ESTRO 2024 - Abstract Book

S3773

Physics - Image acquisition and processing

ESTRO 2024

By addressing this challenge, we aim to elevate the potential of MR to CT image synthesis approaches, especially in the realm of on-board imaging guidance.

Material/Methods:

We propose the Diffusion Schrödinger Bridge Models (DSBM) for the generation of synthetic CT images from MR data. The DSBM operates on the principle of constructing diffusion bridges between MR and CT image distributions. It diverges from traditional diffusion methods [4] that initiate generative processes with Gaussian white noise. Instead, DSBM starts from MR images, utilizing their structural information to build informed diffusion bridges for efficient CT synthesis. Mathematically, DSBM is grounded in the Schrödinger Bridge theory. It solves the optimal transport problem between two probability distributions, in this case, MR and CT images, by minimizing the Kullback-Leibler divergence subject to certain constraints. The model employs a set of linear Stochastic Differential Equations (SDEs) to facilitate this process, aided by a score network that parametrizes the gradients of the log densities of the distributions involved [5]. The score network was implemented using a 3D U-Net model [6], processing input data segmented into patches of size 128 × 128 × 4. Data augmentation, involving slicing training data at random angles, was employed to enhance model learning and generalization. The DSBM approach was trained on a dataset of skull MR images from 77 patients, and tested on a separate set of 10 cases. For validation, we conducted image-level evaluations using Mean Absolute Error (MAE) and Pearson correlation. Additionally, dosimetry evaluations with proton dose calculations verified the method's accuracy in proton therapy scenarios. From the evaluation of the 10 distinct cases, the MAE values for our method were recorded as 123.18 ± 10.93 HU for the body region, 65.69 ± 6.70 HU for soft tissues, and 354.80 ± 43.70 HU for bones (figure 1b-e). To validate the geometry-preserving capabilities of our approach, we assessed the Pearson correlation coefficient [7] between the synthesized CT images and the original MR modality, and then compared these metrics with those obtained between the ground truth CT and MR images. While the Pearson correlation coefficient between the ground truth CT and MR was 0.8311 ± 0.0493, our synthesized CT achieved a value of 0.8433 ± 0.0467. These enhanced metrics emphasize our method's superior ability in preserving and translating geometric details from MR to CT (figure 1f-h). Proton-based dosimetry assessments, using reference doses calculated from the ground truth CT plans, also revealed commendable results with a 1%/1 mm gamma pass rate of 96.49% ± 3.25% and a 2%/2 mm rate of 98.93% ± 1.44%. The relative error for D95 was 2.01 ± 2.93% of the CTV (figure 2). Results: