Challenge 3

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

Team: 

Sentigrate BV, Belgium

Proposal title: 

DUET (DUal-modality Engine for Trustworthy synthetic cohorts)

Organisation

Sentigrate BV is a Belgian data-science and data-engineering SME, founded in 2016 in Leuven. It delivers production AI systems for customers in more than 16 countries, with particular depth in health, hospital data and medical devices, including privacy-preserving synthetic clinical data and MDR Class IIa imaging software.

Project info

DUET generates coherent synthetic subjects in which the thorax CT and the 29-variable clinical record describe the same patient, so a subject’s stage, metastasis pattern and smoking status match the accompanying scan. One shared latent per subject conditions both branches through bidirectional cross-attention.

Five modules run inside the Secure Processing Environment. A driver module uses gradient-boosted surrogates with SHAP to identify which variables drive cohort structure. A tabular module generates mixed-type clinical data and models event times and censoring jointly, making the cohort usable for survival research; diffusion is benchmarked against PrivSyn and MICE/missForest rather than assumed to win. An imaging module fine-tunes released pretrained CT weights with rectified-flow latent generation and anatomical masks, needing about 30 sampling steps instead of 1,000. A series-completion module translates between available series, with a kVp-conditioned variant for harmonisation that runs alongside ComBat and is selected per radiomic feature on measured concordance. An audit gate checks every cohort.

Targets exceed the challenge thresholds, for example a 70% cut in subgroup imbalance, and add privacy (membership-inference AUC of 0.55 or less), blinded realism and survival-model checks.

Details

EUCAIM’s lung-cancer cohorts are large but incomplete, imbalanced and heterogeneously acquired. Today, building a balanced cohort means dropping incomplete subjects, imputing variables one by one, or generating tables and images with unrelated models, which yields synthetic patients whose CT contradicts their record.

DUET, developed by Sentigrate, generates synthetic subjects as single, internally consistent units: a thorax CT volume and a full oncology record driven by one shared latent. The engine also completes missing values and CT series, harmonises scans acquired at different tube voltages, and generates event times and censoring together with clinical covariates so that the synthetic cohort supports prognostic research. Harmonisation is decided by measurement: generative translation and the established ComBat method run side by side, and the better one is kept for each radiomic feature. Efficiency is a design goal: the imaging branch fine-tunes released CT weights with rectified flow instead of training from scratch, and inference is quantised to run on a single-GPU EUCAIM node.

Sentigrate brings experience in privacy-preserving synthetic clinical data, generative modelling and MDR Class IIa medical-imaging software, and will take the combined system from TRL 3–4 to TRL 5 on the 1,088-study CHAIMELEON cohort. It targets results above every challenge threshold: fidelity on at least 90% of variables, a 70% reduction in subgroup imbalance, 98% imaging completion and median radiomic concordance of 0.88 under noise. Four further KPIs guard the result: membership-inference AUC of 0.55 or less, expert readers telling synthetic from real no better than 60% of the time, survival-model accuracy within 0.02 C-index of real data, and a full cohort generated in under 24 GPU-hours.