3112 - A User Prompt-Based Statistical Biomechanical Model Architecture for Rapid Digital Twin Simulation of Pelvic Anatomy for Population-Level Intrafraction Motion Evaluation During Radiation Therapy

12:30pm - 01:45pm ET

Poster Hall - Exhibit Hall A

Screen: 33

POSTER

Presenter(s)

Rene Rodriguez-Perez, MS, MSc - Benemerita Universidad Autonoma de Puebla, Puebla, Puebla

R. E. Rodriguez-Perez¹, G. U. Perez Rojas¹, B. D. C. Alonso¹, and K. Singhrao²; ¹Faculty of Mathematical Physical Sciences, Benemerita Universidad Autonoma de Puebla, Puebla, Mexico, ²Department of Radiation Oncology, Brigham and Women’s Hospital, Dana-Farber Cancer Institute, Harvard Medical School, Boston, MA

Purpose/Objective(s): Population-level quantification of intrafraction motion is challenging because in vivo assessment relies on onboard imaging with poor soft-tissue contrast such as planar kV imaging or limited volumetric information from 2D cine MRI. Population-level statistical studies incorporating anatomical motion therefore require robust, user-directed biomechanical simulation capable of realizing prescribed anatomical transformations. Here we developed and validated a fast, user-prompted generative statistical biomechanical deep learning model to produce physically accurate deformed anatomy by learning implicit organ physics.

Materials/Methods: The overall model architecture involved (1) a CT-to-label-map (LM) model, (2) a user-prompted model to simulate biomechanical organ deformations, and (3) a deep learning model to convert deformed organ structures into synthetic CT (sCT). Training data used 126 patients from a public dataset (100/13/13 split) to generate ground-truth LMs for prostate, urinary bladder, and rectum. The biomechanical simulation was trained using text prompts describing known volume changes, assembled for all LM pair permutations per data split. This dataset trained a 3D attention UNet conditioned on pre-trained DistilBERT embeddings to predict deformed LMs based on prescribed transformations for up to three organs of interest. LM-to-sCT conversion was performed using a 2.5D generative deep learning model (cycleGAN), enabling deformed sCT generation while preserving static patient anatomy.

Architecture validation was performed by computing Dice coefficient differences between predicted and baseline LMs against synthetically generated reference deformations. For evaluation, 210 test LMs were deformed using their corresponding prompts, and DICE coefficients were compared per organ of interest.

Results: Mean DICE coefficient differences between model-predicted and simulated ground-truth bladder, prostate, and rectum volumes were 0.16 ± 0.10, 0.04 ± 0.14, and 0.12 ± 0.10, respectively, where a DICE difference of 0 indicates no difference in spatial overlap; paired t-tests showed no statistically significant differences between generated and ground-truth volumes. These results demonstrate reproducible generation of anatomically consistent digital twins suitable for population-level analysis. The user-prompted biomechanical model generated physically realistic pelvic anatomy consistent with prescribed organ displacements, producing full CT-like volumes in under 4 seconds per volume.

Conclusion: The proposed user-prompted statistical biomechanical model rapidly generates physically realistic pelvic digital twins that accurately reproduce prescribed organ motion. Future work will use this framework for population-level intrafraction motion simulations to characterize the dosimetric impact of pelvic anatomy motion across image-guided radiotherapy techniques.