Background

Efforts in the European Union (EU) to develop a legal framework for safe and trustworthy artificial intelligence began in 2019 and paved the way for the Artificial Intelligence Act (AIA) in 2024. The AIA uses a risk-based approach to address ethical, safety, and transparency concerns for AI products and services from the development phase to their application. In high-risk AI systems used in the healthcare sector, the manufacturer has to ensure data governance along with risk management.

In the EU, medical devices are regulated under (EU) 2017/745 Medical Device Regulation (MDR). While MDR also uses a risk-based approach, the requirements related to risk related to medical devices do not explicitly address the risks specific to AI systems. Here. The AIA introduces a complimentary approach with requirements to address hazards and risks for the health, safety, and rights specific to AI systems.

Important Dates

• 1st of August 2024 – The EU AI Act came into force.
• 2nd of February 2025 - Prohibitions of AI systems deemed to present an unacceptable risk apply.
• 2nd of August 2025 – Deadline for the Member States to designate national competent authorities, who will oversee the application of the rules for AI systems and carry out market surveillance
• 2nd of August 2026 – Date on which the majority of rules of the AI Act will start applying.
• 2nd of August 2027 - Rules for high-risk AI systems (e.g., medical devices) under Article 6(1) become applicable.

The Commission has launched the AI pact, aimed to bridge the transitional period before full implementation. Through the AI pact, AI developers are invited to voluntarily adopt key obligations of the AI Act ahead of the legal deadlines.

Applicability to Radformation Products

Certain Radformation products (e.g., AutoContour) are considered Software as a Medical Device (SaMD) and are subject to a number of regulations in the EU. They are Class IIa under MDR and therefore, third-party conformity assessment by a Notified Body (NB) is required.

As such, the products are considered high-risk and fall within the scope of the AIA. However, when multiple rules apply (MDR, AIA), the more specific one takes precedence. Per Article 8, Paragraph 2 of the AIA, in order to minimize the burden on manufacturers, the required testing and reporting processes, information, and documentation under the AIA will be incorporated into existing MDR documentation to be evaluated by the Notified Body.

This document serves to describe the steps taking by Radformation to ensure compliance with the EU AIA, as well as the following:

• Regulation (EU) 2016/679 of the European Parliament and of the Council of 27 April 2016 on the protection of natural persons with regard to the processing of personal data and on the free movement of such data (GDPR)
• Belgian law of July 30, 2018 relating to the protection of individuals with regard to the processing of personal data.
• Regulation (EU) 2017/745 on Medical Devices (MDR)

Radformation Compliance with EU

The AI Act came into force during AutoContour’s Conformity Assessment. With the date of applicability being the 2nd of August, 2027, assessment procedures for the AIA are still under development, as discussed in the Team-NB paper on AIA adopted in April 2025.

However, a dedicated AI Technical review was already added to the Notified Bodies evaluation team for the purpose of assessing compliance with the EU AI act as well as existing related regulations. AutoContour completed this assessment process in December 2024, successfully passing the Clinical, Technical and a newly added preliminary AI technical review processes. A formal "Regulatory Letter” from our Notified Body, confirming the favorable review of our AI processes is included in the “References’ section.

Specific AIA Considerations Relevant to Radformation Products

No use of “prohibited practices”

Radformation products do not employ any prohibited practices as defined by the AI Regulation. Although the capability of the software, within the scope of its Intended Use, can produce a 3D set of coordinates representing the external surface of the individual represented in the input medical image, the functionality is limited solely to this form of output and there is no further functionality within the software to support visualization or any of the categorizations defined within the “prohibited practices” of Article 5 of the EU AI Act.

No practices per Article 50, Paragraph 3 of the AIA

As with the prohibited practices above, there is no further functionality within the software to support emotion recognition system or biometric categorization as defined within the Act.

Patient-centric approach, including the transparency of devices to users

Transparency implies being open about the process details that are relevant for the users. Radformation closely follows the international contouring guidelines for developing AutoContour. Having this approach, the users can rely on what has been used as reference. Radformation also includes relevant information about the algorithm limitations in the instructions for use, and shares additional details about the development and test data within the Appendix of our User Manual, Structure Model Atlas and AutoContour AI White Paper documents available to its customers.

Methods for the elimination of ML algorithm bias and algorithm improvement

Radformation’s AutoContour product implements a rigorous data curation and training framework that is detailed within our SOP-00031-01 - AI Model Development Policy. This includes sourcing training data from a diverse set of institutions across multiple countries, encompassing a wide range of patient demographics, imaging modalities, and clinical scenarios. Ground truth contours are validated by expert clinicians and aligned with standardized contouring guidelines. Additionally, data augmentation techniques are applied to address imaging-related biases such as patient position, size, and noise level.

Radformation encourages and facilitates an active customer feedback loop to help identify scenarios where models may underperform within previously unidentified patient sub-populations or because of advances in medical image acquisition where the input images differ significantly from the training datasets. Based on this feedback, and where a single generalizable model is found insufficient, new training dataset information or model variations (e.g. female specific pelvis OAR models) can be incorporated into new AI modelling that occurs between sequential releases of AutoContour software, allowing appropriate optimization of the model performance for those patient sub-groups.

Real-world performance monitoring pilots

Radformation maintains a system for continuously monitoring and documenting usage, user feedback, feature requests and/or complaints, as well as actively performing customer surveys and monitoring adverse event databases. Independent clinical validation studies are encouraged and shared by customers, allowing the company to learn from their results.

Software usage (number of runs) is monitored via the Radformation website. Customer feedback/complaints are continuously collected via the Radformation Support and Success teams and are analyzed for trends annually. Since May 2022, the number of annual AutoContour runs has exceeded 330,000 and there have been no complaints which resulted in patient harm or adverse events.

Radformation AutoContour is being evaluated as part of the 3-year UK NICE health technology evaluation and was approved for use by the early value assessment (CE Mark now granted). This evaluation aims to reveal the real-world benefits via an evidence creation program.

Radformation uses customer feedback in scenarios where models are discovered to underperform for specific patient sub-populations or because of advances in medical image acquisition. Based on this feedback, and where a single generalizable model is found insufficient, new training dataset information or model variations (e.g. female specific pelvis OAR models) can be incorporated into new AI modelling that occurs between sequential releases of AutoContour software, allowing appropriate optimization of the model performance for those patient sub-groups

High relevance of available data to the clinical problem and current clinical practice

Radformation training datasets are composed of images acquired from the general population of patients receiving radiotherapy treatment. The datasets are not restricted on the basis of age, ethnicity, race, gender, or disease states. A mix of healthy patients and patients with disease were used as training data. By sampling a large volume of data from a variety of institutions, countries, and including publicly available datasets; a general population of patients receiving radiotherapy treatment was created, representing the target population.

Consistency in data collection that does not deviate from the SaMD’s intended use

Data collection is consistent with the intended use, which is the auto-segmentation of radiation therapy structures.

Planned modification pathway

Radformation manages products over the entire lifecycle under a certified ISO 13485:2016 quality system. Radformation has planned modifications for minor and major changes, and each AutoContour update includes a clear change plan prior to modification.

Appropriate boundaries in the datasets used for training, tuning, and testing the AI algorithms

Radformation defines the boundaries for dataset use at the patient level: all data from a single patient may only be used for a single purpose (e.g. training, or validation, or testing, etc.). Radformation defines this single purpose use as “a single data split”. The details of the data splits, including source and demographic information belonging to the training and validation datasets, are provided in the SOP-00031-01 - AI Model Development Policy and a summary for customers is presented in the AutoContour User Manual, Appendix A.

Conclusion

Radformation has taken all necessary steps to ensure conformity with the AI Act and the aforementioned regulations. Radformation provides this signed declaration declaring that it is in compliance and will continue to comply with the aforementioned legislation in the context of the use of AI in our Medical Device Applications and in the fulfillment of our contractual agreements. In the event that the user needs to report any incident, they can email Radformation support (support@radformation.com) and we will attend your request as soon as possible.

References

• Team-NB Position Paper, “European Artificial Intelligence Act”, April 2025
• AIB 2025-1/MDCG 2025-6, “Interplay between the Medical Devices Regulation (MDR) & In vitro Diagnostic Medical Devices Regulation (IVDR) and the Artificial Intelligence Act (AIA)”
• BSI Radformation Regulatory Letter - AI Review (2025-07-24)