Challenge 3

Generative AI for Enhancement of Clinical Datasets_EUropean Federation for CAncer IMages (EUCAIM)

Team: 

Idara research group — IIS Galicia Sur (Galicia Sur Health Research Institute), Spain

Proposal title: 

SynCohort: Generative AI for synthetic cohorts, gender-bias reduction and completion of clinical and CT imaging datasets (EUCAIM)

Organisation

Idara (Inteligencia Digestiva y Salud) is the AI research group of the Digestive Medicine Service at the University Hospital Complex of Pontevedra (CHUP), part of the ISCIII-accredited Galicia Sur Health Research Institute, Spain. It runs nationally funded projects on health-data spaces and synthetic clinical data.

Project info

SynCohort will generate coherent synthetic lung-cancer patients that link clinical records with thorax CT, working entirely inside the AI-BOOST/EUCAIM secure environment on the challenge’s 1,088 CT studies. One pipeline addresses four problems: subgroup imbalance, with gender as the priority; missing clinical values; missing CT acquisitions; and heterogeneous kVp protocols.

Clinical records are generated with conditional tabular models (CTGAN/TVAE or tabular diffusion), which also support targeted sampling to rebalance subgroups and impute missing variables. For imaging, the team extends the pretrained NV-Generate-CT (MAISI) 3D latent-diffusion model with clinical and acquisition conditioning through ControlNet modules, using a patient’s available CT to preserve anatomy when completing missing non-contrast, contrast-enhanced or low-dose series. Harmonisation models conditioned on kVp and DICOM parameters reduce intensity differences between protocols, with differentiable texture losses tested to keep radiomic biomarkers stable and ComBat as a comparison baseline.

A sensitivity-analysis tool lets clinicians inspect how demographic, clinical, imaging and acquisition inputs drive the generated outputs. Targets include halving the largest-to-smallest subgroup ratio, at least 90% accuracy on masked categorical values and median GLCM Joint Entropy concordance of at least 0.85 under noise.

Details

Clinical AI is often trained and tested on cohorts that under-represent some patient groups, and real imaging datasets are incomplete and acquired with different protocols. The result is uncertainty about how models perform for women and other low-prevalence subgroups, which matters for the safety and equity of care.

SynCohort, developed by the Idara group at IIS Galicia Sur, is a generative pipeline that produces synthetic oncology patients whose clinical records and CT scans belong together. It rebalances subgroups with gender as the priority, fills in missing clinical values and CT series, and harmonises scans acquired at different tube voltages while checking that radiomic biomarkers are preserved. The method combines conditional tabular generators with a pretrained 3D CT diffusion model extended with clinical and acquisition conditioning, plus an auditable tool that shows clinicians which input variables drive the outputs. All work runs inside the EUCAIM secure processing environment, and the Challenge Owner will test the enhanced dataset with its own multimodal survival model.

The team builds on Aidatamed, its operational synthetic health-data infrastructure, which has already achieved a fidelity score of 0.93, train-on-synthetic utility ratios of 0.94–0.97 and low re-identification risk on clinical data. The project will reach TRL 5 at mid-term and finish with a prototype demonstrated inside EUCAIM. Targets include at least a 50% reduction in subgroup imbalance, at least 90% accuracy and 20% lower error than median imputation when reconstructing masked clinical values, 95% completion of masked CT acquisitions, and a 30% reduction in cross-protocol intensity differences while preserving radiomic features.