Related Experiment Video
Updated: Jul 1, 2026
06:08
A Cognitive Fusion-guided Prostate Biopsy Using Multiparametric Magnetic Resonance Imaging and Transrectal Ultrasound
Published on: March 21, 2025
Deep learning-based dose prediction for MR-guided prostate SIB: Supporting rapid feasibility assessment and adaptive
Hao-Wen Cheng1, Wen-Chih Tseng2, Guanghua Yan1
1Department of Radiation Oncology, College of Medicine, University of Florida, Gainesville, Florida, USA.
Medical Physics
|June 30, 2026
Summary
Deep learning models accurately predict radiation dose for prostate SIB treatment on MR-Linacs. This enables a new contouring strategy to improve online adaptive planning efficiency.
Area of Science:
- Radiation Oncology
- Medical Physics
- Artificial Intelligence in Medicine
Background:
- Template-based treatment planning on Unity MR-Linac simplifies workflows but dose optimization is time-consuming for complex cases.
- Deformable image registration (DIR) in adaptive planning can yield inaccurate contours, necessitating efficient contour editing for online adaptation, especially with adjacent organs at risk.
Purpose of the Study:
- To address clinical challenges in prostate simultaneous integrated boost (SIB) treatment planning and online adaptive planning on the MR-Linac using deep learning (DL) dose prediction.
- To develop and evaluate DL models for rapid dosimetric feasibility assessment and guide a novel contouring strategy to streamline adaptive planning.
Main Methods:
- Six DL models were developed for 3D dose prediction using 80 reference and adaptive plans from 20 SIB intensity-modulated radiation therapy (SIB-IMRT) patients on the Unity MR-Linac.
- A minimal manual editing margin (MMEM) strategy was introduced to address DIR inaccuracies, with margin sizes (5-30 mm) evaluated by the top DL model.
- MMEM was determined using 20 adaptive plans with DIR inaccuracies, and 10 plans were evaluated for feasibility in the Monaco treatment planning system.
Main Results:
- DL models showed small dosimetric differences (≤1.5 Gy or 2%) compared to treatment planning system calculations, with strong spatial agreement (Dice Similarity Coefficients > 0.9).
- The attention-gated (AG) U-Net model performed best and was used for MMEM determination.
- A 10-mm editing margin for the rectum, semi-automatic segmentation for the bladder, and no manual edits for the bowel bag were recommended based on MMEM and feasibility analyses.
Conclusions:
- The AG U-Net model facilitates accurate dose prediction for rapid dosimetric feasibility assessment in prostate SIB treatment on the Unity MR-Linac.
- An MMEM-based contouring strategy, guided by the AG U-Net, improves contour modification and enhances the efficiency of online adaptive planning.