Deep Learning-Based Estimation of Minimal Editing Margins for Online Adaptive MR-Linac Prostate SIB
Abstract
Purpose
Online adaptive prostate SIB treatment on the Unity MR-Linac is challenged by the limited robustness of deformable image registration (DIR)-based contour propagation. This study investigates the use of a deep learning (DL) dose prediction model to guide the development of a novel contouring strategy that streamlines online adaptive planning while maintaining dosimetric accuracy.
Methods
A contour-editing strategy using a minimal manual editing margin (MMEM) around target structures was developed and evaluated with six margin sizes (5–30 mm, 5-mm increments). An attention-gated U-Net DL dose prediction model, trained on 80 prostate SIB cases, was used to quantify dose variations resulting from various editing margins. Twenty adaptive plans with DIR inaccuracies in the bladder, rectum, and bowel bag were analyzed for MMEM determination. Hybrid structure sets were created by combining DIR-propagated contours, MMEM-modified regions, and semi-automatically segmented (SAS) bladder contours generated using Monaco treatment planning system (TPS). Dosimetric accuracy was assessed by comparison with fully manually edited structure sets. Planning feasibility using MMEM-based hybrid structures was further evaluated in the Monaco TPS on ten additional adaptive plans.
Results
SAS bladder contours showed dosimetric differences ≤0.8 Gy. Rectal dosimetric differences were ≤2 Gy (or 2%) when using 5-mm and 10-mm MMEMs. Bowel bag dosimetric differences remained ≤1.5 Gy without manual editing, aside from correcting contour overlaps. Based on these findings, feasibility evaluation was performed using hybrid structure sets incorporating SAS bladder contours, no manual bowel bag editing, and 5-mm and 10-mm MMEMs for the rectum. Average dosimetric differences for the bladder, rectum, and bowel bag were ≤1.0 Gy with a 10-mm MMEM and ≤1.5 Gy with a 5-mm MMEM.
Conclusion
A conservative MMEM-based hybrid contouring strategy, with a recommended 10-mm rectal MMEM, improves the robustness and efficiency of online adaptive prostate SIB planning and supports streamlined MR-Linac clinical workflows.