ESTRO 2024 - Abstract Book

S4168

Physics - Intra-fraction motion management and real-time adaptive radiotherapy

ESTRO 2024

1 Paul Scherrer Institut, Center for Proton Therapy, Villigen, Switzerland. 2 ETH Zurich, Department of Computer Science, Zurich, Switzerland. 3 ETH Zurich, Department of Physics, Zurich, Switzerland. 4 University Hospital of Zurich, Department of Radiation Oncology, Zurich, Switzerland. 5 Inselspital, Bern University Hospital, University of Bern, Department of Radiation Oncology, Bern, Switzerland

Purpose/Objective:

Intra-fraction motions are commonly estimated using deformable image registration (DIR). However, the conventional voxel-based methods can compromise accuracy, especially for large changes, affecting processes like dose accumulation and 4D dose optimization. Therefore, we introduce a novel way of representing the registration process, conceptualizing it as a continuous flow in both spatial and temporal realms. This innovative model effectively addresses the inherent limitations of voxel-based discretization, while offering a robust solution to challenges posed by large deformations. Most significantly, the sliding boundary problem, a challenge for classical methods like B splines [1], can be solved effectively.

Material/Methods:

To address the spatial discretization of voxel-based approaches, we incorporated the end-to-end (E2E) Implicit Neural Representation (INR) [2,3] to train a multilayer perceptron (MLP) network. This network maps a 3D coordinate to its corresponding deformation vector (∆x,∆y,∆z), shown in Fig1. Unlike voxel-based methods that offer piecewise continuity using bilinear interpolation, our approach ensures a second-order spatial continuity. Additionally inspired by LDDMM [4], we decompose large deformations (LDD) into smaller, more manageable movements. The network was designed to predict a velocity vector (vx,vy,vz) based on an input coordinate and timestep t within [0, 1]. By modeling Velocity Fields (VF) instead of the Deformation Vector Field (DVF), we facilitate the infinite decomposition of deformation steps, allowing for comprehensive temporal continuity in flow modeling. DVFs are subsequently integrated using the Euler-Maruyama method [5] over the VFs within the time span [0, 1]. This explicit flow modeling ensures that motion is seamlessly represented across both spatial and temporal dimensions. Our algorithm utilizes a 5-layer MLP for the INR, each layer housing 256 hidden states. For every 4DCT case, without relying on pre-training datasets, we optimized the networking using the normalized cross-correlation between the deformed moving and stationary images. We incorporated Jacobian regularization in E2E to ensure that the determinant of the Jacobian matrix remains near 1 across all locations. Furthermore, to maintain trajectory smoothness, we minimized the norm of each velocity vector. Notably, even though our model employs forward integration, achieving a backward trajectory is straightforward by inverting the integration direction, setting it apart from traditional methods which necessitate double optimisation for both directions. We evaluated our proposed method using 10 lung 4DCT cases from the DIR-Lab dataset [6]. The classic B-splines method was utilized as the baseline for method comparison. Motion was extracted between the End-Exhale-Phase (moving) and End-Inhalation-Phase (fixed). The DIR performance was evaluated and compared for the Target Registration Error (TRE) of the landmarks in lung, the Dice Coefficient for ribcage, and the Mean Absolute Error (MAE) for both ribcage and body regions.