Constraint-aware generative AI for balanced, complete and harmonised cancer imaging cohorts

Cancer AI development is constrained by incomplete, heterogeneous and imbalanced clinical and imaging cohorts, limiting representativeness, fairness and downstream research reliability. Our proposed solution directly addresses this challenge by creating more complete and balanced multimodal cohorts while preserving clinical and imaging consistency.

We already have a working HTTPS demonstrator that generates mixed-type clinical cohorts, reconstructs masked fields, supports controlled balancing and processes lung CT data through DICOM inventory, HU/geometry QC, reconstruction and harmonisation checks. This is engineering feasibility evidence, not clinical validation.

During Advance, the prototype will be extended and validated inside the SPE to: generate realistic synthetic clinical profiles; create complete new synthetic patients with corresponding CT studies for under-represented groups; reconstruct artificially withheld clinical variables and CT acquisitions for existing patients; and reduce acquisition-related image variability while protecting radiomic properties. A fail-closed 29-field adapter and subgroup, anatomy, HU, radiomics, diversity and copy safeguards will keep uncertainty and human review explicit.

Our differentiation is not a single generative architecture but a failure-visible multimodal qualification system. Clinical synthesis, subgroup balancing, new-patient CT generation, existing-patient acquisition completion, harmonisation, quantitative robustness and memorisation safeguards are evaluated as distinct non-substitutable capabilities. The same platform exposes which variable drives generation, which cohort intervention changed balance, which imaging gate failed and whether post-processing improved one KPI at the expense of another. This creates an auditable route from synthetic generation to production ready evidence rather than an opaque collection of plausible synthetic samples.

The prototype combines synthetic cohort generation with configurable, human-governed dependency experimentation. Predefined drivers, subgroup targets and dependency hypotheses can be compared using marginal, joint, conditional and pre/post-repair aggregate measurements. Fidelity and balancing remain separate objectives; interventions, shortages and trade-offs are visible.

The system treats semantic uncertainty as a computable stop condition. In an artificial stage-metastasis investigation it identified and quantified the only technically triggered repair predicate, but did not reinterpret it as a clinical contradiction because event timestamps were absent. Authoritative review, rather than automatic rule revision, is required.

The multimodal workflow links clinical synthesis to metadata-first DICOM processing, HU/geometry QC, interactive and all-slice reconstruction evidence, harmonisation/robustness harnesses and acquisition readiness. Diagnostics expose range compression, class imbalance and repair effects. The 29-field hash-bound SPE adapter, deterministic manifests and zero-fit tests make the transfer boundary executable. Novelty lies in failure-visible integration: algorithms, post-processing, privacy proxies, semantics and utility are tested separately before promotion.