Periodic Reporting for period 1 - FoundRa (Foundation Model-Based Biomarker Platform for Radiology and Clinical Trials)
Okres sprawozdawczy: 2024-11-01 do 2025-10-31
The FoundRa project is based on 2 challenges: (1) Due to aging and population growth, there is an increase in pathologies that require radiology, from diagnosis to follow-up. (2) The ambition to make personalized medicine the standard of the future requires all kinds of biomarkers.
Raidium is an agile, young and ambitious French company, born from the mission of the co-founders to accelerate the application of precision medicine and relieve radiologists from their enormous workload. Raidium delivers the first scalable interactive AI platform for radiology, supporting radiologists in their work, enabling a broader group of clinicians to benefit from radiology insights, and making clinical trials faster and more cost-effective. The innovation is a multimodal foundation model (FM) that uses self-supervised learning based on unlabeled data, creating scalable biomarkers; unlike current offerings where biomarker-generating AI can only target one segment, due to the time- and cost-consuming effort of supervised learning based on manually labeled data. The human-centric platform offers an intelligent interactive and intuitive human-machine interface that can be trained by the radiologist, leveraging their experience while performing automatic segmentation, annotation and analytical tasks, providing reports and decision support.
This initiative aims to revolutionize and bring about significant advances in the fields of precision medicine in respect to medical practice and clinical trials.
Raidium is an agile, young and ambitious French company, born from the mission of the co-founders to accelerate the application of precision medicine and relieve radiologists from their enormous workload. Raidium delivers the first scalable interactive AI platform for radiology, supporting radiologists in their work, enabling a broader group of clinicians to benefit from radiology insights, and making clinical trials faster and more cost-effective. The innovation is a multimodal foundation model (FM) that uses self-supervised learning based on unlabeled data, creating scalable biomarkers; unlike current offerings where biomarker-generating AI can only target one segment, due to the time- and cost-consuming effort of supervised learning based on manually labeled data. The human-centric platform offers an intelligent interactive and intuitive human-machine interface that can be trained by the radiologist, leveraging their experience while performing automatic segmentation, annotation and analytical tasks, providing reports and decision support.
This initiative aims to revolutionize and bring about significant advances in the fields of precision medicine in respect to medical practice and clinical trials.
The activities performed during this period focused on advancing the technical and scientific capabilities of the FoundRa project, specifically through the development of large-scale foundation models (FMs), automated biomarker extraction, and sophisticated cohort management tools.
1. Development of the Curia Foundation Model
A major achievement was the release of Curia, a multimodal radiological foundation model.
● Large-Scale Training: The model was trained using self-supervised learning on over 200 million CT and MRI images, representing 130 TB of anonymized clinical data.
● Technical Specifications: Curia utilizes a Vision Transformer (ViT) backbone and the DINOv2 algorithm. It is available in "Base" (86M parameters) and "Large" (300M parameters) configurations.
● Proven Performance: The model was evaluated against CuriaBench, a benchmark of 19 tasks across six clinical domains. It achieved state-of-the-art (SotA) results, consistently outperforming existing models like BiomedCLIP and MedImageInsight.
2. Advanced Imaging and Biomarkers
Research focused on quantifying advanced biomarkers through automated analysis:
● The Liver-to-Spleen Volume Ratio (LSVR) was developed as a predictor for liver fibrosis. Validation on internal cohorts confirmed its ability to discriminate between early and late-stage fibrosis.
● Portal Hypertension Prediction: In collaboration with AP-HP Beaujon Hospital, the team developed automated non-invasive tests for predicting Hypertension Venous-Portal Gradient (HVPG). Raidium’s models emerged as the top-performing fully automated method for this task.
● Oncology biomarkers: the team developed a complete panel of advanced biomarkers for oncology workflows.
3. Platform Technical Infrastructure
New components were integrated into the Raidium Viewer to support clinical research workflows:
● Cohort Builder: A new component allows researchers to form specific patient cohorts for analysis.
● Cohort-Level Biomarker Application: The platform's allows to handle large scale studies
● External Tool Integration: The system supports the integration of third-party biomarkers through API interfaces.
1. Development of the Curia Foundation Model
A major achievement was the release of Curia, a multimodal radiological foundation model.
● Large-Scale Training: The model was trained using self-supervised learning on over 200 million CT and MRI images, representing 130 TB of anonymized clinical data.
● Technical Specifications: Curia utilizes a Vision Transformer (ViT) backbone and the DINOv2 algorithm. It is available in "Base" (86M parameters) and "Large" (300M parameters) configurations.
● Proven Performance: The model was evaluated against CuriaBench, a benchmark of 19 tasks across six clinical domains. It achieved state-of-the-art (SotA) results, consistently outperforming existing models like BiomedCLIP and MedImageInsight.
2. Advanced Imaging and Biomarkers
Research focused on quantifying advanced biomarkers through automated analysis:
● The Liver-to-Spleen Volume Ratio (LSVR) was developed as a predictor for liver fibrosis. Validation on internal cohorts confirmed its ability to discriminate between early and late-stage fibrosis.
● Portal Hypertension Prediction: In collaboration with AP-HP Beaujon Hospital, the team developed automated non-invasive tests for predicting Hypertension Venous-Portal Gradient (HVPG). Raidium’s models emerged as the top-performing fully automated method for this task.
● Oncology biomarkers: the team developed a complete panel of advanced biomarkers for oncology workflows.
3. Platform Technical Infrastructure
New components were integrated into the Raidium Viewer to support clinical research workflows:
● Cohort Builder: A new component allows researchers to form specific patient cohorts for analysis.
● Cohort-Level Biomarker Application: The platform's allows to handle large scale studies
● External Tool Integration: The system supports the integration of third-party biomarkers through API interfaces.
The impact of the results (the Curia Model and the Viewer product) are three fold:
● Scientific Impact: Advancing the field of precision medicine through quantitative 3D imaging and establishing a blueprint for sovereign European AI that upholds data protection standards while being global state of the art.
● Economic Impact: Significant reduction in specialist workload through AI-assisted functionalities, freeing time for more and deeper examinations.
● Societal Impact: Contribution to improved patient care through providing standardized, reproducible biomarkers for early diagnosis and treatment monitoring. Augmented access to care through improvement and harmonisation of average radiologist read level, as well as increased radiologist bandwidth.
Key needs include:
Further Research: integrating advanced vision-language modeling and agentic capabilities for more comprehensive clinical utility.
Demonstration: scaling pilots across diverse clinical environments to validate real-world stability and interoperability with various PACS vendors.
Regulatory & Standardisation: achieving mandatory MDR and FDA certifications for commercialization. Advocating for the inclusion of AI solutions in official clinical guidelines to augment long term uptake and pave the way towards reimbursement.
Commercialisation: Development of structured reimbursement strategies and economic models for AI-driven quantitative diagnostics within healthcare systems.
● Scientific Impact: Advancing the field of precision medicine through quantitative 3D imaging and establishing a blueprint for sovereign European AI that upholds data protection standards while being global state of the art.
● Economic Impact: Significant reduction in specialist workload through AI-assisted functionalities, freeing time for more and deeper examinations.
● Societal Impact: Contribution to improved patient care through providing standardized, reproducible biomarkers for early diagnosis and treatment monitoring. Augmented access to care through improvement and harmonisation of average radiologist read level, as well as increased radiologist bandwidth.
Key needs include:
Further Research: integrating advanced vision-language modeling and agentic capabilities for more comprehensive clinical utility.
Demonstration: scaling pilots across diverse clinical environments to validate real-world stability and interoperability with various PACS vendors.
Regulatory & Standardisation: achieving mandatory MDR and FDA certifications for commercialization. Advocating for the inclusion of AI solutions in official clinical guidelines to augment long term uptake and pave the way towards reimbursement.
Commercialisation: Development of structured reimbursement strategies and economic models for AI-driven quantitative diagnostics within healthcare systems.