Challenge 3

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

Team: 

PONS Ltd Sti, Turkiye

Proposal title: 

Ultrasound AI – Device Agnostic Medical Imaging AI

Organisation

PONS is a Rutgers-origin start-up, founded in 2023 and headquartered in Newark, New Jersey and Istanbul Turkiye, that develops physics-guided AI software to make ultrasound images consistent and AI-ready on any scanner. Its patented technology has been validated in a multicentre clinical study of 62,912 scans.

Project info

PONS treats the challenge’s three problems separately because they fail differently: cohort imbalance is a sampling problem, missing clinical variables a correlation problem, and kVp heterogeneity a physics problem, since tube voltage changes the HU values that identical tissue returns.

Before any generation, each CT study is decomposed by the company’s patented engine (US 11,308,590) into a structural local-phase channel, which encodes tissue interfaces and is invariant to intensity and contrast, and an attenuation channel from a Beer-Lambert forward model conditioned on the scan’s kVp, with filter parameters optimised per volume by AI. A slab-based 3D latent diffusion model of about 350M parameters then works in this decomposed space, conditioned on kVp, series type and clinical variables. Harmonisation becomes a mapping to a pseudo-monoenergetic reference rather than a style transfer; a missing series is synthesised by editing the attenuation channel while the phase channel stays fixed, so anatomy cannot drift; and a differentiable radiomic-stability term is optimised during training. A masked cross-modal transformer conditioned on CT embeddings imputes clinical variables, with SHAP attributions providing the variable-importance tool.

The autoencoder is pretrained on public thorax CT, and only model weights enter the secure environment.

Details

Lung-cancer imaging cohorts skew male, are incomplete and are acquired at different tube voltages. Generative models working directly on CT voxels often produce images that look right but quietly shift the radiomic biomarkers researchers rely on.

PONS proposes to move generation into a physically grounded representation. Each study is first split into a structural channel that captures tissue interfaces regardless of intensity or contrast, and an attenuation channel modelled from the X-ray physics of the scan. A 3D latent diffusion model then works in that space to rebalance subgroups, complete missing clinical variables and CT series, and harmonise scans to a common reference. Because anatomy sits in a channel the generator leaves unchanged when synthesising a missing series, and radiomic stability is optimised during training, biomarkers are designed to survive the process. The decomposition engine already underpins PONS’s commercial ultrasound product, where it has been validated in a multicentre clinical study across three sites and three scanner manufacturers.

In the Advance Phase, PONS will port this engine to thorax CT and validate the pipeline on the CHAIMELEON cohort inside the EUCAIM secure processing environment. Targets go beyond the challenge thresholds for fidelity, with at least 85% of variables within tolerance, and include halving imbalance by sex, age band and tumour location, a 30% improvement over median imputation for continuous variables, 95% imaging completion and median radiomic concordance of 0.85 under noise. In future clinical use, synthesising contrast-enhanced series from non-contrast scans could also avoid repeat scans, radiation dose and contrast media.