S2779
RTT - RTT contouring, target definition, and treatment planning
ESTRO 2026
1 Department of Diagnostics and Intervention, Umeå University, Umeå, Sweden. 2 Department of Radiation Oncology, Medical University of Vienna, Vienna, Austria. 3 Department of Computing Science, Umeå University, Umeå, Sweden Purpose/Objective: Radiotherapy treatment planning remains time- intensive and requires iterative expert intervention despite automation advances in individual workflow tasks. Current deep learning approaches require validation within commercial treatment planning systems (TPS), creating dependencies that limit research flexibility and clinical translation. We developed a fully-differentiable, physics-informed dose engine with DICOM RTPLAN import/export capabilities, enabling end-to-end learning from medical images to deliverable machine parameters, eliminating TPS dependency for radiotherapy AI research. Material/Methods: We implemented a differentiable dose engine in PyTorch that computes delivered dose from treatment delivery parameters (MLC leaf positions, jaw settings, monitor units). The engine decomposes physical beam transport into modular, GPU-accelerated layers: fluence map generation from aperture parameters, 3D fluence propagation, radiological depth calculation, pencil beam kernel convolution[1], and dose volume reconstruction. Hardware constraints (MLC physical limits, jaw collimator boundaries, monitor unit specifications) are differentiably enforced, ensuring all predictions are inherently deliverable. The DICOM RTPLAN import/export functionality enables easy integration with clinical systems, while the collected physics-informed loss functions allow for robust plan optimization. The dose engine was validated against clinical TPS results using water phantom measurements and retrospective patient plans. Direct gradient-based optimization was performed to generate treatment plans requiring only a patient CT, structure contours and clinical treatment goals as input. Results: Water phantom depth-dose curves matched measurements with mean absolute error of 0.303 cGy (0.2% of maximum dose). Imported RTPLAN files demonstrated a dosimetric accuracy with gamma pass rates of 91.5% (3%/3mm criteria) compared to original TPS dose distributions. The gamma analysis results are primarily limited by the current lack of in-field inhomogeneity corrections, which is under active development. Direct gradient-based optimization successfully generated clinically deliverable plans on 20 prostate cases, converging in 2500 iterations using AdamW optimization. The inference time of the dose engine for a 168x188x188 CT volume is 0.3 seconds on
patient’s body.Dosimetric parameters evaluated for the PTV included V95%, V100%, V105%, and V107% (volume receiving x% of the prescribed dose), as well as Dmin, Dmean, and Dmax (minimum, mean, and maximum dose). Differences for each parameter were tested for normality (Shapiro–Wilk), followed by a t- test for normal data or the Wilcoxon test otherwise. Results: The differences in Dmax were normally distributed and statistically significant ( μ = 0.88%, σ = 0.92%, p < 0.001). The remaining non-normally distributed dosimetric parameters showed significant differences in Dmean ( μ = 0.58%, σ = 0.92%, p = 0.004), V107% ( μ = 0.58%, σ = 0.95%, p = 0.004), and V105% ( μ = 5.19%, σ = 6.14%, p < 0.001). No significant differences were found for V95% ( μ = − 0.26%, σ = 1.05%, p = 0.242), V100% ( μ = 2.00%, σ = 9.47%, p = 0.083), or Dmin ( μ = 0.12%, σ = 0.05%, p = 0.317). Although V100% showed mean differences above 2%, it did not reach significance (p > 0.05). The largest variation was observed in V105%, with a mean increase of ~5%. Conclusion: These results show that excluding the VB from the calculation volume affects PTV dose distribution, overestimating the calculated dose while maintaining target coverage. Considering that TPS uncertainties may reach 2% (V105%), total uncertainty could exceed the ±5% accuracy recommended by the ICRU. Therefore, keeping the VB within the calculation BC is essential to ensure dose accuracy in CSI. References: Olch AJ, Gerig L, Li H, Mihaylov I, Moran JM, Pawlicki T, et al. Dosimetric effects caused by couch tops and immobilization devices: Report of AAPM Task Group 176. Med Phys. 2014;41(6):061501.Jabbari K, Almasi T, Rostampour N, Tavakoli MB, Amouheidari A. Evaluating the effect of the vacuum bag on the dose distribution in radiation therapy. J Cancer Res Ther. 2018;14(6):1245–1250.Oulhouq Y, Rrhioua A, Bakari D, Zerfaoui M, Krim D. The effect of carbon fibre treatment couch with and without immobilisation devices on radiotherapy dose calculation using three different planning algorithms and photon beam energies. J Radiother Pract. 2021;1–5. Keywords: Vacuum Bag, Body Contour, Dosimetric evaluation Digital Poster 2237 A Differentiable Dose Engine for AI-Driven Radiotherapy. Attila Simkó 1 , Minh Vu 1 , Matthias Kronsteiner 2 , Simon Glatzer 2 , Josef A Lundman 1 , Joakim Jonsson 1 , Jörgen Olofsson 1 , Kristina Sandgren 1 , Wolfgang Lechner 2 , Dietmar Georg 2 , Anders Garpebring 1 , Tommy Löfstedt 3 , Tufve Nyholm 1 , Gerd Heilemann 2
Made with FlippingBook - Share PDF online