Democratizing AI in Healthcare with Open Medical Inference (OMI): Protocols, Data Exchange, and AI Integration

Abstract

Background

The integration of artificial intelligence (AI) into healthcare is transforming clinical decision-making, patient outcomes, and workflows. AI inference, applying trained models to new data, is central to this evolution, with cloud-based infrastructures enabling scalable AI deployment. The Open Medical Inference (OMI) platform democratizes AI access through open protocols and standardized data formats for seamless, interoperable healthcare data exchange. By integrating standards like FHIR and DICOMweb, OMI ensures interoperability between healthcare institutions and AI services while fostering ethical AI use through a governance framework addressing privacy, transparency, and fairness.

Method

OMI’s implementation is structured into work packages, each addressing technical and ethical aspects. These include expanding the Medical Informatics Initiative (MII) Core Dataset for medical imaging, developing infrastructure for AI inference, and creating an open-source DICOMweb adapter for legacy systems. Standardized data formats ensure interoperability, while the AI Governance Framework promotes trust and responsible AI use.

Conclusion

The project aims to establish an interoperable AI network across healthcare institutions, connecting existing infrastructures and AI services to enhance clinical outcomes.

Key Points

• OMI develops open protocols and standardized data formats for seamless healthcare data exchange.

• Integration with FHIR and DICOMweb ensures interoperability between healthcare systems and AI services.

• A governance framework addresses privacy, transparency, and fairness in AI usage.

• Work packages focus on expanding datasets, creating infrastructure, and enabling legacy system integration.

• The project aims to create a scalable, secure, and interoperable AI network in healthcare.

Citation Format

• Pelka O, Sigle S, Werner P et al. Democratizing AI in Healthcare with Open Medical Inference (OMI): Protocols, Data Exchange, and AI Integration. Rofo 2025; DOI 10.1055/a-2651-6653

Zusammenfassung

Hintergrund

Die Integration von Künstlicher Intelligenz (KI) im Gesundheitswesen verändert die klinische Entscheidungsfindung, Patientenergebnisse und Abläufe. KI-Inferenz, die Anwendung trainierter Modelle auf neue Daten, ist zentral für diese Entwicklung, wobei Cloud-Infrastrukturen skalierbare KI-Einsätze ermöglichen. Die Open Medical Inference (OMI)-Plattform demokratisiert den Zugang zu KI-Diensten durch offene Protokolle und standardisierte Datenformate für einen nahtlosen, interoperablen Austausch von Gesundheitsdaten. OMI integriert Standards wie FHIR und DICOMweb zur Sicherstellung der Interoperabilität zwischen Gesundheitseinrichtungen und KI-Diensten und fördert die ethische KI-Nutzung durch ein Governance-Framework für Datenschutz, Transparenz und Fairness.

Methode

Die Umsetzung von OMI erfolgt in Arbeitspaketen, die technische und ethische Aspekte abdecken. Dazu gehören die Erweiterung des Kerndatensatzes der Medizininformatik-Initiative (MII) für medizinische Bildgebung, die Entwicklung von Infrastrukturen für KI-Inferenz und die Erstellung eines Open-Source-DICOMweb-Adapters für Altsysteme. Standardisierte Datenformate gewährleisten Interoperabilität, während das KI-Governance-Framework Vertrauen und verantwortungsvolle Nutzung stärkt.

Schlussfolgerung

Das Projekt zielt darauf ab, ein interoperables KI-Netzwerk für Gesundheitseinrichtungen zu etablieren, bestehende Infrastrukturen und KI-Dienste zu verknüpfen und so die klinischen Ergebnisse zu verbessern.

Kernaussagen

• OMI entwickelt offene Protokolle und standardisierte Datenformate für den nahtlosen Austausch von Gesundheitsdaten.

• Die Integration mit FHIR und DICOMweb gewährleistet Interoperabilität zwischen Gesundheitssystemen und KI-Diensten.

• Ein Governance-Framework adressiert Datenschutz, Transparenz und Fairness in der KI-Nutzung.

• Arbeitspakete konzentrieren sich auf Datensatzerweiterung, Infrastrukturerstellung und Altsystemintegration.

• Das Projekt strebt ein skalierbares, sicheres und interoperables KI-Netzwerk im Gesundheitswesen an.

Keywords

AI Ethics in Healthcare - Artificial Intelligence in Healthcare - Healthcare Data Interoperability - Medical Inference

Introduction

Rapid advancements in artificial intelligence (AI) have revolutionized healthcare by enhancing clinical decision-making, improving patient outcomes, and optimizing operational efficiency. AI inference – the process of using trained models to predict outcomes from new data – plays a central role in this transformation. However, its effective implementation in healthcare requires robust technical infrastructure capable of handling large-scale data and enabling real-time processing. Cloud-based platforms have emerged as key enablers, offering scalable and secure environments that support high-performance AI models. These platforms significantly enhance clinical workflows by delivering timely and precise diagnostic capabilities [1].

Despite its potential, integrating AI inference into routine clinical practice remains challenging. Concerns about reliability, transparency, and the need for rigorous validation in diverse clinical settings often hinder adoption. Healthcare providers emphasize the importance of thorough testing to ensure the trustworthiness and safety of AI systems, as highlighted in [2]. Addressing these concerns is critical to fostering trust and achieving widespread clinical adoption.

Ethical considerations further complicate the integration of AI. The FAIR principles advocate for open access to data and algorithms, promoting transparency, collaboration, and reproducibility [3]. However, issues such as patient privacy, data security, and the risk of algorithmic bias must be addressed to prevent inequities in healthcare outcomes [4]. Research has shown that algorithmic bias, particularly in underrepresented patient populations, can exacerbate existing disparities, underscoring the need for equitable AI model development [5] [6].

The Open Medical Inference (OMI) [7] platform aims to address these challenges by providing innovative solutions for the discovery and utilization of remote AI services. OMI focuses on creating open protocols and data formats that enable semantically interoperable peer-to-peer exchanges of multimodal healthcare data and remote AI inference. Initially, OMI emphasizes image-centered multimodal AI models, leveraging the Data Sharing Framework (DSF) established by the Medical Informatics Initiative (MII) [8] [9].

OMI is designed to address specific challenges in AI integration, such as standardization, interoperability, and democratization of access to AI services. By building upon established national and international standards, OMI fosters harmonization across healthcare systems. Where no standards exist, OMI develops new open specifications to ensure seamless integration. Additionally, OMI plays a key role in advancing the medical imaging extension module of the MII Core Dataset through its active participation in the MII Working Group on Interoperability (WG IOP) [10].

By tackling the technical, ethical, and standardization challenges, OMI provides a comprehensive framework that bridges the gap between research and clinical practice. This socio-technical approach positions OMI as an innovative and practical solution for integrating AI into healthcare, ultimately advancing patient care and research outcomes.

Technological Foundations of OMI

The Open Medical Inference (OMI) platform advances healthcare interoperability by aligning with established standards such as HL7 FHIR and DICOMweb. OMI enables RESTful access to imaging data by defining FHIR endpoints for DICOMweb-enabled systems, simplifying integration with legacy PACS at German Data Integration Centers (DICs).

By leveraging the Data Sharing Framework (DSF) from the Medical Informatics Initiative (MII), OMI integrates pseudonymization, encryption, and secure data transport using modern web technologies (e.g., REST, TLS). The platform’s architecture supports modular deployment through containerized microservices, ensuring scalability, robustness, and adaptability across evolving clinical environments.

Exchange of Structured Multimodal Data

Successful AI deployment in healthcare depends on the availability of structured, high-quality multimodal data. OMI builds upon the MII’s semantically interoperable core dataset and uses the DICOM standard to model and exchange complex formats including medical images, waveforms, and encapsulated documents [11].

Combining imaging data with demographic or clinical metadata has been shown to enhance predictive performance and model reliability [12]. To support training and benchmarking, OMI incorporates best practices from curated resources such as the ROCO dataset [13], which provides multimodal radiological image annotations. Broader methodological surveys further emphasize the critical role of robust multimodal data integration for smart healthcare systems [14].

Big Data and Analytics Infrastructure

OMI aligns with broader trends toward data-driven healthcare ecosystems that integrate electronic health records (EHRs), imaging, genomics, and wearable sensor data. Studies have shown that while these data offer great potential, their ethical and technical integration must be carefully managed [15].

Scoping reviews on the translation of big data capabilities into clinical value emphasize the importance of interoperable, standardized platforms like OMI to support clinical decision-making, outcome prediction, and population health analytics [16].

Cloud-Enabled Inference and Scalability

To reduce the burden on local infrastructure, OMI supports the remote execution of AI inference using cloud-based or edge–cloud hybrid environments. These deployments follow architectural models that combine the Internet of Medical Things (IoMT), artificial intelligence, and distributed computation to enable real-time diagnostics and analytics [17] [18].

This flexible deployment model enables institutions to choose between local, cloud-based, or federated execution environments depending on performance, security, and governance requirements.

Data Privacy and Security in Healthcare

Ensuring the security and privacy of patient data is central to OMI’s architecture. The platform enforces GDPR-compliant practices including pseudonymization, encrypted communication channels, access controls, and mutual authentication protocols.

These implementations draw on state-of-the-art privacy-preserving machine learning techniques [19], and advanced cryptographic schemes such as chaos-based encryption models designed to protect sensitive clinical data in distributed environments [20].

Interoperability Standards, DICOMweb, and FHIR

OMI harmonizes several established interoperability standards to support seamless integration across heterogeneous health IT systems. Clinical data are structured using HL7 FHIR resources such as the Observation, Condition, and Imaging Study [21]. FHIR’s flexibility is extended via custom profiles and implementation guides that adapt to specific healthcare domains [22] [23].

DICOMweb – a RESTful extension of the traditional DICOM protocol – enables the efficient transmission, retrieval, and storage of imaging data using standard HTTP APIs such as WADO-RS and STOW-RS [24]. Together, these standards form the technical backbone of OMI’s inference architecture.

AI and Machine Learning in Medical Imaging

Machine learning, particularly deep learning, has revolutionized radiology through applications such as image segmentation, classification, and disease detection. Models trained on medical imaging data have demonstrated performance comparable to human experts with respect to certain diagnostic tasks [25].

OMI supports the clinical translation of such models by managing data preprocessing, inference orchestration, and structured reporting of model results [26]. Through its standard-based gateway architecture, OMI enables scalable and reproducible deployment of AI models within clinical workflows.

Ethical Considerations in AI for Healthcare

Despite technical advancements, the ethical deployment of AI remains a key challenge. Many biomedical AI models lack transparency and explainability, making it difficult to assess fairness or assign responsibility [27].

Recent policy reports and academic reviews have called for stronger institutional oversight, including the establishment of AI governance boards and ethical auditing frameworks [28] [29] [30]. OMI responds to these needs by incorporating explainability modules, fairness assessments, and a governance model aligned with international ethical standards. This approach reflects current debates on stakeholder responsibility, human–AI collaboration, and bias mitigation in healthcare AI [31] [32] [33] [34].

Remote AI Services and Market Fragmentation

Although multiple commercial platforms offer remote AI services – such as Aidoc [35], Arterys [36], Nuance Precision Imaging Network [37], and deepcOS [38] – these solutions remain largely proprietary and fragmented. This raises concerns about vendor lock-in, interoperability barriers, and monopolistic tendencies.

Calls for open, transparent alternatives have been growing [39], especially as AI becomes embedded in clinical infrastructure. OMI addresses these issues by providing a vendor-neutral, standards-based architecture with openly defined participation rules, fostering long-term accessibility, innovation, and trust across healthcare institutions [40].

OMI Approach and Framework

Primary Goals of OMI

OMI is designed as a methodological platform to advance AI integration and interoperability in healthcare. Its key objectives are supported by specific steps, tools, and validation procedures to ensure successful implementation.

1. Specification of Open Protocols and Data Formats

   • Steps: Develop technical specifications for open protocols and standardized data formats using international standards like FHIR and DICOMweb. Conduct iterative reviews with stakeholders to ensure broad compatibility and alignment with healthcare IT requirements.

   • Tools: Use protocol specification tools and validate compliance with existing interoperability standards using automated testing tools.

   • Validation: Simulated data exchanges between partner institutions to test semantic and syntactic interoperability.

2. Extension of the MII Core Dataset

   • Steps: Identify gaps in the current MII Core Dataset for representing medical imaging data. Define FHIR profiles for imaging data and AI-derived metadata, integrating these into the MII data catalog.

   • Tools: Utilize FHIR implementation guides and data modeling tools to develop and refine the dataset.

   • Validation: Engage with healthcare providers and researchers through collaborative sessions to evaluate the dataset’s usability and completeness.

3. Development of Infrastructure Components

   • Steps: Implement reference components such as a DIC-integrated client, gateway, and service registry. Conduct iterative development cycles with feedback from technical teams at partner DICs.

   • Tools: Use microservices architecture with containerization tools and orchestration platforms for scalability.

   • Validation: Perform system integration tests in controlled environments to ensure compatibility and reliability.

4. Creation of Additional DIC Infrastructure Components

   • Steps: Develop and test de-identification services for DICOM data using privacy-preserving algorithms. Incorporate pseudonymization and encryption for secure data transport.

   • Tools: Apply frameworks like Keycloak for authentication and DICOM Toolkits (e.g., dcm4che) for DICOM-specific operations.

   • Validation: Perform privacy compliance testing, including adherence to GDPR and local regulatory frameworks.

5. Deployment and Demonstration

   • Steps: Gradually deploy the infrastructure across partner DICs, beginning with pilot sites to identify and resolve initial challenges. Conduct technical demonstrations of remote AI inference.

   • Validation: Develop acceptance criteria, including uptime, latency, and accuracy metrics, and monitor performance during demonstration phases.

6. Publication of an AI Governance and Ethics Framework

   • Steps: Collaborate with ethicists, clinicians, and technologists to draft the framework. Address key issues like algorithmic transparency, bias, and fairness.

   • Validation: Conduct peer reviews and stakeholder consultations to ensure the framework’s relevance and comprehensiveness.

7. Connection of NUM-RACOON Nodes

   • Steps: Map the existing NUM-RACOON[1] [41] nodes to OMI’s architecture. Test data exchange and AI service integration across the nodes.

   • Validation: Measure success by evaluating data flow consistency and AI inference performance in real-world use cases.

Success Metrics and Evaluation

The effectiveness of OMI will be evaluated using a combination of technical and operational metrics:

• Interoperability: Number of successfully exchanged datasets between systems.

• Scalability: Performance of infrastructure under varying loads.

• User Acceptance: Feedback from clinicians and IT professionals at partner DICs.

• Ethical Compliance: Adoption rate of the AI Governance and Ethics Framework by participating institutions.

Testing and Validation Approaches

OMI will follow a structured testing process:

1. Pilot Testing: Evaluate individual components in isolated environments to ensure functionality.

2. Integration Testing: Test interoperability between components and partner systems.

3. End-to-End Testing: Simulate real-world use cases to validate system reliability and performance.

4. User Testing: Gather feedback from clinicians and IT teams to refine usability and workflow integration.

By pursuing these methods and addressing potential challenges, OMI aims to establish a comprehensive, reliable, and ethical platform for AI services in healthcare.

Implementation Approach

[Fig. 1] illustrates the interconnected components of the OMI project, highlighting how its design supports seamless data exchange, AI service integration, and ethical governance. OMI is organized into work packages that align with specific project objectives, including protocol development, MII Core Dataset extension, infrastructure implementation, and AI governance.

Fig. 1 OMI Inference Infrastructure: Hospitals or partners (shown on the left) without their own inference systems can request AI-based evaluations by submitting multimodal data to inference service providers (shown on the right). The OMI Service Registry (shown in the center) provides details on availability and estimated computation times prior to submission.

Specification of Open Protocols and Data Formats for Remote AI Services

Our primary objective is to develop an open, neutral, distributed, and secure data exchange system leveraging the existing DIC data sharing infrastructure for both non-commercial and commercial remote medical AI services. This system is based on an FHIR specification [27] of open protocols and data formats, ensuring semantic interoperability and leveraging of existing MII infrastructure. The specification will be continuously refined to accommodate new and advanced use cases, incorporating modern security concepts such as perimeterless security, mutual authentication, and strong encryption. The goal is to minimize entry barriers by reusing existing standards and focusing on widely used technologies like REST, TLS, FHIR, and DICOMweb. The design emphasizes interoperability with both internal and external MII activities, expanding upon existing FHIR specifications like the MII Core Dataset, Medical Information Objects, and hospital IT systems. Special focus is placed on integrating the DSF and ensuring compatibility and facilitating nationwide integration within the NUM and MII networks.

The protocol will also include an extensible specification for data sharing agreements, service terms, and license terms. Although the routine clinical use of remote AI services is beyond the project’s scope, future use will be considered in the protocol specifications, with guidance from the MII “Fit 4 Translation” [42] project for the safe development of medical software as a medical device.

Extension of the MII Core Dataset

A central contribution of OMI is the expansion of the MII Core Dataset to include a module for the semantically interoperable exchange of medical imaging data and structured representation of inference results. This will be achieved through close collaboration with MII stakeholders and associated partners (such as DRG [43], DGN [44], EuSoMII [45], gematik [46], NVIDIA, PlanetAI [47], PrivateAIM [48], and mio42 [49]). The FHIR-based imaging extension module will link FHIR and DICOM, ensuring compatibility with other disciplines and initiatives. This module will specify FHIR endpoint resources describing DICOMweb services, making DICOM data referenceable from the FHIR metadata description. Additionally, a generic FHIR framework for structured reporting of multimodal, image-based AI services will be developed, focusing on interoperability, reproducibility, and explainability.

Enabling Access to Legacy DICOM Networking

OMI provides an open-source DICOMweb adapter that enables RESTful access to legacy PACS and other DICOM nodes, as displayed in [Fig. 2]. This gateway server will translate traditional DICOM communication protocols to a vendor-neutral DICOMweb service, allowing client access. It will handle tasks like authentication, authorization, data conversion, and de-identification, simplifying access to imaging data for researchers and developers. The server will support media transcoding into common image and video formats, enabling lightweight viewers on devices lacking DICOM support.

Fig. 2 Overview of the architecture of the generic adapter component, which enables RESTful access to PACS and other DICOM nodes via a subset of the DICOMweb API specification. Client access to retrieve images is managed through existing infrastructure (dark blue) forwarding the request to the DICOMweb adapter (light blue) which handles image retrieval and de-identification from PACS.

Connection of Established AI Infrastructures to OMI

Inference of AI models on specialized hardware, such as graphical processing units (GPUs), requires careful orchestration of inputs, hardware availability, and outputs. Technologies like the NVIDIA Triton inference server [50] allows multiple AI models to run simultaneously on shared hardware. To connect OMI to inference servers using vendor-independent predict protocols, a reference gateway server will be developed. The components of the gateway server are displayed in [Fig. 3]. This server will enable communication between local components, leveraging the RACOON NODES network. It will include a local synchronization and routing module, allowing local communication between clinical users, RACOON-NODEs, DICs, and the central service registry. The gateway server will support a plug-in mechanism for AI services, providing common tasks like authentication, authorization, and monitoring. Beyond initial pilots within the RACOON network, OMI is being extended to demonstrate compatibility with previously published AI applications. Planned integrations include:

Fig. 3 The OMI gateway server (red box) connects inference clusters (blue box) via a plugin architecture. It handles requests like registering and inference for the OMI service registry and client.

• A deep learning model for tumor segmentation in brain MRI, previously validated in research settings, now being deployed via the OMI gateway for use in radiology departments.

• A lightweight AI classifier for chest X-ray triage, integrated into clinical workflow using FHIR-based data exchange.

These concrete case studies will serve as reusable reference implementations, enabling wider adoption beyond the project’s original imaging focus.

Integration of the OMI Platform into DIC Infrastructures

OMI will develop a client reference implementation to integrate the platform into the local MII-compliant infrastructure of partner sites. The client will interface with local DIC data transfer mechanisms, managing access to research data and enforcing consent and use & access policies. It will support multimodal inference with FHIR resources as input or output data, interface with local FHIR servers, and include a user interface component for image metadata review and quality assurance. The client will interface with the service registry to discover available services and manage inference jobs. It will forward relevant patient data to the OMI gateway server for AI inference and receive results, undergoing extensive testing for integration, performance, and usability.

Central Registry of AI Services

OMI proposes a central registry of available AI services, allowing clients to query and connect to services without manual configuration. The registry supports filters and tracks metrics like queue length and system load. OMI gateway instances register with the service registry, providing service descriptions and usage statistics. The registry is implemented for high availability, zero-downtime deployments, and failover. It conforms to the OMI API description and FHIR specification, with integration, unit testing, and load testing ensuring compatibility and performance.

AI Governance Guidelines and Ethical Principles

OMI aims to establish a federated network architecture for AI services across German university hospitals, offering an open and scalable interface for additional AI providers and users. A governance framework, based on international principles, will address legal, regulatory, clinical, ethical, and organizational aspects. Workshops with stakeholders from partner hospitals will refine the framework to meet the specific needs of these institutions.

A pragmatic version of this framework, including an AI implementation guideline and checklist, will be introduced in one or two partner hospitals, supported by a comprehensive training concept. This will cover fundamental AI concepts, the opportunities and risks of AI systems, and will be tailored for clinicians, nurses, and hospital staff. The day-to-day application of the AI guidelines, along with clinical feedback, will inform revisions to the framework before its broader rollout across additional OMI partner institutions.

Implementation Outline

This section outlines the implementation process for OMI, addressing challenges and practical feasibility in diverse clinical environments.

Practical Integration in Clinical Settings

OMI’s implementation strategy takes into account the variability of IT environments in healthcare settings. To ensure broad compatibility, the platform includes:

• Approach: The OMI gateway server and client reference implementation are designed to integrate with existing PACS, DICOM nodes, and FHIR servers, regardless of vendor-specific configurations.

• Example: A hospital using legacy PACS can adopt the OMI DICOMweb adapter, which translates traditional DICOM protocols into modern RESTful APIs, ensuring compatibility.

• Testing: Pilot studies at partner hospitals simulate diverse IT scenarios, including those with limited IT resources, to validate compatibility.

Addressing Implementation Barriers

OMI addresses several key challenges to ensure successful integration:

• Technical Compatibility: Variability in IT systems, including outdated hardware and software, can hinder integration. To mitigate this, OMI provides open-source tools like the DICOMweb adapter and offers integration guides tailored to diverse environments.

• Limited IT Resources: Smaller facilities may lack robust IT teams. OMI overcomes this by deploying lightweight tools requiring minimal local infrastructure, alongside remote training sessions and a dedicated helpdesk for ongoing support.

• Data Privacy Concerns: To address concerns about sharing sensitive data, OMI employs advanced de-identification, encryption, and mutual authentication to secure data exchange.

Enhancements to Foster Practical Feasibility

• Customizable Features: The OMI platform includes modular plugins for the gateway server, allowing hospitals to adapt functionalities to their specific needs.

• Iterative Testing Phases: Each component undergoes thorough integration testing in simulated environments to ensure usability and functionality before live deployment.

• Stakeholder Collaboration: Continuous engagement with healthcare professionals helps refine the platform and address real-world challenges proactively through workshops and feedback sessions.

Case Studies in Implementation

OMI will develop a repository of case studies documenting successful integrations across diverse clinical settings. These case studies will serve as reference materials, offering best practices and troubleshooting insights.

Focus on Long-Term Sustainability

OMI is designed with scalability and future readiness:

• Integration with national initiatives like NUM ensures alignment with broader healthcare goals.

• The use of established standards like FHIR and DICOM minimizes reliance on proprietary systems, ensuring long-term compatibility and reducing future integration challenges.

By addressing potential barriers and emphasizing collaboration with healthcare professionals, OMI ensures practical feasibility and smooth integration across various clinical settings.

Deployment Strategy

Implementation Guides and Standardization Efforts

The OMI platform aligns its tools, standards, and frameworks with national efforts to harmonize data exchange, ensuring compatibility and scalability. For instance:

• FHIR Implementation Guides: Detailed documentation for representing algorithms and specifying input/output parameters facilitates semantic searching and interoperability [51].

• Custom Plugin Architecture: The OMI client leverages state-of-the-art frameworks like Celery [52], an asynchronous task queue that ensures efficient and distributed task processing.

• Data Transmission: Secure middleware, such as the DSF, manages decentralized peer-to-peer (p2p) communication, ensuring robust automated process handling.

Insights from Early Adopters

Feedback from early adopters at participating hospitals provided critical insights:

• Practical Applicability: Clinical users emphasized the platform’s potential to streamline workflows, particularly for radiology and oncology. However, they noted the importance of robust training materials to support adoption.

• Integration Challenges: Some facilities faced initial difficulties integrating legacy systems, which were resolved through technical support and the deployment of the DICOMweb adapter.

Rollout and Evaluation

In the final phase of the project, OMI components will be deployed across partner institutions, acting as both service providers and users. Key evaluation steps include:

1. Functionality Testing: Comprehensive test suites will assess data exchange, inference accuracy, and protocol adherence.

2. Security Audits: Regular evaluations will ensure compliance with privacy regulations and security best practices.

3. User Experience Assessments: Feedback from clinicians and IT administrators will guide iterative improvements.

Test datasets and results will be published in the OMI source code repository to support transparency and collaboration.

Organizational Structure

OMI’s organizational structure ([Fig. 4]) ensures effective coordination across scientific and technical teams, leveraging expertise from diverse stakeholders. Thematic groups (TGs) oversee specification, implementation, and evaluation, thereby fostering synergy between MII [53] and NUM [54] initiatives. Regular workshops ensure that the platform evolves to meet user needs while adhering to regulatory standards.

Fig. 4 Overview of stakeholder interaction and collaboration within the OMI project, highlighting the integration of scientific and technical expertise through thematic groups and workshops. The image also illustrates the close cooperation with MII leveraging a nationwide medical imaging infrastructure to ensure synergy between MII and NUM initiatives. Additionally, it showcases the inclusion of key partners from medical societies, established infrastructures, and AI application manufacturers.

Ethical Principles and Governance Framework

OMI’s ethical framework emphasizes transparency, fairness, and data protection. To ensure effective implementation, the following measures are proposed:

Integration of Ethical Principles into Clinical Practice

• Policy Development: Hospitals should adopt clear policies based on OMI’s ethical guidelines, covering data handling, AI validation, and decision transparency.

• Training Programs: Regular training sessions should educate healthcare professionals on AI ethics, data protection, and AI result interpretation.

Compliance Monitoring Mechanisms

• Auditing and Reporting: Conduct periodic audits to ensure fairness, detect bias, and verify compliance with data usage policies.

• Real-Time Monitoring: Implement tools to track AI performance, logging decisions, outcomes, and data usage for accountability.

Stakeholder Engagement

• Workshops and Feedback Sessions: Engage clinicians, data scientists, and patients to refine governance practices and address ethical challenges.

• Transparency in Governance: Publish audit summaries and corrective measures to maintain trust and accountability.

Governance Framework Implementation

• Phased Rollout: Introduce governance gradually, starting with pilot sites across diverse clinical settings.

• Support from MII and DICs: Utilize MII and university hospital DICs to integrate governance tools into existing infrastructures.

• Incentives for Adoption: Encourage implementation by offering training resources and funding support.

Critical Success Factors

Achieving OMI’s objectives requires success in four key areas:

1. Support from MII: Continued collaboration ensures the effective deployment of the imaging extension module

2. Engagement with Data Integration Centers (DICs): Adapting OMI nodes to existing hospital infrastructures is crucial for seamless integration.

3. Adoption of the OMI Protocol: A flexible and evolving specification ensures compatibility and widespread acceptance.

4. Collaboration with Commercial AI Platforms: Advocacy from AI customers will drive platform providers to adopt OMI, thus expanding its reach.

Case Study Repository and Broader Integration

To support adoption across diverse clinical contexts, OMI is building a publicly accessible repository of implementation case studies. These case studies document successful deployment beyond the initial RACOON imaging network, showcasing OMI’s flexibility and interoperability. Current examples include the integration of a deep learning model for brain tumor segmentation on MRI and a chest X-ray triage classifier. Both models, previously validated in research settings, are now being deployed in hospital environments via OMI’s gateway infrastructure. These implementation cases serve as reference blueprints, demonstrating how external AI applications can be reused and integrated into clinical workflows using OMI’s open protocols and standards-based architecture.

OMI integrates state-of-the-art tools, frameworks, and standards to tackle challenges in machine-to-machine communication, process automation, and decentralized service deployment. Its structured approach to algorithm input and output preserves data semantics, unlocking the potential of legacy data for AI applications. Collaboration with the Medical Informatics Initiative (MII) and its interoperability working group (WG IOP) ensures alignment with national standards for the imaging extension module of the MII Core Dataset (CDS). Partnerships with NUM, particularly RACOON, strengthen OMI’s medical imaging infrastructure, while collaboration with gematik and commercial AI developers supports AI governance and ethics frameworks. We recommend that OMI university hospital partners pragmatically implement the AI governance framework, starting with concise AI guidelines and a checklist to assist clinicians with respect to AI adoption. These should evolve iteratively, incorporating technological advances and clinical feedback.

Outlook

The Open Medical Inference (OMI) protocol is transforming AI-driven healthcare by enabling seamless remote AI inference in research and clinical settings. Through open protocols and standardized data formats, OMI facilitates secure multimodal medical data exchange, leveraging AI for better patient outcomes. Once implemented, OMI will streamline interactions between healthcare providers and AI vendors, simplifying AI model management. Its machine-to-machine communication will optimize pipeline development, automating data retrieval and AI-based analysis to accelerate discoveries in personalized medicine, diagnostics, and operational efficiency.

Realistic Assessment of Long-Term Goals

Achieving OMI’s vision requires overcoming technical and regulatory obstacles. The following roadmap outlines key steps:

• Short-Term Goals:

  • Pilot Deployments: Initial tests in imaging applications must validate functionality and reliability. Collaboration with stakeholders ensures seamless integration with existing infrastructures.

  • User Adoption: Encouraging adoption by healthcare providers and AI vendors requires demonstrating tangible benefits, such as improved diagnostics and streamlined operations, through case studies and pilot programs.

  • Regulatory Compliance: Early efforts should secure regulatory guidance and establish compliance mechanisms for data protection, AI validation, and decision transparency.

• Medium-Term Goals:

  • Expansion to Additional Disciplines: OMI must adapt to diverse medical fields, supporting various data formats and AI models.

  • Interoperability and Standardization: Standardizing protocols and ensuring compatibility with current and future healthcare infrastructures are essential.

  • Building Trust and Transparency: Governance frameworks must ensure fairness and accountability. Feedback from early adopters will refine the system and support ethical AI deployment.

• Long-Term Goals:

  • Widespread Healthcare Adoption: Advocacy, technical support, and incentives will drive integration into institutional workflows.

  • Global Scalability: International collaboration will require compliance with global data regulations and partnerships for large-scale AI-driven research.

  • Continuous Innovation: Adapting to evolving medical needs, new data types, and AI advancements will ensure OMI’s long-term relevance.

Technical and Regulatory Challenges

To achieve these goals, key challenges must be addressed:

• Integration with Existing IT Systems: Many healthcare institutions rely on legacy systems incompatible with OMI. Ensuring seamless data exchange requires significant effort.

• Data Privacy and Security: Compliance with healthcare data protection laws and robust cybersecurity measures is crucial for maintaining trust.

• Regulatory Approval for AI in Healthcare: The evolving AI regulatory landscape necessitates strategic planning and collaboration with regulatory bodies to secure approvals across medical disciplines.

Vision and Impact

By aligning technological innovation with ethical oversight and open standards, OMI contributes to a sustainable and trustworthy digital health ecosystem. The platform’s modular, standards-based design positions it as a blueprint for national and international efforts to democratize access to clinically meaningful AI services. The ongoing publication of open tools, datasets, and protocols ensures that OMI remains adaptable, extensible, and responsive to real-world needs.

Conclusion

OMI’s open, federated network architecture offers an exciting future for AI in healthcare, with the potential to drive significant advancements across a wide range of medical disciplines. However, achieving this vision requires a step-by-step approach, addressing technical challenges, navigating regulatory hurdles, and fostering collaboration among stakeholders. By setting realistic short- and long-term goals, OMI can successfully scale and contribute to the transformation of healthcare through AI-driven innovation.

List of Abbreviations

OMI: Open Medical Inference

FHIR: Fast Healthcare Interoperability Resources

DSF: Data Sharing Framework

p2p: Peer-to-Peer

DIC: Data Integration Centers

EHR: Electronic Health Records

DICOM: Digital Imaging and Communications in Medicine

IHE: Integrating the Healthcare Enterprise

HL7: Health Level Seven International

CNN: Convolutional Neural Networks

JSON: JavaScript Object Notation

WADO-RS: Web Access to DICOM Objects – Retrieve Service

QIDO-RS: Query for DICOM Objects – Retrieve Service

STOW-RS: Store to DICOM Objects – Retrieve Service

HIPAA: Health Insurance Portability and Accountability Act

TG: Topic Groups

MII: Medical Informatics Initiative

NUM: Network of University Medicine

CDS: Core Dataset

WG IOP: Working Group for Interoperability

AI: Artificial Intelligence

PACS: Picture Archiving and Communication System

RACOON: Radiological Cooperative Network

REST: Representational State Transfer

TLS: Transport Layer Security

GPU: Graphics Processing Unit

DRG: Deutsche Röntgengesellschaft (German Radiological Society)

DGN: Deutsche Gesellschaft für Nuklearmedizin (German Society for Nuclear Medicine)

EuSoMII: European Society of Medical Informatics

mio42: Medical Intelligence and Operations 42

GDPR: General Data Protection Regulation

Conflict of Interest

The authors declare that they have no conflict of interest.

¹ RACOON is funded by “NUM 2.0” (FKZ: 01KX2121). The Network of University Medicine (NUM) was established in April 2020 as part of the COVID-19 crisis response, with the goal of coordinating clinical research on COVID-19 across academic medical centers.

• References

• 1 Topol EJ. High-performance medicine: the convergence of human and artificial intelligence. Nature Medicine 2019; 25 (01) 44-56
• 2 McKinney SM, Sieniek M, Godbole V. et al. International evaluation of an AI system for breast cancer screening. Nature 2020; 577: 89-94
• 3 Wilkinson MD, Dumontier M, Aalbersberg IJ. et al. The FAIR Guiding Principles for scientific data management and stewardship. Scientific Data 2016; 3: 160018
• 4 Chen F, Wang L, Hong J. et al. Unmasking bias in artificial intelligence: a systematic review of bias detection and mitigation strategies in electronic health record-based models. J Am Med Inform Assoc 2024; 31 (05) 1172-1183
• 5 Boniface M, Haines R, Tan K. et al. Empowering the edge with AI: the AI4EU experiment. AI Communications 2020; 33 (02) 239-253
• 6 Boeckhout M, Zielhuis GA, Bredenoord AL. The FAIR guiding principles for data stewardship: fair enough?. Eur J Hum Genet 2018; 26: 931-936
• 7 Open Medical Inference (OMI) Homepage. Zugriff am 05. März 2025 unter: https://omi.ikim.nrw/#partners
• 8 Data Sharing Framework (DSF) Homepage. Zugriff am 05. März 2025 unter: https://github.com/datasharingframework/dsf
• 9 Hund H, Wettstein R, Heidt CM. et al. Executing Distributed Healthcare and Research Processes – The HiGHmed Data Sharing Framework. Stud Health Technol Inform 2021;
• 10 Ammon D, Kurscheidt M, Buckow K. et al. Arbeitsgruppe Interoperabilität: Kerndatensatz und Informationssysteme für Integration und Austausch von Daten in der Medizininformatik-Initiative. Bundesgesundheitsbl 2024; 67: 656-667
• 11 Pelka O, Friedrich CM, Nensa F. et al. Sociodemographic data and APOE-ε4 augmentation for MRI-based detection of amnestic mild cognitive impairment using deep learning systems. PLOS ONE 2020; 15 (09)
• 12 Cai Q, Wang H, Li Z. et al. A survey on multimodal data-driven smart healthcare systems: approaches and applications. IEEE Access 2019; 7: 133583-133599
• 13 Pelka O, Koitka S, Rückert J. et al. Radiology objects in context (ROCO): a multimodal image dataset. In CVII-STENT & LABELS 2018 Workshops. In: . Springer; 180-189
• 14 Shaik T, Tao X, Li L. et al. A survey of multimodal information fusion for smart healthcare: mapping the journey from data to wisdom. Information Fusion 2024; 102040
• 15 Ienca M, Ferretti A, Hurst S. et al. Considerations for ethics review of big data health research: a scoping review. PLOS ONE 2018; 13 (10) e0204937
• 16 Brossard PY, Minvielle E, Sicotte C. The path from big data analytics capabilities to value in hospitals: a scoping review. BMC Health Services Research 2022; 22 (01) 134
• 17 Sun L, Jiang X, Ren H. et al. Edge-cloud computing and artificial intelligence in internet of medical things: architecture, technology and application. IEEE Access 2020; 8: 101079-101092
• 18 Monteiro ACB, França RP, Arthur R. et al. An overview of medical Internet of Things, artificial intelligence, and cloud computing employed in health care from a modern panorama. In The Fusion of IoT, AI, and Cloud Computing in Healthcare 2021; 3-23
• 19 Guerra-Manzanares A, Lopez LJL, Maniatakos M. et al. Privacy-preserving machine learning for healthcare: open challenges and future perspectives. In Trustworthy Machine Learning for Healthcare. In: . Springer; 25-40
• 20 Rehman MU, Shafique A, Ghadi YY. et al. A novel chaos-based privacy-preserving deep learning model for cancer diagnosis. IEEE Trans. on Network Science and Engineering 2022; 9 (06) 4322-4337
• 21 Ayaz M. et al. The Fast Healthcare Interoperability Resources (FHIR) standard: systematic literature review. JMIR Med. Inform 2021; 9 (07) e21929
• 22 de Mello BH, Rigo SJ, da Costa CA. et al. Semantic interoperability in health records standards: a systematic review. Health and Technology 2022; 12 (02) 255-272
• 23 Nan J, Xu LQ. Designing interoperable healthcare services based on FHIR: literature review. JMIR Med. Inform 2023; 11 (01) e44842
• 24 Genereaux BW. et al. DICOMweb: background and application of the web standard for medical imaging. J Digit Imaging 2018; 31 (03) 321-326
• 25 Schaffter T. et al. Evaluation of combined AI and radiologist assessment to interpret screening mammograms. JAMA Netw Open 2020; 3 (03)
• 26 Shen D, Wu G, Suk HI. Deep learning in medical image analysis. Annu Rev Biomed Eng 2017; 19: 221-248
• 27 Han H, Liu X. The challenges of explainable AI in biomedical data science. BMC Bioinformatics 2021; 22 (Suppl. 12) 443
• 28
• 29 Chen Y, Clayton EW, Novak LL. et al. Human-centered design to address biases in AI. J Med Internet Res 2023; 25: e43251
• 30 Borkowski AA, Jakey CE, Thomas LB. et al. Establishing a hospital artificial intelligence committee to improve patient care. Fed Pract 2022; 39 (08) 334-336
• 31 Karimian G, Petelos E, Evers SM. The ethical issues of AI in healthcare: a systematic scoping review. AI and Ethics 2022; 2 (04) 539-551
• 32 Čartolovni A, Tomičić A, Mosler EL. Ethical, legal, and social considerations of AI-based decision-support. Int J Med Inform 2022; 161: 104738
• 33 Apfelbacher T, Koçman SE, Prokosch HU. et al. A governance framework for AI applications in hospitals. Stud Health Technol Inform 2022; 316: 776-780
• 34 Bekbolatova M, Mayer J, Ong CW. et al. The transformative potential of AI in healthcare: navigating ethics and public views. Healthcare 2024; 12 (02) 125
• 35 AIDOC Homepage. Zugriff am 05. März 2025 unter: https://www.aidoc.com/
• 36 Arterys Homepage. Zugriff am 05. März 2025 unter: https://www.arterys.com/
• 37 Nuance Precision Imaging Network. Zugriff am 05. März 2025 unter: https://www.nuance.com/healthcare/
• 38 Deepc Homepage. Zugriff am 25. September 2024 unter: https://www.deepc.ai/
• 39 Alowais SA, Alghamdi SS, Alsuhebany N. et al. Revolutionizing healthcare: the role of AI in clinical practice. BMC Med Educ 2023; 23: 689
• 40 Bajwa J, Munir U, Nori A. et al. Artificial intelligence in healthcare: transforming the practice of medicine. Future Healthc J 2021; 8 (02) e188-e194
• 41 Radiological Cooperative Network (RACOON) Homepage. Zugriff am 05. März 2025 unter: https://www.netzwerk-universitaetsmedizin.de/projekte/racoon
• 42 Fit 4 Translation Homepage. Zugriff am 05. März 2025 unter: https://fit4translation.de/
• 43 Deutsche Röntgengesellschaft (DRG) Homepage. Zugriff am 05. März 2025 unter: https://www.drg.de/
• 44 Deutsche Gesellschaft für Neurologie (DGN) Homepage. Zugriff am 05. März 2025 unter: https://dgn.org/uber-uns
• 45 European Soceity of Medical Imaging Informatics (EUSOMII) Homepage. Zugriff am 05. März 2025 unter: https://www.eusomii.org/
• 46 Gematik Homepage. Zugriff am 05. März 2025 unter: https://www.gematik.de/
• 47 PlanetAI Homepage. Zugriff am 05. März 2025 unter: https://planet-ai.com/
• 48 PrivateAim Homepage. Zugriff am 05. März 2025 unter: https://privateaim.de/
• 49 mio42 GmbH Homepage. Zugriff am 05. März 2025 unter: https://mio42.de/
• 50 NVIDIA Triton Inference Server. Zugriff am 05. März 2025 unter: https://developer.nvidia.com/triton-inference-server
• 51 Sigle S, Werner P, Schweizer S. et al. Bridging the Gap Between (AI-) Services and Their Application in Research and Clinical Settings Through Interoperability: the OMI-Protocol. 2024.
• 52 Celery python framework. Zugriff am 05. März 2025 unter: https://github.com/celery
• 53 Medical Informatics Initiative Homepage. Zugriff am 05. März 2025 unter: https://www.medizininformatik-initiative.de/de/start
• 54 Network of University Medicine Homepage. Zugriff am 05. März 2025 unter: https://www.netzwerk-universitaetsmedizin.de/

Publikationsverlauf

Eingereicht: 28. März 2025

Angenommen nach Revision: 03. Juli 2025

Artikel online veröffentlicht:
29. September 2025