests are present. We propose that the principle of human-centered AI development and application—structured into modules that include qualitative and ethical validation tailored to the respective use case—also be evaluated in this context. However, it is quite possible that other ethical catalogs are more applicable there [4, 32] than that applied in the current clinical use case.

The HUMAINE joint project in occupational science is funded by the German Federal Ministry of Research, Technology, and Space (BMFTR) FKZ 02L19C200 and supervised by the Project Management Agency Karlsruhe (PTKA). The authors are responsible for the content of this publication.

The original German version of this article can be accessed via DOI: 10.30844/I4SD.26.1.136

You might also be interested in

Tacit Talk A speech-based hybrid AI system for capturing tacit maintenance knowledge Vincent Philipp Brandhoff , Philipp Besinger, Joscha Zaremba , Daniel Valtiner , Michael Necemer , Safa Omri , Jens Neuhüttler , Fazel Ansari, Katharina Hölzle

The tacit expertise required for industrial maintenance is increasingly at risk due to demographic change and system complexity. This article presents Tacit-TALK, a speech-based system that externalizes tacit knowledge through AI-guided conversation. Spoken input is transcribed, analyzed by large language models (LLMs), and persisted in a knowledge graph, linking new insights to existing organizational data. Using a design science research process grounded in sociotechnical theory and instantiated for semiconductor maintenance, the article contributes design principles for AI-supported tacit knowledge capture derived from value-based stakeholder engagement, a reference architecture combining LLMs with knowledge graphs, and preliminary prototyping insights into user acceptance and organizational implications.

Industry 4.0 Science | Volume 42 | 2026 | Edition 5 | Pages 82-90 | DOI 10.30844/I4SE.26.5.16

AI-Based Identification of Knowledge Transfer Situations An adaptive multi-agent system for agile product development in engineering Georg David Ritterbusch , Ravil Goetzke, Norbert Gronau

Although it has been demonstrated that organizational knowledge transfer can be improved in principle, there is still no automated approach for identifying patterns in complex, context-dependent, and domain-specific knowledge transfer situations. This conceptual article therefore examines and characterizes knowledge transfer situations using the real-world example of product development in engineering. Furthermore, a concept for an adaptive, AI-based, real-time multi-agent system uses data to recognize recurring patterns in knowledge transfer situations and enables context-sensitive interventions. Finally, an outlook is provided on AI-based learning mechanisms (reinforcement learning) that can be used to adapt interventions for higher effectiveness in the long term.

Industry 4.0 Science | Volume 42 | 2026 | Edition 5 | Pages 110-116 | DOI 10.30844/I4SE.26.5.13

Experiential Knowledge Powered by AI  Practical insights from industry Martin Schmauder , Gritt Ott , Bianca Windisch

The use of experiential knowledge is a key success factor for companies. Based on four corporate case studies, this article analyzes the technical, organizational, and personnel challenges associated with the use of retrieval-augmented generation (RAG) systems. The results show that the success of such systems depends on the strategic development and maintenance of the knowledge base, as well as on employee engagement.

Industry 4.0 Science | Volume 42 | 2026 | Edition 5 | Pages 128-135 | DOI 10.30844/I4SE.26.5.15

Inclusive Work System Design Automation, standardization, and adaptability Sebastian Schlund

The design of inclusive work systems is gaining importance due to demographic change and the increasing digital penetration of value creation processes. While traditional ergonomic approaches are based primarily on percentile logic and thus address only a portion of the user population, the integration of digital technologies opens up new possibilities for the dynamic and individualized adaptation of work systems. This article presents a conceptual framework for inclusive work system design that integrates standardization, automation, and adaptability. Methodologically, the article is based on a conceptual analysis of existing approaches from ergonomics and human-centered design. The article makes a theoretical contribution to the systematization of inclusive work system design and identifies areas of focus for further research and industrial practice.

Industry 4.0 Science | Volume 42 | 2026 | Edition 5 | Pages 70-76 | DOI 10.30844/I4SE.26.5.8

MTM Analyses with AI and Rule-Based Algorithms An approach to interpreting textual process descriptions Constantin Eckart , Martin Benter, Peter Kuhlang

MTM methods are a proven standard for analyzing and designing human work processes. However, work planners continue to face challenges in applying these methods correctly and efficiently. Artificial intelligence — particularly in the case of large language models — holds significant potential for combatting these challenges. However, the use of AI raises legitimate questions regarding the reliability and traceability of the results. The approach presented here combines LLMs with a rule-based algorithm to extract the information required for MTM analyses from textual process descriptions such as work instructions. This information is then translated into MTM analyses in accordance with the MTM methodology. This approach ensures that the resulting analyses can be transparently traced back to the original input data by the user.

Industry 4.0 Science | Volume 42 | 2026 | Edition 5 | Pages 62-69 | DOI 10.30844/I4SE.26.5.7

AI-Powered Lubrication Strategies for Thread Forming Adaptive spray jet control to increase process reliability and tool life Reinhard Schmied, Marco Susic, Christian Donhauser

Thread forming requires precise lubricant application because high contact pressures and process temperatures strongly influence tool loading, friction, and process stability. Although minimum quantity lubrication (MQL) systems are widely used, current spray-based approaches can still suffer from spray losses, insufficient wetting of the thread grooves, and unstable droplet transport. This article presents a concept for adaptive precision lubrication in thread forming based on computational fluid dynamics (CFD)-supported flow analysis, experimental validation, and artificial intelligence (AI)-assisted optimization. The focus is on droplet size, spray jet geometry, nozzle position, ambient flow conditions, and their influence on wetting intensity. Preliminary simulation-based investigations indicate that data-driven optimization can help identify wetting deficiencies and support the development of future control strategies for resource-efficient lubricant application.

Industry 4.0 Science | Volume 42 | 2027 | Edition 3 | Pages 76-83