Training-free generation of 4D CT and other rare acquisitions improves segmentation across modalities
Abstract
RationaleRoutine clinical imaging captures anatomy under a limited set of physiologic and acquisition conditions, leaving segmentation models with sparse examples of how lesions and surrounding tissues deform with compression, respiration, or vascular pulsation. Standard image augmentation expands training datasets but does not explicitly model the mechanics underlying these changes. Patient-specific biomechanical simulation offers an approach to generate anatomically consistent images and corresponding segmentation masks under conditions absent from the original scan, without additional imaging or training a generative model. We hypothesized that incorporating this mechanistically grounded variability into training improves segmentation across imaging modalities.
MethodsWe evaluate across three segmentation tasks: breast cancer on 3D digital breast tomosynthesis, lung tumors on chest CT, and carotid plaque on ultrasound. From each routine scan, patient-specific biomechanical models generate acquisitions absent from the original study--a range of breast compressions, the respiratory cycle, and arterial wall distension across the cardiac cycle. The resulting displacement fields deform both images and masks. Segmentation models were trained with and without the generated data across a range of standard image augmentation techniques, evaluated via cross-validation and held-out test sets. Realism was assessed through blinded reader studies, and generated respiratory motion was validated against real 4D CT.
ResultsTraining on the generated acquisitions improved segmentation across modalities, increasing Dice scores by 6.6% for breast (p<0.05), 10.1% for lung (p<0.05), and 4.5% for carotid plaque (p>0.05) relative to the routine scans alone. Generated respiratory motion closely matched real gated 4D CT, with Dice of 0.94 {+/-} 0.02 for lung volume overlap. Blinded readers discriminated generated from real images at near chance (AUC: breast 0.61, lung 0.41, carotid 0.46).
ConclusionsMechanistic simulation generates imaging acquisitions that clinical systems do not supply, without training data or additional scanning, and improves AI-based disease detection across modalities.
Related
- Salivary glands auto-segmentation in head and neck radiotherapy: imaging modalities, models, and clinical relevance
- Advances and perspectives of next-generation antitumor antibody-based modalities: Structural innovation and target-driven strategies.
- Comparative effects of different exercise training modalities on exercise capacity, pulmonary function, and health-related quality of life in patients with interstitial lung disease: a systematic review and network meta-analysis.
- Automated voxel-based prostate volumetry improves PSA density-based detection of clinically significant prostate cancer in the PSA gray zone: comparison with manual segmentation and the ellipsoid formula.
- Lightweight Frequency-Gated Enhancement of nnU-Net Improves Data Efficiency in Nasal Sinus CT Segmentation
BibTeX
@misc{sadee2026trainingfree,
title = {{Training-free generation of 4D CT and other rare acquisitions improves segmentation across modalities}},
author = {C. Sadee and A. Dils and N. S. Hundal and K. Wan and A. Chun and A. Shukla and I. L. Xi and M. Van Puyvelde and K. Hartmann and P. Shah and M. Hirschvogel and W. Zhou and C. Lin and V. Gangaram and A. G. Er and M. Seweryn and W. Van Criekinge and T. N. Nandi and J. I. Aniegbe and L. Shen and A. C. Ilo and F. Immel-Sadee and A. O. Nnamani and D. L. Levin and M. Hao and E. S. McDonald and N. C. Achilefu and N. I. Iloanusi and S. P. Poplack and B. Z. Dashevsky and S. M. Damrauer and I. Okoye and O. Gevaert},
howpublished = {medRxiv},
year = {2026},
month = oct,
doi = {10.64898/2026.10.22.26365566},
url = {https://www.medrxiv.org/content/10.64898/2026.10.22.26365566v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - Training-free generation of 4D CT and other rare acquisitions improves segmentation across modalities
AU - C. Sadee
AU - A. Dils
AU - N. S. Hundal
AU - K. Wan
AU - A. Chun
AU - A. Shukla
AU - I. L. Xi
AU - M. Van Puyvelde
AU - K. Hartmann
AU - P. Shah
AU - M. Hirschvogel
AU - W. Zhou
AU - C. Lin
AU - V. Gangaram
AU - A. G. Er
AU - M. Seweryn
AU - W. Van Criekinge
AU - T. N. Nandi
AU - J. I. Aniegbe
AU - L. Shen
AU - A. C. Ilo
AU - F. Immel-Sadee
AU - A. O. Nnamani
AU - D. L. Levin
AU - M. Hao
AU - E. S. McDonald
AU - N. C. Achilefu
AU - N. I. Iloanusi
AU - S. P. Poplack
AU - B. Z. Dashevsky
AU - S. M. Damrauer
AU - I. Okoye
AU - O. Gevaert
PY - 2026
DA - 2026/10/22
PB - medRxiv
DO - 10.64898/2026.10.22.26365566
UR - https://www.medrxiv.org/content/10.64898/2026.10.22.26365566v1
AB - RationaleRoutine clinical imaging captures anatomy under a limited set of physiologic and acquisition conditions, leaving segmentation models with sparse examples of how lesions and surrounding tissues deform with compression, respiration, or vascular pulsation. Standard image augmentation expands training datasets but does not explicitly model the mechanics underlying these changes. Patient-specific biomechanical simulation offers an approach to generate anatomically consistent images and corresponding segmentation masks under conditions absent from the original scan, without additional imaging or training a generative model. We hypothesized that incorporating this mechanistically grounded variability into training improves segmentation across imaging modalities. MethodsWe evaluate across three segmentation tasks: breast cancer on 3D digital breast tomosynthesis, lung tumors on chest CT, and carotid plaque on ultrasound. From each routine scan, patient-specific biomechanical models generate acquisitions absent from the original study--a range of breast compressions, the respiratory cycle, and arterial wall distension across the cardiac cycle. The resulting displacement fields deform both images and masks. Segmentation models were trained with and without the generated data across a range of standard image augmentation techniques, evaluated via cross-validation and held-out test sets. Realism was assessed through blinded reader studies, and generated respiratory motion was validated against real 4D CT. ResultsTraining on the generated acquisitions improved segmentation across modalities, increasing Dice scores by 6.6% for breast (p<0.05), 10.1% for lung (p<0.05), and 4.5% for carotid plaque (p>0.05) relative to the routine scans alone. Generated respiratory motion closely matched real gated 4D CT, with Dice of 0.94 {+/-} 0.02 for lung volume overlap. Blinded readers discriminated generated from real images at near chance (AUC: breast 0.61, lung 0.41, carotid 0.46). ConclusionsMechanistic simulation generates imaging acquisitions that clinical systems do not supply, without training data or additional scanning, and improves AI-based disease detection across modalities.
ER -