S2286
Physics - Machine learning and AI algorithms
ESTRO 2026
Digital Poster 1808
the planning MR to the treatment MR images. The contour propagation accuracy was quantitatively evaluated against physician-delineated references using the Dice similarity coefficient (DSC). For benchmarking, DIR results obtained from the clinical treatment planning system (Monaco, Elekta) were also assessed. Results: The model achieved the inference time of 0.5–1.0 s per MRI pair. In the test dataset, the mean DSCs for the proposed method and Monaco were 0.81 ± 0.14 / 0.72 ± 0.25 for the bladder, 0.93 ± 0.03 / 0.97 ± 0.02 for the prostate, 0.83 ± 0.09 / 0.87 ± 0.12 for the seminal vesicles, and 0.88 ± 0.05 / 0.91 ± 0.07 for the rectum, respectively. A representative case is shown in Figure 1. Although complete agreement was not achieved for the bladder, the proposed method produced contours that were closer to the reference delineation than those generated by the commercial used DIR software. The proposed method achieved higher bladder contour accuracy than the clinical DIR software, showing comparable trends for the prostate and rectum, while maintaining fast inference.
Interactive radiotherapy dose prediction for prostate cancer using a 3D text-guided diffusion model YENJUNG CHEN 1 , Lin-Shan Chou 1 , Shu-Sheng Li 2 , Ti- Hao Wang 3,2 1 Department of Heavy Particles & Radiation Oncology, Taipei Veterans general hospital, Taipei, Taiwan. 2 Department of Radiation Oncology, China Medical University Hospital, Taichung, Taiwan. 3 Department of Medicine, China Medical University, Taichung, Taiwan Purpose/Objective: Current deep learning models for dose prediction, such as U-Nets and GANs, function as non-interactive "black boxes." They often produce blurry, dose- averaged predictions or suffer from training instability. Furthermore, many 2D models fail to ensure dose continuity across the z-axis, limiting clinical realism. We aimed to overcome these limitations by developing the 3D, text-guided diffusion model for prostate radiotherapy. This model generates high-fidelity 3D dose distributions from CTs and contours and allows for interactive plan refinement using natural language prompts. Material/Methods: A retrospective dataset from 133 patients was used. To capture planning variability, the training set included 3-8 different clinically-approved plan styles per patient (e.g., IMRT, VMAT, rectal-sparing), totaling over 742 unique 3D dose distributions. Inputs included the 3D CT volume and contours for the PTV, bladder, femoral heads, and rectum. The output was the corresponding 3D dose matrix. We developed a 3D latent diffusion model conditioned on both the anatomical segmentations and text prompts. A pre- trained CLIP text-encoder was used to convert text inputs (e.g., "standard VMAT plan" or "IMRT to minimize rectum dose") into embeddings. These embeddings were integrated into the model's U-Net backbone via cross-attention layers, allowing the text to guide the iterative denoising and dose-generation process. The model was evaluated by comparing its baseline predictions (using a neutral prompt) against the clinically-approved plans using standard DVH metrics. To validate the interactive feature, we applied specific textual prompts (e.g., "reduce bladder dose") and quantified the resulting change in OAR DVH parameters compared to the baseline. Results: In 42 independent test cases, the 3D diffusion model generated volumetrically smooth dose distributions with high dosimetric accuracy. Baseline predictions were highly comparable to clinical plans: mean PTV V95% was 98.7± 0.02% (model) vs. 98.0± 0.02% (clinical), difference of rectum V50Gy was 1.11 ± 1.24
Conclusion: A deep learning–based DIR with adaptive loss improved bladder contour propagation over clinical used DIR software and sub-second inference, supporting its feasibility for MRI-guided online adaptive prostate radiotherapy. Keywords: deep learning, deformable image registration
Made with FlippingBook - Share PDF online