Physics-Informed Synthetic Tumor Modeling In CT for Training and Stress-Testing Target Delineation AI
Abstract
Purpose
To test whether a physics-informed “time-machine” tumor synthesis pipeline can generate realistic small pancreatic ductal adenocarcinoma (PDAC) targets on contrast-enhanced CT for training and stress-testing AI models intended to support CT-based target delineation, where true early-stage tumors (<2 cm) are rare.
Methods
We curated 3,144 contrast-enhanced CT scans (2,098 PDAC with voxel-wise tumor masks; 1,046 normal controls). Tumor stage distribution was T1=285, T2=396, T3=367, T4=45, and unknown=1,005. A published multi-structure segmenter generated anatomical context masks (pancreas, vessels, ducts, and 20 additional organs). A reverse-temporal cellular-automata model, conditioned on patient factors and scan phase, simulated plausible backward progression from late-stage PDAC to earlier-stage size, shape, and location. A diffusion model then inserted the synthesized small PDAC into the original CT while preserving local anatomy and enhancement patterns, producing paired synthetic early-stage CT scans with voxel-wise lesion masks. An nnU-Net was trained using real PDAC scans augmented with their synthetic early-stage counterparts. Internal evaluation used 650 patients (51 T1 PDAC; 599 controls). External validation used 3,592 CT scans from six international hospitals. A retrospective prediagnostic subset (n=19; 3–36 months before clinical diagnosis) was used to assess sensitivity under subtle disease.
Results
On internal testing for small PDAC (<2 cm), the augmented model achieved AUC 0.975 (95% CI: 0.939–1.000), sensitivity 95.3% (95% CI: 84.5–98.7%), and specificity 98.7% (95% CI: 96.6–99.5%). External testing yielded AUC 0.940 (95% CI: 0.922–0.956), sensitivity 93.2% (95% CI: 89.4–95.6%), and specificity 86.2% (95% CI: 84.9–87.3%). On prediagnostic scans, sensitivity was 36.8% (95% CI: 19.1–59.0%) where original clinical reads reported 0% of these cases.
Conclusion
Physics-informed reverse-temporal synthesis creates anatomically consistent small PDAC targets with voxel-wise masks, enabling data-efficient training and targeted stress-testing of CT-based lesion delineation models under domain shift and subtle disease.