ESTRO 2024 - Abstract Book

S4062

Physics - Inter-fraction motion management and offline adaptive radiotherapy

ESTRO 2024

Deformable image registration (DIR) is an essential technique for various tasks in state-of-the-art radiation therapy. However, the ill-posed nature of DIR causes inherent uncertainties in the estimated deformation, e.g., due to the specific choice of DIR hyper-parameters or algorithm limitations. Understanding these uncertainties and integrating them into downstream applications such as dose accumulation is paramount for informed clinical decision-making. In this work, we propose a versatile methodology using principal component analysis (PCA) to compute the spatial distribution of the DIR uncertainty and directly integrate it into dose accumulation for adaptive radiotherapy of lung cancer.

Material/Methods:

We utilized data from 20 locally-advanced NSCLC patients receiving conventional intensity-modulated radiation therapy. Each patient received six consecutive weekly CBCT scans. DIRs between planning CT (pCT) and weekly CBCTs were used to accumulate the delivered dose. All DIRs were performed using the B-spline free-form deformation algorithm of the open-source package Elastix [1]. To derive the dosimetric impact related to DIR uncertainty, a three step procedure was implemented for each registration. First, we performed a series of 63 independent DIRs, each with slightly perturbed hyper-parameters that can noticeably affect the registration (i.e., B-spline knot spacing, bending energy penalty weight, and optimizer step size), selected around the standard settings for thoracic CT-CBCT DIR. All results were validated using expert-delineated contours available on each CBCT to ensure reasonable registrations. For each pCT voxel, the cloud of homologous points in the CBCT anatomy obtained from the DIRs reflects discrete samples of the underlying DIR uncertainty distribution. Next, we used PCA to construct voxel-specific confidence ellipsoids that contain a user-defined coverage (here: 95%) of the DIR uncertainty samples by scaling the semi-axis length according to the square root of the eigenvalue times the critical value of the chi-square distribution (Figure 1). Finally, when pulling the dose distribution to the pCT for dose accumulation, we mapped the largest and smallest dose values within the ellipsoids to estimate the upper and lower dose limits for the particular voxel. The weekly doses were accumulated, and the associated DVH uncertainty bands were calculated. We evaluated the DIR uncertainty distribution provided by the PCA model and compared the original treatment plan with the accumulated dose distribution. A DIR uncertainty map for an example patient is shown in Figure 1. The largest semi-axis length of the 95% confidence ellipsoid was 2.5±1.5 mm, 3.7±2.3 mm, and 7.2±4.8 mm for the esophagus, ipsilateral lung, and GTV, respectively. The high uncertainty around the GTV reflects the progressive tumor changes in response to radiotherapy. Overall, the DIR uncertainty was highly spatially dependent and correlated. The average confidence ellipsoid was strongly elongated, with the largest semi-axis being 5.5 and 2.5 times the length of the others. The corresponding dose-volume histogram of the example patient (Figure 1) demonstrated good agreement between planned and accumulated doses. However, the uncertainty (95% confidence) in the esophagus mean dose was 4.1 Gy, which would result in exceeding the institutional constraint of 34 Gy. The relevant institutional OAR DVH parameters (Table 1) were slightly higher for the accumulated doses compared to the original treatment plan but within the estimated uncertainty due to the DIR limitations. Average dosimetric uncertainties derived from the confidence ellipsoids were mostly within 3 Gy, but were highly patient-specific due to interplay between DIR uncertainty and dose gradient. When the DIR uncertainty was considered within the dose accumulation, institutional OAR dose constraints were violated in six (30%) cases, indicating the importance of considering the inherent uncertainty associated with DIR. The average computational time for computing the DIRs, PCA confidence ellipsoids, and performing dose search via high-resolution interpolation was around 30 min, making the approach clinically feasible for automated offline evaluations. Results: