Transforming a clinical study database into a structured database adapted to artificial intelligence applications

Thibault Sauron; Carole Lazarus; Camille Kurtz; Florence Cloppet; Isabelle Thomassin Naggara; EURAD Study Group; Laure Fournier

doi:10.1186/s13244-025-02087-2

Transforming a clinical study database into a structured database adapted to artificial intelligence applications

Insights Imaging. 2026 Feb 16;17(1):43. doi: 10.1186/s13244-025-02087-2.

Authors

Thibault Sauron¹, Carole Lazarus², Camille Kurtz¹, Florence Cloppet¹, Isabelle Thomassin Naggara³; EURAD Study Group; Laure Fournier⁴

Collaborators

EURAD Study Group:
Edouard Poncelet, Aurelie Jalaguier-Coudray, Ingrid Millet, Valerie Juhan, Corinne Balleyguier, Caroline Malhaire, Nicolas Perrot, Marc Bazot, Patrice Taourel, Emile Darai

Affiliations

¹ LIPADE, Université Paris Cité, Paris, France.
² Philips Research France, Paris, France.
³ Sorbonne Université, AP-HP, Department of Diagnostic and Interventional Imaging, Hôpital Tenon, Paris, France.
⁴ Université Paris Cité, AP-HP, Department of Radiology, Hôpital Européen Georges Pompidou, PARCC UMRS 970, INSERM, Paris, France. Laure.fournier@aphp.fr.

Abstract

Objective: Medical imaging databases suitable for training machine learning/computer vision algorithms are scarce, limiting the potential for development and generalisation of clinical tools. Clinical trial databases are a source of data, known for their high-quality data and reliable annotations. However, they are not tailored to the needs of machine learning or deep learning models. Our objective was to develop a methodology and tools that enable the curation of these databases specifically for the training or testing of artificial intelligence tools.

Materials and methods: MRIs from the French centres of the EURAD clinical trial (MRI of women with pelvic adnexal lesions) were used to constitute the database. We developed the steps required to curate a clinical trial database: definition of inclusion and exclusion criteria, removal of unnecessary data according to the principle of parsimony, quality control, and harmonisation.

Results: A total of 713 patients were included in our study. The directory structure was simplified, and the number of files and folders decreased by 44% and 95% respectively. Only 62 DICOM fields were considered necessary for artificial intelligence (AI) model applications. Quality control was implemented in repeated cycles of automatic checks, followed by a final manual random inspection. Finally, sequence names were harmonised for easy identification when developing models.

Conclusion: Using a clinical trial database, we propose a methodology to build a database suitable to train or test AI algorithms. This study underlines the need for a more global and systematic framework for the secondary use of health data to develop AI imaging tools for patient care.

Critical relevance statement: We propose and detail a framework and tools to curate a clinical trial database to allow secondary use of the high-quality annotated data generated in clinical trials for the training and testing of artificial intelligence models.

Key points: Clinical trial imaging databases are not adapted for AI model development. A curation process of MRI databases was developed for machine learning applications. We share the open-source tools and methodology developed in this study.

Keywords: Artificial intelligence; Clinical trial; Data curation; MRI; Medical computer vision.