PyDoseRT: A physics-informed, plug-and-play dose engine for gradient-based radiotherapy treatment planning

Attila Simkó, Matthias Kronsteiner, Simon Glatzer, Minh Hoang Vu, Josef A. Lundman, Joakim Jonsson, Jörgen Olofsson, Kristina Sandgren, Wolfgang Lechner, Dietmar Georg, Tommy Löfstedt, Tufve Nyholm, Anders Garpebring, Gerd Heilemann

Keywords: Radiotherapy, dose engine, differentiable physics, treatment planning, deep learning

TL;DR: Developed a PyTorch-based differentiable dose engine for deep learning applications

Registration Requirement: Yes

Abstract: We present PyDoseRT, a GPU-accelerated, differentiable dose engine implemented in PyTorch that computes 3D dose distributions from machine-deliverable parameters (MLC leaf positions, jaw settings, gantry angles, monitor units) while preserving gradients throughout. The engine was validated on 181 clinical VMAT prostate plans from two institutions, achieving mean gamma pass rates of 99.6% and 97.5% (2%/2mm). We further trained a deep learning model on the LUND-PROBE dataset to predict delivered dose from patient anatomy, using PyDoseRT as a differentiable layer for end-to-end training. The model inherently produced deliverable plans with Dice coefficients of 0.87+-0.02 and 0.92+-0.03 for the 50% and 95% isodose volumes on a held-out validation set. PyDoseRT enables TPS-independent, gradient-based radiotherapy optimization and provides a platform for deep learning-based treatment planning.

Reproducibility: https://github.com/UMU-DDI/PyDoseRT

Visa & Travel: Yes

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 40