RadWorld: A World Foundation Model for 3D Tomographic Medical Imaging

The human body.
A shared generative foundation.

Pretrained on 232,781 curated scans from 152 datasets across 23 countries, RadWorld learns a shared generative prior across anatomical regions and imaging modalities.

Many tasks.
One generative framework.

CT and MRI tasks differ in the information available and the examination required. RadWorld brings scan generation, imaging translation and treatment simulation into a common formulation: generate a target 3D volume from the supplied conditions.

Generation

Generate virtual 3D CT and MRI volumes.

Generate CT and MRI volumes, and guide CT synthesis with radiology reports or segmentation masks.

Translation

Translate scans across imaging settings.

Condition on an available scan to generate a target modality, MRI sequence or contrast phase.

Setting:CBCTCT

Input cone-beam CT, chest — Source modality
Cone-beam CT

RadWorld

CT

Target modality

CT

Setting:MRICT

Input MR, head and neck — Source modality
MRI

RadWorld

CT

Target modality

CT

Setting:Non-contrast CTContrast-enhanced CT

Input non-contrast chest CT — Source modality
Non-contrast CT

Arterial-phase CT generated by RadWorld from the non-contrast CT — RadWorld
Arterial-phase CT

Real arterial-phase CT of the same patient, aligned to the input — Target modality
Arterial-phase CT

Setting:T1 + T2 + FLAIRT1CE

Input T1 MRI — Source modality
T1 + T2 + FLAIR

T1CE generated by RadWorld from T1, T2 and FLAIR, post-treatment glioma — RadWorld
T1CE MRI

Real T1CE of the same patient — Target modality
T1CE MRI

Simulation

Simulate post-treatment CT volumes.

Condition on pre-treatment CT and TACE details to generate virtual post-treatment scans for hepatocellular carcinoma.

Setting: TACE · Patient 1

Pre-TACE CT of patient 1 — Pre-treatment CT
CT before TACE

Virtual post-TACE CT of patient 1, generated by RadWorld — Virtual post-treatment CT
Generated by RadWorld

Real post-TACE CT of patient 1 — Acquired post-treatment CT
Real CT after TACE

Treatment details

Chemotherapy: Pirarubicin 12 mg, lobaplatin 15 mg
Lipiodol: 5 mL
Target artery: Superior mesenteric artery

Setting: TACE · Patient 2

Pre-TACE CT of patient 2 — Pre-treatment CT
CT before TACE

Virtual post-TACE CT of patient 2, generated by RadWorld — Virtual post-treatment CT
Generated by RadWorld

Real post-TACE CT of patient 2 — Acquired post-treatment CT
Real CT after TACE

Treatment details

Chemotherapy: Pirarubicin 20 mg, oxaliplatin 100 mg
Lipiodol: 20 mL
Target artery: Right hepatic artery

Visual Turing test

Can you tell real from synthetic?

Explore ten tumor scans and decide which are real and which are generated. View the radiologist reader study to see how experts performed.

Axial CT

10 cases · CT & MRI · Tumor outlines

Your answers stay in this browser.

Downstream applications

Generate scans. Support downstream tasks.

Use RadWorld-generated scans to augment training data for classification, segmentation and report generation, or as inputs for survival prediction.

Classification

Improve chest abnormality classification with synthetic training scans.

Tumor segmentation

Improve segmentation across eight tumor types with synthetic training examples.

Report generation

Train report generation models with paired synthetic scans and reports.

Survival prediction

Use virtual post-TACE CT to support progression-free survival prediction.

Classification

Synthetic-to-real training ratio: 2:1.

Classification performance

AUC for chest CT abnormalities, higher is better

Abnormality	Real data only	+ GenerateCT	+ MedSyn	+ RadWorld
Hiatal hernia	0.477	0.483	0.489	0.609
Lung opacity	0.546	0.559	0.613	0.632
Atelectasis	0.566	0.582	0.608	0.690
Adenopathy	0.623	0.612	0.623	0.650
Pericardial effusion	0.633	0.641	0.635	0.672
Calcification	0.579	0.614	0.680	0.688
Lung nodule	0.609	0.647	0.673	0.673
Emphysema	0.535	0.586	0.676	0.772
Bronchiectasis	0.610	0.679	0.615	0.784
Consolidation	0.644	0.657	0.724	0.723
Fibrotic sequela	0.645	0.756	0.737	0.826
Pleural effusion	0.790	0.813	0.852	0.902
Overall	0.613	0.650	0.665	0.723

Tumor segmentation

Synthetic-to-real training ratio: 2:1.

Tumor segmentation performance

Dice for eight cancers, higher is better

Cancer	Real data only	+ DiffTumor	+ FreeTumor	+ RadWorld
NPC	0.576	0.590	0.598	0.629
HNSCC	0.632	0.653	0.654	0.675
NSCLC	0.625	0.651	0.691	0.723
HCC	0.599	0.700	0.703	0.730
PDAC	0.487	0.541	0.535	0.588
CRC	0.438	0.476	0.476	0.531
RCC	0.729	0.762	0.764	0.783
BLCA	0.673	0.726	0.724	0.764

Report generation

Synthetic-to-real training ratio: 2:1.

Report generation performance

Text similarity and clinical accuracy, higher is better

Region	Real data only	+ GenerateCT	+ MedSyn	+ RadWorld
Chest	0.215	0.263	0.264	0.270
Abdomen	0.320	0.340	0.342	0.358
Pelvis	0.193	0.252	0.250	0.258
Chest	0.311	0.320	0.316	0.324
Abdomen	0.413	0.426	0.429	0.442
Pelvis	0.373	0.377	0.372	0.388
Chest	0.322	0.336	0.339	0.352
Abdomen	0.373	0.399	0.395	0.410
Pelvis	0.424	0.428	0.429	0.435
Chest	0.266	0.274	0.288	0.314
Abdomen	0.421	0.442	0.446	0.501
Pelvis	0.665	0.686	0.690	0.741

Survival prediction

Progression-free survival after TACE, evaluated in two cohorts.

Survival prediction performance

C-index for progression-free survival, higher is better

Cohort	Pre-TACE CT only	+ patient and procedural variables	+ real post-TACE CT	+ virtual post-TACE CT
Cohort A	0.577	0.595	0.732	0.727
Cohort B	0.571	0.595	0.726	0.721

View survival curves

Cohort A

Progression-free survival of risk groups predicted from virtual post-TACE CT

Hazard ratio 2.73, P < 0.0001

Cohort B

Progression-free survival of risk groups predicted from virtual post-TACE CT

Hazard ratio 2.45, P < 0.0001

Research team

The team behind RadWorld.

Jin Ye^1,3‡, Chenglong Ma^3,5,7‡, Lu Zhang^6‡, Yanzhou Su⁸, Jiyao Liu^3,7, Wei Li³, Yicheng Wu¹, Zhihao Chen⁷, Yirong Chen³, Hongming Shan⁷, Zhaolin Chen¹, Bohan Zhuang¹, Yu Qiao³, Junjie Zhang², Junjun He^2,3,5*, Jianfei Cai^1*, Yuanfeng Ji^2,4*

‡ Equal contribution · * Corresponding authors

Affiliations

¹Monash University ²Ant Healthcare ³Shanghai Artificial Intelligence Laboratory ⁴Stanford University School of Medicine ⁵Shanghai Innovation Institute ⁶The First Affiliated Hospital of Jinan University ⁷Fudan University ⁸Fuzhou University

Resources

Build with RadWorld.

Cite RadWorldBibTeX

@article{ye2026radworld,
  title   = {A World Foundation Model for 3D Tomographic Medical Imaging},
  author  = {Ye, Jin and Ma, Chenglong and Zhang, Lu and others},
  journal = {arXiv preprint},
  year    = {2026}
}

RadWorld

The human body.A shared generative foundation.

Many tasks.One generative framework.

Generate virtual 3D CT and MRI volumes.

Translate scans across imaging settings.

Setting:CBCTCT

Source modality

RadWorld

Target modality

Setting:MRICT

Source modality

RadWorld

Target modality

Setting:Non-contrast CTContrast-enhanced CT

Source modality

RadWorld

Target modality

Setting:T1 + T2 + FLAIRT1CE

Source modality

RadWorld

Target modality

Simulate post-treatment CT volumes.

Setting: TACE · Patient 1

Pre-treatment CT

Virtual post-treatment CT

Acquired post-treatment CT

Setting: TACE · Patient 2

Pre-treatment CT

Virtual post-treatment CT

Acquired post-treatment CT

Can you tell real from synthetic?

Generate scans. Support downstream tasks.

Classification

Tumor segmentation

Report generation

Survival prediction

Classification

Classification performance

Tumor segmentation

Tumor segmentation performance

Report generation

Report generation performance

Survival prediction

Survival prediction performance

Cohort A

Cohort B

The team behind RadWorld.

Build with RadWorld.

Contact

The human body.
A shared generative foundation.

Many tasks.
One generative framework.