Challenge 3
Generative AI for Enhancement of Clinical Datasets_EUropean Federation for CAncer IMages (EUCAIM)
Team:
UCBM – ARCO: Università Campus Bio-Medico di Roma (coordinator), Eskişehir Osmangazi University (ESOGU) and Saint Camillus International University of Health Sciences (UniCamillus)
Proposal title:
REGEN — REsilient GENerative Multimodal AI for Clinical Data
Organisation
The consortium is coordinated by Università Campus Bio-Medico di Roma (UCBM), Italy, with Eskişehir Osmangazi University (ESOGU), Türkiye, and Saint Camillus International University of Health Sciences (UniCamillus), Italy. It combines expertise in resilient AI, multimodal learning, generative medical imaging, clinical AI and scalable deep-learning engineering.
Project info
REGEN replaces separate imputers, generators and harmonisers with one resilient multimodal framework that builds each patient’s representation only from the information actually available, so errors do not propagate between sequential steps.
Clinical variables are encoded by NAIM, a transformer that handles missing features without imputation, and each available CT acquisition by the frozen MONAI MAISI autoencoder. Missing-aware vision-tabular attention fuses both into a shared patient state. This state conditions two Latent Flow Matching branches: a tabular branch for clinical completion and synthetic-patient generation, and a 3D branch in MAISI latent space for missing-CT reconstruction, kVp harmonisation and synthetic CT generation. Task and target-modality tokens select the job, kVp is injected as a continuous condition, and available CT latents supply patient-specific anatomy. Because the state can be built from any subset of data, the same model handles both completion and targeted cohort balancing. Generated outputs are re-encoded to check consistency with the conditioning patient.
Beyond the official KPIs, the team targets a lung-mask Dice score of at least 0.95 on reconstructed CT and non-inferior performance of the Challenge Owner’s survival model on enhanced cohorts.
Details
Real-world clinical imaging cohorts are incomplete, heterogeneous and imbalanced: CT studies lack acquisitions, records miss variables, and protocols differ between patients. Existing pipelines treat each problem with a separate tool, so errors in one step carry into the next and the modalities are never coordinated.
REGEN, developed by a consortium led by Università Campus Bio-Medico di Roma with Eskişehir Osmangazi University and UniCamillus, is a unified generative framework for clinical tables and 3D CTs. It first builds a resilient patient representation from observed evidence only, using missing-aware encoders the team has already developed and validated, including on lung-cancer cohorts. That representation then drives latent flow-matching models that complete missing clinical variables, reconstruct missing CT series, harmonise scans across tube voltages and generate new synthetic patients matching specified subgroup conditions. Working in the latent space of a frozen, pretrained CT autoencoder keeps computation and energy use down.
Starting from TRL 2–3, the framework will be validated in stages and demonstrated as a containerised, reproducible pipeline in the CINECA/EUCAIM secure processing environment at TRL 5. The team commits to all official KPIs, including at least a 50% reduction in demographic imbalance, at least 95% completion of missing CT series and at least a 30% reduction in cross-kVp intensity differences while preserving radiomic features. It will also check anatomical accuracy, consistency between images and records, downstream survival-model performance and the absence of memorised training samples. The modular design allows new variables, protocols and cancer datasets to be added without rebuilding the system.
