S1775
Physics - Dose prediction/calculation, optimisation and applications for particle therapy planning
ESTRO 2026
large patients. Designing these geometries is time- consuming and highly dependent on the medical physicist (MP) expertise. Previous studies have explored automation using deep learning [2]. We propose a non-rigid, deformation-based method that automatically adapts field geometry to patient anatomy through deformable mapping, determining the optimal number and layout of isocenters and adding two when needed to improve arm coverage. Material/Methods: The dataset included 121 patients treated with TMI/TMLI between 2011 and 2024. For each patient, the PTV and RT plan were exported in DICOM format. The deformation model [3] uses PTV-derived meshes generated via the marching cubes algorithm after normalization and remeshing (Fig. 1a). Each mesh is deformed onto four distinct templates, representing different geometry configurations that vary according to patient largeness and the iliac–rib distance (Fig. 1b). The alignment between the patient PTV and each template is quantified using Chamfer distance, the most similar template was selected. A point-to-point correspondence is established using the K-nearest neighbors algorithm (Fig. 1c) to adapt the template geometry to the new patient (Fig. 1d). Model evaluation was conducted on a test set of 12 patients randomly collected over a 12-year period, comparing manual and automatically generated field geometries, using a four-point scale (1 = not adequate, 4 = adequate) in a blinded review by two MPs.
refines these parameters. With the goal of maximizing lattice density, a matrix correlating geometric center, rotation angle, and lattice density is established, ultimately determining the optimal distribution of lattice points in terms of both quantity and spatial arrangement. Results: Statistical analysis was performed on clinical data from 115 cases treated with hexagonal close-packed (HCP) lattice radiotherapy. The peak-to-valley dose ratio (PVDR) of the clinical plans was 2.04 ± 0.37, and the ablation-to-dose ratio (ADR) was 1.94 ± 0.54. The cases were categorized into five anatomical regions: head & neck, thorax, abdomen, pelvis, and extremities, with lattice density indices of 5.44 ± 1.63, 3.94 ± 2.24, 3.12 ± 3.21, 2.97 ± 1.34, and 6.04 ± 0.29, respectively. For five representative cases, a dual-degree-of-freedom optimized lattice arrangement was applied in physical dose planning. The resulting PVDR was 1.97 ± 0.22, ADR was 2.43 ± 0.37, and the lattice density index improved to 6.56 ± 1.16. Conclusion: This study employed a dual-degree-of-freedom dynamic lattice optimization approach to achieve optimal lattice distribution tailored to individualized tumor anatomy, significantly enhancing the dosimetric performance of lattice radiotherapy with marked improvements in both ablation-to-dose ratio and lattice density index. The research advances a paradigm shift in spatially fractionated radiotherapy— transitioning from empirical fixed templates to personalized dose sculpting in lattice arrangement— which will actively contribute to improved tumor treatment efficacy. Keywords: Dosimetry, radiotherapy Automatic Generation of Field Geometry for Total Marrow Irradiation Using a Non-Rigid Deformation Model Giorgio Longari 1 , Nicola Lambri 2 , Pietro Mancosu 2 , Simone Melzi 1 1 Computer Science, University of Milano-Bicocca, Milano, Italy. 2 Radiotherapy and Radiosurgery Department, IRCCS Humanitas Research Hospital, Milan, Italy Purpose/Objective: Total Marrow (Lymphoid) Irradiation (TMI/TMLI) is an advanced RT technique targeting bony structures, major lymph node chains, and the spleen [1]. VMAT- based TMI/TMLI requires complex field geometries with multiple isocenters and jaw settings. Typically, five isocenters and ten overlapping fields cover the upper body, with additional isocenters needed for Digital Poster 4319
Results: Differences were found between manual and automatic geometries for both MPs, with median scores of 3/3 and 4/3 (Figure 2). The automatic plans were rated at least as adequate as the manual ones in 11 and 3 cases, respectively, with no “not adequate” (score = 1) ratings. These results underscore the lack of consensus among experts. For two test patients, the model proposed a different isocenter configuration that both MPs rated higher. Automated configurations
Made with FlippingBook - Share PDF online