Deep learning‐based dose prediction for MR‐guided prostate SIB: Supporting rapid feasibility assessment and adaptive editing margin selection
Medical Physics · 2026-07 · 53 (7)
DOI: 10.1002/mp.70547 ↗At a glance
Journal article in Medical Physics, published online on 2026-06-30. Written by Hao-Wen Cheng and 4 co-authors. Institution: Department of Radiation Oncology, College of Medicine, University of Florida, Gainesville, Florida, USA. Major MEDLINE topics: Magnetic Resonance Imaging, Deep Learning, Radiotherapy Planning, Computer-Assisted, Prostatic Neoplasms. 27 references.
Abstract (original language)
Background
The template‐based treatment planning workflow in the Unity magnetic resonance‐guided linear accelerator (MR‐Linac) helps facilitate the treatment planning process; however, dose optimization and calculation remain time‐consuming in complex cases. In addition, the deformable image registration (DIR) algorithm used during adaptive planning may produce inaccurate results, thereby making efficient contour editing essential for online adaptation, especially in cases involving multiple organs at risk adjacent to the target.
Purpose
This study aims to resolve key clinical challenges in treatment planning and online adaptive planning for prostate simultaneous integrated boost (SIB) treatment on the MR‐Linac using deep learning (DL)‐based dose prediction. Multiple DL‐based dose prediction models are developed and evaluated for rapid dosimetric feasibility assessment, and the top‐performing model is used to guide the development of a novel contouring strategy to streamline the adaptive planning workflow.
Methods
Six DL models were developed and evaluated for 3D dose prediction using 80 reference and adaptive plans from 20 patients previously treated with SIB intensity‐modulated radiation therapy (SIB‐IMRT) on the Unity MR‐Linac. To address DIR‐introduced contour inaccuracies during adaptation, a contour editing strategy based on a minimal manual editing margin (MMEM) around the target structures was introduced. The MMEM was determined using six margin sizes (5‐30 mm, in 5‐mm increments) evaluated by the top‐performing DL model. Twenty adaptive plans with DIR inaccuracies in bladder, rectum, and bowel bag from 12 patients were analyzed for MMEM determination, and 10 adaptive plans were further evaluated for feasibility in the Monaco treatment planning system.
Results
The dosimetric differences between model predictions and the treatment planning system‐calculated ground truth were small for most models (≤1.5 Gy or 2%), and the average Dice Similarity Coefficients across the 10–100% isodose volumes exceeded 0.9, demonstrating strong spatial agreement. Among the evaluated models, the attention‐gated (AG) U‐Net achieved the best overall performance and was used for the MMEM determination. Based on the MMEM and feasibility analyses, a 10‐mm editing margin is recommended for the rectum, with the use of the semi‐automatic segmentation tool for the bladder and no manual edits for the bowel bag.
Conclusion
The AG U‐Net enabled accurate dose prediction for rapid dosimetric feasibility assessment in prostate SIB treatment on the Unity MR‐Linac. Leveraging this model, an MMEM‐based contouring strategy was developed to guide contour modification and improve the efficiency of online adaptive planning.
Subjects
MeSH terms (MEDLINE)
Author keywords
Authors and institutions
- Department of Radiation Oncology, College of Medicine, University of Florida, Gainesville, Florida, USA.
- Department of Radiation Medicine, Medical University of South Carolina, Charleston, South Carolina, USA.
Cite
APA
Cheng, H. W., Tseng, W. C., Yan, G., Li, J., & Liu, C. (2026). Deep learning‐based dose prediction for MR‐guided prostate SIB: Supporting rapid feasibility assessment and adaptive editing margin selection. Medical Physics, 53(7). https://doi.org/10.1002/mp.70547
Vancouver
Cheng HW, Tseng WC, Yan G, Li J, Liu C. Deep learning‐based dose prediction for MR‐guided prostate SIB: Supporting rapid feasibility assessment and adaptive editing margin selection. Medical Physics. 2026;53(7). doi:10.1002/mp.70547
BibTeX
@article{Cheng2026,
author = {Cheng, Hao-Wen and Tseng, Wen-Chih and Yan, Guanghua and Li, Jonathan and Liu, Chihray},
title = {Deep learning‐based dose prediction for MR‐guided prostate SIB: Supporting rapid feasibility assessment and adaptive editing margin selection},
journal = {Medical Physics},
year = {2026},
volume = {53},
number = {7},
doi = {10.1002/mp.70547}
}
EUVIMED