3101 - A Patient-Specific 3-Dimensional Gaussian Representation Framework for Volumetric Imaging and Tumor Tracking from a Single kV Projection

Purpose/Objective(s):

Respiratory motion introduces anatomical changes that can compromise target coverage and increase dose to organs at risk. This study aims to investigate the feasibility of a novel, patient-specific method using a 3-D Gaussian representation framework to achieve real-time volumetric imaging and tumor tracking using only a single kV x-ray projection.

Materials/Methods:

A volumetric imaging neural network was developed, consisting of a projection-based conditioning module and a deformation module. The images of the mean respiratory phase were modeled using 3-D Gaussians, where each Gaussian was parameterized by its position, scale, rotation, and density. The conditioning module extracts spatial features from the input single 2-D kV projection. Driven by these features, the deformation module predicts the changes in the 3-D Gaussian parameters. Then, a differentiable voxelization process was employed to render the deformed Gaussians into 3-D voxel images representing the target respiratory phase. Following reconstruction, an auto-contouring network was applied to segment and track the target. The framework was validated on two lung cancer patients. To generate the dataset, a total of 360 uniformly distributed projection angles were simulated for each of the ten respiratory phases. The dataset was randomly split into training and validation sets at a 9:1 ratio. To simulate real clinical treatment for testing, patient-specific real-time position management (RPM) traces were used to map respiratory phases and corresponding testing projections. Image reconstruction quality was evaluated using Structural Similarity Index Measure (SSIM) and Peak Signal-to-Noise Ratio (PSNR). Segmentation and tracking accuracy were evaluated using Dice Similarity Coefficient (DSC), 95% Hausdorff Distance (HD95), and Center of Mass (COM) Root Mean Square Error (RMSE).

Results:

For volumetric reconstruction, the method achieved an SSIM of 0.991 ± 0.001 and 0.992 ± 0.002, with a PSNR of 40.83 ± 0.45 dB and 43.60 ± 1.39 dB for patients 1 and 2, respectively. DSCs between the predicted contours and ground truth were 0.95 ± 0.01 and 0.88 ± 0.02 for patients 1 and 2, respectively, with corresponding HD95 values of 2.15 ± 0.24 mm and 3.37 ± 0.59 mm. The COM RMSE in the X, Y, and Z directions were 0.14 mm, 0.18 mm, and 0.27 mm for patient 1, and 0.74 mm, 0.82 mm, and 1.04 mm for patient 2. Image reconstruction and target segmentation required only about 0.1 s and 0.2 s per frame, respectively, using an NVIDIA RTX 3090 GPU.

Conclusion:

The proposed 3D Gaussian representation framework effectively reconstructs high-quality volumetric images and tracking tumor from a single 2D kV projection. It demonstrates significant potential for real-time intrafraction monitoring and markerless tumor tracking in lung radiotherapy.