Periodic Reporting for period 3 - ProCAncer-I (An AI Platform integrating imaging data and models, supporting precision care through prostate cancer’s continuum)
Période du rapport: 2023-10-01 au 2025-03-31
The ProCAncer-I project was conceived to develop a trustworthy, AI-powered platform for precision prostate cancer (PCa) care by integrating and analyzing large-scale imaging and clinical data. Its overarching goal was to create ProstateNet, the largest interoperable repository of multiparametric MRI (mpMRI) data related to PCa, paired with clinically relevant metadata. This initiative aimed to enable the development and validation of explainable, generalizable, and robust AI models that could support clinical decision-making throughout the continuum of PCa care.
The project tackled unmet clinical needs through nine clearly defined use cases, ranging from cancer detection to patient stratification and toxicity prediction. These use cases guided data selection, model development, and validation strategies. ProCAncer-I successfully harmonized retrospective and prospective data collection from 14,349 unique patients, 105,443 imaging series, and 9,114,562 images, while ensuring full compliance with GDPR and ethical standards.
The technical platform developed includes advanced anonymization pipelines, radiomics and deep learning (DL) pre-processing tools, the development of an AI model passport to document the model development parameters for traceability purposes, and a secure, cloud-based MLOps infrastructure. Models were developed using both vendor-specific and vendor-neutral approaches, with a strong focus on robustness and trustworthiness.
A major success of the project was the clinical validation of selected AI models in real-world diagnostic scenarios through multicenter reader studies. These demonstrated improved diagnostic sensitivity and specificity when AI support was used, even among less-experienced radiologists. Furthermore, segmentation models reached performance levels comparable to expert human annotations.
The project also produced substantial contributions to standards for medical imaging metadata, extending OMOP-CDM and DCAT-AP to suit radiology use cases. Sustainability has been planned via integration with the European Cancer Imaging Initiative (EUCAIM), ensuring long-term accessibility and reuse of the ProstateNet repository and associated AI tools.
The project tackled unmet clinical needs through nine clearly defined use cases, ranging from cancer detection to patient stratification and toxicity prediction. These use cases guided data selection, model development, and validation strategies. ProCAncer-I successfully harmonized retrospective and prospective data collection from 14,349 unique patients, 105,443 imaging series, and 9,114,562 images, while ensuring full compliance with GDPR and ethical standards.
The technical platform developed includes advanced anonymization pipelines, radiomics and deep learning (DL) pre-processing tools, the development of an AI model passport to document the model development parameters for traceability purposes, and a secure, cloud-based MLOps infrastructure. Models were developed using both vendor-specific and vendor-neutral approaches, with a strong focus on robustness and trustworthiness.
A major success of the project was the clinical validation of selected AI models in real-world diagnostic scenarios through multicenter reader studies. These demonstrated improved diagnostic sensitivity and specificity when AI support was used, even among less-experienced radiologists. Furthermore, segmentation models reached performance levels comparable to expert human annotations.
The project also produced substantial contributions to standards for medical imaging metadata, extending OMOP-CDM and DCAT-AP to suit radiology use cases. Sustainability has been planned via integration with the European Cancer Imaging Initiative (EUCAIM), ensuring long-term accessibility and reuse of the ProstateNet repository and associated AI tools.
Since its inception, the ProCAncer-I project has aimed to establish a robust, ethically compliant AI platform—ProstateNet—to revolutionize prostate cancer (PCa) care through data-driven decision support. During the implementation period (M1–M54), the consortium successfully orchestrated a multifaceted research and development program involving 36 partners across Europe.
The work began with the design of a data governance and ethical framework, securing approvals for retrospective and prospective data collection, leading to the creation of the ProstateNet image repository that maintains 14,349 unique patients, 105,443 imaging series, and 9,114,562 images. This dataset is among the largest of its kind worldwide and serves as the foundation for AI model development and/or the external validation of models already developed using other proprietary data.
Key technical developments that took place during the project implementation include, i) A standards-compliant cloud infrastructure with advanced data curation, anonymization, and data harmonization tools; ii) the development of AI models addressing distinct clinical use cases, spanning the prostate cancer care continuum, i.e. from the detection of prostate cancer with high accuracy both in peripheral and transitional zone (UC1), to the prediction of prediction of post radical prostatectomy and/or radiation-induced urinary toxicity (UC7); iii) the implementation of preprocessing pipelines and segmentation models to support AI workflows; and iv) a novel AI Model Passport system for traceability, explainability, and regulatory alignment.
Model validation was a central focus of our work. Clinical validation of selected models in a multicenter radiologist reader study demonstrated improved diagnostic performance and high user acceptance. Federated learning experiments confirmed the feasibility of decentralized, privacy-preserving model training.
In terms of dissemination, the project achieved significant visibility, with over 30 high-impact scientific publications; major presentations at ECR, EuSoMII, and AI for Health Imaging (AI4HI) cluster events; contributions to consensus guidelines like the “FUTURE-AI” framework for trustworthy AI, and active participation in standardization bodies such as OHDSI and W3C, with notable extensions to OMOP-CDM and DCAT-AP.
On the exploitation front, the project finalized a two-phase sustainability plan. In the short term, a FORTH-led four-year roadmap will maintain the infrastructure. For long-term sustainability, ProstateNet will be integrated into the EUCAIM platform, enabling ongoing reuse, and regulatory-grade AI development. A regulatory roadmap aligned with the EU MDR and AI Act ensures readiness for clinical deployment.
Overall, ProCAncer-I has developed a unique platform that represents a significant step toward advancing AI-enabled prostate cancer diagnostics and personalized care in Europe.
The work began with the design of a data governance and ethical framework, securing approvals for retrospective and prospective data collection, leading to the creation of the ProstateNet image repository that maintains 14,349 unique patients, 105,443 imaging series, and 9,114,562 images. This dataset is among the largest of its kind worldwide and serves as the foundation for AI model development and/or the external validation of models already developed using other proprietary data.
Key technical developments that took place during the project implementation include, i) A standards-compliant cloud infrastructure with advanced data curation, anonymization, and data harmonization tools; ii) the development of AI models addressing distinct clinical use cases, spanning the prostate cancer care continuum, i.e. from the detection of prostate cancer with high accuracy both in peripheral and transitional zone (UC1), to the prediction of prediction of post radical prostatectomy and/or radiation-induced urinary toxicity (UC7); iii) the implementation of preprocessing pipelines and segmentation models to support AI workflows; and iv) a novel AI Model Passport system for traceability, explainability, and regulatory alignment.
Model validation was a central focus of our work. Clinical validation of selected models in a multicenter radiologist reader study demonstrated improved diagnostic performance and high user acceptance. Federated learning experiments confirmed the feasibility of decentralized, privacy-preserving model training.
In terms of dissemination, the project achieved significant visibility, with over 30 high-impact scientific publications; major presentations at ECR, EuSoMII, and AI for Health Imaging (AI4HI) cluster events; contributions to consensus guidelines like the “FUTURE-AI” framework for trustworthy AI, and active participation in standardization bodies such as OHDSI and W3C, with notable extensions to OMOP-CDM and DCAT-AP.
On the exploitation front, the project finalized a two-phase sustainability plan. In the short term, a FORTH-led four-year roadmap will maintain the infrastructure. For long-term sustainability, ProstateNet will be integrated into the EUCAIM platform, enabling ongoing reuse, and regulatory-grade AI development. A regulatory roadmap aligned with the EU MDR and AI Act ensures readiness for clinical deployment.
Overall, ProCAncer-I has developed a unique platform that represents a significant step toward advancing AI-enabled prostate cancer diagnostics and personalized care in Europe.
The ProCAncer-I project has achieved significant advancements beyond the current state of the art in AI-driven medical imaging and prostate cancer (PCa) management. At its core, the project created ProstateNet, a high-quality, ethically sourced, GDPR-compliant dataset of over 1.5 million mpMRI images—by far the largest PCa imaging dataset in Europe—coupled with harmonized clinical data. This scale and standardization enable the training and validation of generalizable AI models with unprecedented robustness and transparency.
Technically, the project introduced novel contributions such as i) Development of AI models, including deep learning and radiomics-based approaches across nine clinically defined use cases; ii) Integration of innovative preprocessing tools and model uncertainty estimation techniques; iii) Implementation of self-supervised and multiple instance learning (MIL) strategies to address data scarcity, enabling effective model training even with limited annotated data; and iv) Establishment of the AI Model Passport for full lifecycle traceability and transparency, aligned with W3C standards and EU digital health regulations.
These innovations culminated in clinically validated tools shown to enhance radiologist performance, particularly in PCa detection and stratification tasks. The models demonstrated gains in sensitivity and specificity, with strong usability scores, confirming their readiness for deployment in real-world clinical environments.
Regarding the wider Socio-Economic and Societal Impacts of the project, ProCAncer-I directly addresses the societal burden of PCa, the second most common cancer among men, by aiming to reduce overdiagnosis and unnecessary treatments through precision AI tools. The provision of a secure processing environment and its privacy-preserving design ensures that sensitive health data remains secure while still enabling collaborative innovation through the use of ProstateNET.
Moreover, the project’s emphasis on trustworthiness and explainability—aligned with the FUTURE-AI principles—helps bridge the trust gap between AI technologies and clinicians, patients, and regulators. By grounding its impact in ethical and legal compliance, ProCAncer-I is positioned to contribute meaningfully to the digital transformation of healthcare in Europe.
Technically, the project introduced novel contributions such as i) Development of AI models, including deep learning and radiomics-based approaches across nine clinically defined use cases; ii) Integration of innovative preprocessing tools and model uncertainty estimation techniques; iii) Implementation of self-supervised and multiple instance learning (MIL) strategies to address data scarcity, enabling effective model training even with limited annotated data; and iv) Establishment of the AI Model Passport for full lifecycle traceability and transparency, aligned with W3C standards and EU digital health regulations.
These innovations culminated in clinically validated tools shown to enhance radiologist performance, particularly in PCa detection and stratification tasks. The models demonstrated gains in sensitivity and specificity, with strong usability scores, confirming their readiness for deployment in real-world clinical environments.
Regarding the wider Socio-Economic and Societal Impacts of the project, ProCAncer-I directly addresses the societal burden of PCa, the second most common cancer among men, by aiming to reduce overdiagnosis and unnecessary treatments through precision AI tools. The provision of a secure processing environment and its privacy-preserving design ensures that sensitive health data remains secure while still enabling collaborative innovation through the use of ProstateNET.
Moreover, the project’s emphasis on trustworthiness and explainability—aligned with the FUTURE-AI principles—helps bridge the trust gap between AI technologies and clinicians, patients, and regulators. By grounding its impact in ethical and legal compliance, ProCAncer-I is positioned to contribute meaningfully to the digital transformation of healthcare in Europe.