Periodic Reporting for period 2 - CHAIMELEON (Accelerating the lab to market transition of AI tools for cancer management)
Période du rapport: 2022-03-01 au 2023-08-31
There are 3.7 million new cancer cases in Europe every year, that require the development of cancer management tools to improve current cancer treatment procedures.
In this context, the use of Artificial Intelligence (AI) on health data is currently generating promising tools with a vast potential to become useful in different areas of clinical application.
However, the development of imaging-based AI tools relies on the availability of large, quality-controlled datasets, which still remains a major challenge. This is because the generation of imaging biobanks is a resource-intensive endeavor, facing multiple technical and operational difficulties such as image and data harmonization, curation and annotation, and various legal and ethical restrictions. As a result, the quantity, quality, and representativeness of datasets still remain major limiting factors in the development of cancer management tools.
To address the lack of data availability, the CHAIMELEON project aims to set-up a cancer imaging repository facilitating access to large, high-quality sets of anonymized data. This will be achieved through the creation of a distributed data repository that will be made interoperable with other existing biobanks, enabling secure share and reuse of data as a single-access point resource for the community of AI developers.
Since imaging datasets contain images acquired at different centers (cross-vendor/cross-institution datasets), quantitative image features extracted from images may not be reproducible from once center to another. To ensure the reproducibility of quantitative imaging biomarkers (QIB) and allow scientific reuse of retrospective imaging data from multicenter acquisitions, CHAIMELEON will test image harmonization protocols as one of its main objectives. The project will involve the setup of the IT infrastructure, the creation of protocols for legal compliance, the development of tools for data ingestion and curation and processing pipelines for imaging data annotation and harmonization, as well as methodologies and tools for enhanced interpretability of AI models among others. Since the repository will be targeted to AI developers as end-users, AI-powered pipelines for data processing will be implemented as another main objective to the project, ultimately enhancing the interpretability of the tested AI models.
In this context, the use of Artificial Intelligence (AI) on health data is currently generating promising tools with a vast potential to become useful in different areas of clinical application.
However, the development of imaging-based AI tools relies on the availability of large, quality-controlled datasets, which still remains a major challenge. This is because the generation of imaging biobanks is a resource-intensive endeavor, facing multiple technical and operational difficulties such as image and data harmonization, curation and annotation, and various legal and ethical restrictions. As a result, the quantity, quality, and representativeness of datasets still remain major limiting factors in the development of cancer management tools.
To address the lack of data availability, the CHAIMELEON project aims to set-up a cancer imaging repository facilitating access to large, high-quality sets of anonymized data. This will be achieved through the creation of a distributed data repository that will be made interoperable with other existing biobanks, enabling secure share and reuse of data as a single-access point resource for the community of AI developers.
Since imaging datasets contain images acquired at different centers (cross-vendor/cross-institution datasets), quantitative image features extracted from images may not be reproducible from once center to another. To ensure the reproducibility of quantitative imaging biomarkers (QIB) and allow scientific reuse of retrospective imaging data from multicenter acquisitions, CHAIMELEON will test image harmonization protocols as one of its main objectives. The project will involve the setup of the IT infrastructure, the creation of protocols for legal compliance, the development of tools for data ingestion and curation and processing pipelines for imaging data annotation and harmonization, as well as methodologies and tools for enhanced interpretability of AI models among others. Since the repository will be targeted to AI developers as end-users, AI-powered pipelines for data processing will be implemented as another main objective to the project, ultimately enhancing the interpretability of the tested AI models.
During the first 3 years of the project, the repository infrastructure and main functionalities have been deployed, leading to the release of the final repository prototype which can now be accessed via web interface. Importantly, the repository has been subjected to a thorough validation by AI developers, clinicians as well as legal experts within the Consortium. Currently, an external validation phase – articulated through an Open Challenge – is ongoing and welcomes the worldwide AI community to train and test their models in the CHAIMELEON platform.
In parallel, hospitals continue to populate the repository with cancer imaging and clinical data. This involves: (1) The screening of patients according to eligibility criteria, (2) creation and completion of the Clinical Report Form (eCRF), and (3) export of the locally collected data to the central repository. Importantly, this process was facilitated by technical partners, who developed and implemented the tools to allow data extraction, anonymization, and curation at hospitals prior to its incorporation into the repository.
At the same time, AI developers designed tools to facilitate data analysis within the CHAIMELEON platform. Such tools include: (1) image harmonization approaches, (2) AI security algorithms to protect data integrity, (3) a dataset selection engine to enable users to create their own datasets, (4) data privacy mechanisms via distributed learning, data augmentation, and privacy preserving models, and (5) tools that will help explain the decisions of medical AI to generate trust amongst clinicians and measure a model’s reliability.
All the work done so far has been guided by legal experts in data protection, that have implemented the legal operational model for the repository, data management plan, data protection impact assessment, and data sharing agreement. The first complete verification of the prototype’s compliance with the GDPR was also completed.
Meanwhile, a preliminary plan for the repository’s long-term sustainability and business model has been designed. Lastly, an ambitious project dissemination and communication plan was implemented, leading to the release of 48 project publications and presentations at 33 scientific events, effectively disseminating the project’s results to the wider public.
In parallel, hospitals continue to populate the repository with cancer imaging and clinical data. This involves: (1) The screening of patients according to eligibility criteria, (2) creation and completion of the Clinical Report Form (eCRF), and (3) export of the locally collected data to the central repository. Importantly, this process was facilitated by technical partners, who developed and implemented the tools to allow data extraction, anonymization, and curation at hospitals prior to its incorporation into the repository.
At the same time, AI developers designed tools to facilitate data analysis within the CHAIMELEON platform. Such tools include: (1) image harmonization approaches, (2) AI security algorithms to protect data integrity, (3) a dataset selection engine to enable users to create their own datasets, (4) data privacy mechanisms via distributed learning, data augmentation, and privacy preserving models, and (5) tools that will help explain the decisions of medical AI to generate trust amongst clinicians and measure a model’s reliability.
All the work done so far has been guided by legal experts in data protection, that have implemented the legal operational model for the repository, data management plan, data protection impact assessment, and data sharing agreement. The first complete verification of the prototype’s compliance with the GDPR was also completed.
Meanwhile, a preliminary plan for the repository’s long-term sustainability and business model has been designed. Lastly, an ambitious project dissemination and communication plan was implemented, leading to the release of 48 project publications and presentations at 33 scientific events, effectively disseminating the project’s results to the wider public.
The CHAIMELEON repository along with its related AI-powered tools are being designed to impact the management of the four most prevalent types of cancer worldwide. Due to the social and economic burden these imply, we expect the outcome of this project to have an EU-wide impact at both levels.
Upon successful validation, we expect the repository infrastructure, legal operational model, analysis tools and web-based user interfaces to be adapted to the management of other types of cancer.
The project will contribute to the current state of the art of AI for health imaging by defining a framework for legitimate access to anonymized imaging and related clinical data, making these more openly accessible across the EU for secondary use in research. At the technical level, we hope to contribute to the advancement in robustness of AI systems against malicious attacks, interpretability of AI-based models, and validation of AI tools in clinical observational studies. We also aim to impact the fields of standardization of radiological procedures for image acquisition and analysis, as well as harmonization for extraction of reproducible imaging biomarkers.
From a legal perspective, this project will contribute to the creation of ethical standards for the use of health imaging data in the context of AI tool development. All in all, we expect CHAIMELEON to generate resources that facilitate faster and more successful development of AI-based solutions for cancer management, while promoting actions to foster the evolution of European laws in the reuse of health data for research purposes. By doing so, we expect to increase trust in AI solutions amongst healthcare professionals, patients, and stakeholders in both industry and academia.
Upon successful validation, we expect the repository infrastructure, legal operational model, analysis tools and web-based user interfaces to be adapted to the management of other types of cancer.
The project will contribute to the current state of the art of AI for health imaging by defining a framework for legitimate access to anonymized imaging and related clinical data, making these more openly accessible across the EU for secondary use in research. At the technical level, we hope to contribute to the advancement in robustness of AI systems against malicious attacks, interpretability of AI-based models, and validation of AI tools in clinical observational studies. We also aim to impact the fields of standardization of radiological procedures for image acquisition and analysis, as well as harmonization for extraction of reproducible imaging biomarkers.
From a legal perspective, this project will contribute to the creation of ethical standards for the use of health imaging data in the context of AI tool development. All in all, we expect CHAIMELEON to generate resources that facilitate faster and more successful development of AI-based solutions for cancer management, while promoting actions to foster the evolution of European laws in the reuse of health data for research purposes. By doing so, we expect to increase trust in AI solutions amongst healthcare professionals, patients, and stakeholders in both industry and academia.