Periodic Reporting for period 3 - CHAIMELEON (Accelerating the lab to market transition of AI tools for cancer management)
Berichtszeitraum: 2023-09-01 bis 2025-02-28
Every year, 3.7 million new cancer cases are diagnosed in Europe, underscoring the urgent need for advanced tools to support cancer management and improve treatment pathways. In this context, the use of Artificial Intelligence (AI) applied to health data holds immense promise across a range of clinical applications. Yet, the development of robust, imaging-based AI tools continues to be limited by the availability of large-scale, high-quality, and representative datasets. Building such datasets—imaging biobanks—remains a resource-intensive challenge, often hindered by technical complexity, institutional variability, and stringent legal and ethical requirements.
The CHAIMELEON project was conceived to address these key barriers by establishing a cancer imaging repository designed to facilitate access to anonymized imaging datasets of high quality and clinical relevance. Now completed, the project has successfully delivered a distributed data infrastructure that is interoperable with existing biobanks and enables secure data sharing and reuse, functioning as a single-access resource for the AI development community.
To overcome the challenges of cross-institutional variability in image acquisition—which can compromise the reproducibility of quantitative imaging biomarkers (QIBs)—the project has tested and validated image harmonisation protocols, thereby enabling the scientific reuse of multicentre retrospective imaging data.
Throughout its implementation, the CHAIMELEON project has achieved the following key outcomes:
• Deployment of a data repository enabling secure share and reuse of health data for the AI developer community.
• Development and integration of tools for data ingestion, curation, annotation, and harmonisation.
• Implementation of methodologies to enhance the interpretability and robustness of AI models.
• Creation of legal and ethical compliance protocols for data access and sharing.
These achievements not only contribute to bridging the data gap in AI for cancer imaging but also provide a sustainable foundation for the continued development of clinically relevant and trustworthy AI solutions.
The CHAIMELEON project was conceived to address these key barriers by establishing a cancer imaging repository designed to facilitate access to anonymized imaging datasets of high quality and clinical relevance. Now completed, the project has successfully delivered a distributed data infrastructure that is interoperable with existing biobanks and enables secure data sharing and reuse, functioning as a single-access resource for the AI development community.
To overcome the challenges of cross-institutional variability in image acquisition—which can compromise the reproducibility of quantitative imaging biomarkers (QIBs)—the project has tested and validated image harmonisation protocols, thereby enabling the scientific reuse of multicentre retrospective imaging data.
Throughout its implementation, the CHAIMELEON project has achieved the following key outcomes:
• Deployment of a data repository enabling secure share and reuse of health data for the AI developer community.
• Development and integration of tools for data ingestion, curation, annotation, and harmonisation.
• Implementation of methodologies to enhance the interpretability and robustness of AI models.
• Creation of legal and ethical compliance protocols for data access and sharing.
These achievements not only contribute to bridging the data gap in AI for cancer imaging but also provide a sustainable foundation for the continued development of clinically relevant and trustworthy AI solutions.
Throughout the course of the project, the following main results have been achieved:
- Creation, set-up and external validation of the CHAIMELEON platform equipped with a set of data analysis functionalities including: (1) in-house image harmonisation tools, (2) AI security algorithms, (3) a data analysis engine supporting creation of customised datasets, (4) data privacy mechanisms via distributed learning, data augmentation, and privacy preserving models, and (5) explainability tools to generate trust amongst clinicians and measure a model’s reliability.
- Establishment of a database encompassing 20,000+ prostate, lung, breast, colon and rectum cancers cases following strict inclusion criteria. Of note, all these cases were collected from 9 consortium sites and 7 external data providers across 9 different countries aiding to the heterogeneity of the data.
- AI developers worldwide were invited to participate in an Open Challenge aimed at fostering the development of clinical decision-making solutions using the CHAIMELEON platform. This initiative enabled participants to access the platform to train and refine their AI models. It also served as a critical feedback mechanism, which—alongside internal partner validation—supported iterative optimization cycles that enhanced both platform performance and usability.
- Clinical validation of the developed AI models was conducted using a dedicated, in-house validation platform specifically designed and customised for scalability. This approach enabled 46 clinicians from 12 EU hospitals to test the AI models in silico using real-world data. Their feedback on the models’ clinical utility played a key role in bridging the lab-to-bedside gap and advancing the integration of these innovative technologies into routine clinical practice.
- Agreement for the sustainability of the CHAIMELEON database and tools developed throughout the course of the action through their integration in Cancer Image Europe (EUCAIM).
Altogether, the work performed in the CHAIMELEON project has led to 54 publications (with 5+ currently being prepared) and has created 3 main exploitable assets: the CHAIMELEON platform − including high-quality datasets −, harmonisation tools and AI predictive models.
- Creation, set-up and external validation of the CHAIMELEON platform equipped with a set of data analysis functionalities including: (1) in-house image harmonisation tools, (2) AI security algorithms, (3) a data analysis engine supporting creation of customised datasets, (4) data privacy mechanisms via distributed learning, data augmentation, and privacy preserving models, and (5) explainability tools to generate trust amongst clinicians and measure a model’s reliability.
- Establishment of a database encompassing 20,000+ prostate, lung, breast, colon and rectum cancers cases following strict inclusion criteria. Of note, all these cases were collected from 9 consortium sites and 7 external data providers across 9 different countries aiding to the heterogeneity of the data.
- AI developers worldwide were invited to participate in an Open Challenge aimed at fostering the development of clinical decision-making solutions using the CHAIMELEON platform. This initiative enabled participants to access the platform to train and refine their AI models. It also served as a critical feedback mechanism, which—alongside internal partner validation—supported iterative optimization cycles that enhanced both platform performance and usability.
- Clinical validation of the developed AI models was conducted using a dedicated, in-house validation platform specifically designed and customised for scalability. This approach enabled 46 clinicians from 12 EU hospitals to test the AI models in silico using real-world data. Their feedback on the models’ clinical utility played a key role in bridging the lab-to-bedside gap and advancing the integration of these innovative technologies into routine clinical practice.
- Agreement for the sustainability of the CHAIMELEON database and tools developed throughout the course of the action through their integration in Cancer Image Europe (EUCAIM).
Altogether, the work performed in the CHAIMELEON project has led to 54 publications (with 5+ currently being prepared) and has created 3 main exploitable assets: the CHAIMELEON platform − including high-quality datasets −, harmonisation tools and AI predictive models.
The CHAIMELEON project has been designed to make a significant impact on the management of the five most prevalent types of cancer worldwide through the development of the CHAIMELEON repository and its associated AI-powered tools. Given the substantial social and economic burden posed by these cancers, we anticipate that the outcomes of this project will have a broad EU-wide impact at both societal and economic levels. Furthermore, the project's repository infrastructure, legal and operational framework, analysis tools, and web-based interfaces are scalable and adaptable, offering potential for application in the management of other cancer types as well.
In terms of advancing the state of the art in AI for health imaging, the project has established a framework for legitimate access to anonymized imaging and associated clinical data, thereby supporting broader EU-wide data accessibility for secondary research use. Technically, we have contributed to key areas including the robustness of AI systems against adversarial attacks, model interpretability, and the validation of AI tools through clinical observational studies.
Additionally, the Consortium has made substantial progress in the standardization of radiological procedures for image acquisition and analysis, and in the harmonization of methodologies for extracting reproducible imaging biomarkers.
From a legal and ethical standpoint, the project has helped define ethical standards for the use of health imaging data in the development of AI-based tools. Collectively, these efforts have resulted in resources that support faster training and more clinically relevant AI solutions for cancer care, while also contributing to policy evolution concerning the reuse of health data for research across Europe.
Ultimately, the project has played a key role in fostering trust in AI among healthcare professionals, patients, and stakeholders in both academia and industry.
In terms of advancing the state of the art in AI for health imaging, the project has established a framework for legitimate access to anonymized imaging and associated clinical data, thereby supporting broader EU-wide data accessibility for secondary research use. Technically, we have contributed to key areas including the robustness of AI systems against adversarial attacks, model interpretability, and the validation of AI tools through clinical observational studies.
Additionally, the Consortium has made substantial progress in the standardization of radiological procedures for image acquisition and analysis, and in the harmonization of methodologies for extracting reproducible imaging biomarkers.
From a legal and ethical standpoint, the project has helped define ethical standards for the use of health imaging data in the development of AI-based tools. Collectively, these efforts have resulted in resources that support faster training and more clinically relevant AI solutions for cancer care, while also contributing to policy evolution concerning the reuse of health data for research across Europe.
Ultimately, the project has played a key role in fostering trust in AI among healthcare professionals, patients, and stakeholders in both academia and industry.