Research

Warm starts, validated at clinical scale.

Before Life Operators generalize, they must win somewhere specific. Two disease programs — renal cancer and the heart — have already closed the loop from data to operator to validation.

Program 01 · Renal cancer

RenalCLIP — the disease‑specific foundation model.

Multi-center, multimodal, fully followed-up: a renal-cancer validation base built for prognosis-grade tasks.

12,809patients
27,866CT scans
4,557pre-operative multi-phase CT cases
4,103small renal tumors
20,000+digital pathology slides

Benign vs. malignant diagnosis

RenalCLIP0.871
Radiomics0.810
R.E.N.A.L. score0.490

AUC. Outperforms conventional radiomics and clinical scoring models.

Indolent vs. invasive assessment

RenalCLIP0.783
Radiomics0.736
R.E.N.A.L. score0.704

AUC. Predictions correlate significantly with survival differences.

Against clinicians

On small renal-mass differentiation, performance surpasses specialist urological radiologists — and significantly exceeds non-specialist radiologists and urologists.

Against foundation models

In multi-center evaluation, the disease-specific RenalCLIP outperforms general-purpose foundation models — CT-CLIP, Merlin, CT-FM — on renal tasks.

Program 02 · The heart

A data–solver–simulation loop for cardiology.

In-hospital, million-scale, multimodal, report-paired — the data base beneath the Cardio-World milestone.

3.5M+ECG recordings
1M+CTA studies
1M+echocardiograms
1.5M+chest X-rays
10,000+cardiac MR
5,000+cardiac PET

Cardio-Solver

1000×+ faster

Neural-operator acceleration over traditional FEM simulation — fast, patient-specific mechanical simulation.

Aging dynamics

3 cohorts reproduced

Consistent cardiac-aging trajectories — and candidate targets — recovered across Hammersmith, UK Biobank and CHPP.

Virtual clinical trials

Validated against Phase I

Two commissioned virtual-trial projects for innovative device companies, cross-checked against Phase I clinical results.

Publications

Representative work.

  • Nature Biomedical EngineeringLeveraging large language and vision models for knowledge extraction from large-scale image–text colonoscopy records2025
  • Nature Machine IntelligenceA personalized time-resolved 3D mesh generative model for unveiling normal heart dynamics2025
  • Nature MedicineA population-based phenome-wide association study of cardiac and aortic structure and function2020
  • Nature CommunicationsA disease-centric vision–language foundation model for precision oncology in kidney cancer2026
  • Nature CommunicationsArtificial intelligence links CT images to pathologic features and survival outcomes of renal masses2025
  • RadiologyPrecise differentiation of small renal masses2024
  • Medical Image AnalysisCardiac motion modelling2023
  • Medical Image AnalysisCMRxMotion — organizing the global cardiac MR challenge2026
  • ESC Heart FailureHydrogel-device virtual clinical trial versus Phase I clinical comparison2025

From validated programs to full‑scale simulation.