3307 - Efficacy of an Adaptive Plan Quality Predictor for Decision Support in Online MR-Guided Adaptive Radiotherapy for Prostate Stereotactic Ablative Radiotherapy

S. Wang¹, A. Maniscalco², F. C. Su³, X. Liang¹, B. Washington¹, C. Kabat¹, A. Garant¹, N. B. Desai¹, A. Pompos¹, R. Hannan¹, D. X. Yang¹, D. Nguyen², and M. H. Lin¹; ¹Department of Radiation Oncology, University of Texas Southwestern Medical Center, Dallas, TX, ²MAIA Laboratory, Department of Radiation Oncology, University of Texas Southwestern Medical Center, Dallas, TX, ³University of Texas Southwestern Medical Center, Dallas, TX

Purpose/Objective(s): The current MR-guided online adaptive radiotherapy (oART) system necessitates an initial decision before proceeding into two distinct workflows: Adapt-to-Position (ATP) and Adapt-to-Shape (ATS). While ATS can achieve higher plan quality and accommodate daily anatomical variations, ATP frequently emerges as the preferred clinical option due to its significantly lower resource demands. We hypothesize that a deep learning dose prediction model can be designed to predict the ATS plan quality to assist the clinician in rapid evaluation of the treatment options, enable an informed choice between the adaptive workflows, and optimize resource distribution for MR-guided oART. Materials/

Methods: 60 initial plans and 130 online ATS plans of 60 prostate SAbR patients (treated with 45 Gy and 40 Gy SIB in 5 fractions) were selected to train and validate an in-house deep learning-based adaptive dose predictor. A population model was trained from the MR images, planning target volumes (PTVs), PTV-structure distance map, and dosimetric goal arrays of 40 patients, and then fine-tuned to intentionally overfit the initial plan of the other 20 patients to create individual prediction models for their respective adaptive plans (that were withheld for testing). The predicted ATS plans for these 20 test patients were compared with their clinically delivered ATS plans via dose evaluation metrics commonly used for prostate SAbR.
Results: All predicted ATS plans were normalized to the preplan target coverage, similar to the online workflow. Once trained, the model is capable of generating predictions within 10 seconds. Between the predicted and clinical plans, the mean low dose PTV coverage difference was within 1.0 %. For organs-at-risk, the predicted doses and dose volumes for rectum, rectal wall, bladder, and bladder wall are less than 1.0 Gy and 1.5%, respectively. The only significant deviation was observed for urethra where the predictor overestimated the max dose.
Conclusion: The in-house deep learning-based dose predictor for MR-guided oART was capable of accurately predicting the quality of ATS plans. While the overestimation of urethra dose suggests an area for further refinement, the predictor demonstrated its efficacy as a helpful addition to a developing ATP vs. ATS decision support framework to streamline the online adaptive workflow and lower the resource-intensiveness barrier for MR-guided oART.

Abstract 3307 – Table 1

Structure	Dose Constraint	Predicted - Clinical
PTV 45Gy	V45Gy > 85%	Normalized to Clinical
PTV 40Gy	V40Gy > 95%	0.9 ± 1.8 %
Rectum	V39Gy < 33%	1.5 ± 1.7 %
	V24Gy < 50%	0.2 ± 3.2 %
Rectal Wall	Dmax < 45Gy	-0.5 ± 2.6 Gy
	V37Gy < 20cc	-0.0 ± 0.7 cc
	V24Gy < 50%	0.1 ± 3.4 %
Bladder	Dmean < 30Gy	0.4 ± 1.3 Gy
Bladder Wall	Dmax < 43Gy	1.0 ± 1.5 Gy
	V18.3Gy < 18cc	0.1 ± 2.9 cc