Challenge 3

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

Team: 

Vicomtech (Fundación Centro de Tecnologías de Interacción Visual y Comunicaciones Vicomtech), Spain

Proposal title: 

LUSYN (Lung Cancer Synthetic Data Generation)

Organisation

Vicomtech is a private non-profit technology centre in San Sebastián, Spain, and a member of the Basque Research and Technology Alliance (BRTA). It specialises in artificial intelligence, visual computing and interaction, and transfers applied research to companies, including medical-imaging AI from EU projects such as LUCIA.

Project info

LUSYN builds a shared multimodal patient representation that fuses clinical variables and CT imaging, so every synthetic output stays consistent with a single patient. It combines a 3D CT encoder initialised from Vicomtech’s lung-CT weights, pretrained in the EU LUCIA project, with a clinical-data encoder. Because the 1,088-subject CHAIMELEON cohort is too small to learn multimodal alignment alone, external lung-CT datasets such as NLST are mapped to a harmonised set of clinical and demographic variables to pretrain the encoders before adaptation.

Three modules reuse this representation. An XGeM-style latent diffusion model generates conditional 3D CT and clinical data to augment groups under-represented by sex, metastasis site and stage, imputes missing clinical variables and identifies the variables that drive generation. An image-to-image generator, using unpaired methods such as 3D CycleGAN, synthesises missing non-contrast, contrast-enhanced or low-dose acquisitions. A domain-mapping model harmonises CTs acquired at different kVp settings while preserving radiomic features.

Training runs on CINECA’s Leonardo supercomputer and ADA Cloud, keeping patient data inside the secure processing environment. The modules can be deployed separately, and the team targets all four official challenge KPIs.

Details

Oncology AI models are often trained on incomplete, imbalanced datasets that under-represent patients by sex, metastasis site or stage, and that mix CT scans from different vendors and protocols. This limits robustness, subgroup validation and equitable performance in the clinic.

LUSYN, developed by Vicomtech, addresses these problems in a single framework built around a shared representation of each patient’s CT images and clinical data. Starting from lung-CT encoders already pretrained in the European LUCIA project, and pretraining the clinical branch on external cohorts mapped to a common variable set, the framework drives three tasks: generating balanced synthetic cohorts and imputing missing clinical variables with a latent diffusion model, synthesising missing CT acquisitions, and harmonising scans acquired at different tube voltages. Working from one patient representation, rather than separate tools for each task, keeps generated images and clinical variables consistent with each other.

The framework starts at TRL 3, with its imaging components validated separately. Each new module will be validated on the CHAIMELEON cohort to reach TRL 4, and the integrated workflow will run end-to-end inside the secure processing environment to reach TRL 5, measured against the four challenge KPIs: data fidelity, at least a 50% reduction in subgroup imbalance, data completion and harmonisation, and radiomic robustness under noise. LUSYN is designed to be deployed module by module inside EUCAIM, with Vicomtech offering data-harmonisation and benchmark-generation services to European hospitals and AI developers, and to extend to other cancer cohorts such as breast, prostate and colorectal.