One-for-All Adaptive Radiotherapy Planning Agent: A Foundation Framework for Daily CBCT-Guided Adaptive Radiotherapy Restricted; Files Only

Abstract

Radiotherapy relies on precise anatomical targeting, yet standard practice utilizes a single pretreatment scan that ignores both high-frequency intra-fractional motion and cumulative inter-fractional anatomical shifts. While adaptive radiotherapy (ART) mitigates these discrepancies, it remains clinically impractical due to the labor-intensive requirements of image synthesis, registration, segmentation, and dose calculation. To overcome these barriers, this dissertation explores the evolution of artificial intelligence in ART. We first focus on the intra-fractional challenge by developing ultra-fast CBCT reconstruction models to correct for real-time artifacts and anatomical motion during delivery. We then build the infrastructure for inter-fractional adaptation, starting with unsupervised CBCT-to-CT synthesis to provide daily planning-quality imaging. This is followed by the development of high-performance deep learning models to automate the individual tasks of deformable registration, multi-organ segmentation, and volumetric dose prediction. Building upon these successes, we propose a unified One-for-All foundation network to resolve the systemic inefficiencies and memory overhead of siloed models. This architecture consolidates the entire planning chain into a single, shared-representation manifold—utilizing Stable Diffusion 3.5 VAE latents and DINOv2 features—ensuring total consistency across tasks while significantly reducing computational demand. This full pipeline is orchestrated by an AI agent: One-for-All Adaptive Radiotherapy Planning Agent (A-RPA). By interpreting clinician intent, the agent autonomously selects and executes the optimal adaptation pathway directly from daily imaging. The A-RPA framework completes the entire online adaptive workflow in under two minutes while maintaining strict physician oversight, providing a scalable and fully automated solution for real-time, high-precision adaptive care.

Table of Contents

1 Introduction 1.1 The Evolution and Challenges of Adaptive Radiotherapy 1.2 Deep Learning Across the Radiotherapy Workflow 1.3 Agentic Artificial Intelligence and Workflow Orchestration 1.4 Dissertation Outline and Main Contributions

2 Unconditional Synthesis of Medical Images for Radiotherapy 2.1 Transformer-based Improved Denoising Diffusion Probabilistic Model 2.2 Conclusion

3 Conditional Synthesis of Medical Images for Radiotherapy 3.1 Synthetic CT Generation from MRI using 3D Transformer-based IDDPM 3.2 Cross-modality MRI Synthesis via Mutually Conditioned Cycle-guided Diffusion 3.3 Full-dose Whole-body PET Synthesis using High-efficiency Consistency Models 3.4 Conclusion

4 Sparse Conditional Ultra-fast Volumetric Image Synthesis for Radiotherapy 4.1 Data-Driven Volumetric CT Image Generation from Surface Structures 4.2 Volumetric CBCT Reconstruction using Single X-ray Projection Image 4.3 Patient-Specific CBCT Synthesis for Ultra-fast Tumor Localization 4.4 Conclusion

5 Unsupervised Conditional Synthesis of Medical Images for Radiotherapy 5.1 Schrödinger bridge based Implicit diffusion inversion 5.2 Conclusion

6 Registration of Multimodal Medical Images for Radiotherapy 6.1 FoundationMorph: A Vision-Language Foundation Model for Unsupervised Registration 6.2 F-DiffMorph: A 3D Foundation Model for Unsupervised Diffeomorphic Registration 6.3 Conclusion

7 Deep Learning for Automated Medical Image Segmentation 7.1 Male Pelvic Multi-organ Segmentation Using Token-based Transformer V-Net 7.2 MLP-Vnet: Efficient 3D Abdominal Segmentation via Token-based MLP-Mixer 7.3 Conclusion

8 Dose Prediction for Inter-fractional Adaptive Radiotherapy 8.1 Dose-VFM: A Vision-Language Flow-Matching Diffusion Model 8.1.5 Conclusion

9 Agentic AI for One-for-All Adaptive Radiotherapy Planning 9.1 Experimentation Setup and Implementation Details 9.2 Result and Discussion 9.3 Conclusion

Bibliography

About this Dissertation

Rights statement

• Permission granted by the author to include this thesis or dissertation in this repository. All rights reserved by the author. Please contact the author for information regarding the reproduction and use of this thesis or dissertation.

School	Laney Graduate School
Department	Computer Science and Informatics
Degree	Ph.D.
Submission	Dissertation
Language	English
Research Field	Computer Science Health Sciences, Radiology
Palabra Clave	Image Registration Adaptive Radiotherapy (ART) Image Segmentation Diffusion Models Agentic AI Dose Prediction Foundation Models Medical Image Synthesis Vision-Language Models
Committee Chair / Thesis Advisor	Xiaofeng Yang, Ph.D., Emory University
Committee Members	Saima Rathore, Ph.D., Emory University Liang Zhao, Ph.D., Emory University Justin Roper, Ph.D., Emory University

Última modificación

Primary PDF

Thumbnail	Title	Date Uploaded	Actions
	File download under embargo until 27 May 2028	2026-05-12 15:04:21 -0400	File download under embargo until 27 May 2028

Supplemental Files

Thumbnail	Title	Date Uploaded	Actions