Table 1 The key dimensions of the infrastructure developed for each one of the five AI4HI projects
From: Data infrastructures for AI in medical imaging: a report on the experiences of five EU projects
CHAIMELEON | EuCanImage | INCISIVE | ProCAncer-I | PRIMAGE | |
|---|---|---|---|---|---|
Cancer types | Lung Colorectal Breast Prostate | Colorectal Liver Breast | Lung Colorectal Breast Prostate | Prostate | Neuroblastoma Diffuse intrinsic pontine glioma |
Architecture | Hybrid | Accommodating both decentralized and centralized storage | Hybrid (federated and centralized storage) | Centralized | Centralized |
Data models and types of data | DICOM-MIABIS OMOP CDM (terminology IDs mainly) Structure of the eCRF | DICOM-MIABIS FHIR (and terminologies supported by FHIR + extensions) | FHIR SNOMED-CT LOINC DICOM | DICOM-RT OMOP CDM with extensions | DICOM-MIABIS OMOP CDM |
Deidentification process | Pseudonymized initially for curation and then fully anonymized data at the central repository | Pseudonymized data | Pseudonymized data | Fully anonymized data | Pseudonymized data |
Curation tools | Data completeness and consistency tools, image quality checking, Image anonymization, annotation, segmentation and harmonization | Image anonymization/pseudonymization, quality control and annotation, non-imaging data anonymization and homogenization | Image de-identification tools, quality control, and annotation tools | Image quality control, anonymization, motion-correction, co-registration & annotation. | Image labelling, quality checking, annotation, denoising, motion correction, registration |
Number of potential subjects | 13,000 full cases (images + clinical data), 34,000 image only cases | 25,000 | 8,850 | 17,115 | 1,500 |
Number and type of clinical sites | 8 university hospitals | 6 hospitals (including 4 university hospitals) | 9 data providers (5 universities associated with 5 external hospitals, 2 hospitals, 1 association, 1 company) | 13 data providers | Clinical trials: HR-NBL1, LINES and Society of Paediatric Oncology and Haematology. 16 external hospitals. |
Location of the central repository, when applicable | Universitat Politècnica de València, Spain | European Institute for Biomedical Imaging Research/EIBIR | MAGGIOLI, Greece | FORTH-ICS, Greece | Microsoft Azure, Dublin, Ireland & Universitat Politècnica de València, Spain |
Level of completion (maturity) of the project | +8,500 patients screened | Data deposition & annotation started end of 2022. | Retrospective data collection and annotation finished, prospective data collection finishes end 2023, storage platform almost ready storing all available data so far. The project finishes in March 2024 | Retrospective data collection almost finished. Platform almost ready. AI model development started. | Platform ready. Closing data incorporation (cases for external validation) and AI models tunning and testing. The project finishes in May 2023 |
Involvement of industry | SMEs: Medexprim, Quibim, Bahia Large corporation: General Electric | SMEs: Collective Minds Radiology, Lynkeus, Radiomics. Large corporation: Siemens | SMEs: European Dynamics, Telesto, Squaredev, ADAPTIT, VISARIS, Timelex, White Research Large corporations: MAGGIOLI, Medtronic | SMEs: B3D, Advantis, Quibim Large corporation: Quiron | SMEs: Quibim, Medexprim, Chemotargets. |