ESTRO 2024 - Abstract Book

S3772

Physics - Image acquisition and processing

ESTRO 2024

Keywords: CBCT, HU calibration, radiotherapy planning

References:

1. Changsheng Ma et al., Radiotherapy dose calculation on KV cone-beam CT image for lung tumor using the CIRS calibration, Thoracic Cancer 5 (2014) 68–73,

2. D.Oborska-Kumaszynska, A new calibration method of an Elekta XVI (R.5.0.2) system able to achieve superior image quality, abstract book (conference paper, EP-1721), ESTRO 36 p. 5931.

3. D.Oborska-Kumaszynska, Calibration and optimisation of a XVI system (Elekta Synergy) - Pitfalls, Fizyk Inżynier Medyczny, nr 5/2016, vol.5, p. 274-280.

4. D.Oborska-Kumaszynska, D. Northover, Optymalizacja system XVI dla protokolow klinicznych (eng. Optimisation of a XVI system for clinical protocols), Fizyk Inżynier Medyczny 4 (2016), vol.5, p. 203-215

526

Poster Discussion

Geometry-Preserving MR-based Synthetic CT Generation via Diffusion Schrödinger Bridge Models

Muheng Li 1,2 , Xia Li 3,1 , Sairos Safai 1 , Damien Charles Weber 1,4,5 , Antony John Lomax 1,2 , Ye Zhang 1

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

Purpose/Objective:

Despite the advancements in MR-based CT synthesis [1], a persistent challenge remains: the issue of data alignment during registration for model training. Training generative models mandates the use of meticulously registered MR and CT data. Due to CT and MR images being produced by fundamentally different imaging equipment, inherent discrepancies between the original datasets are substantial. This necessitates algorithmic registration, which inevitably introduces errors and misalignments. Additionally, the absence of a complete common region between CT and MR images further complicates the registration process (figure 1f&g). Conventional generative models directly predict CT images [2], which are susceptible to these registration inaccuracies present in the training datasets. The proposed methodology harnesses the principles of the Schrödinger Bridge [3]. Instead of a direct prediction, this approach focuses on fitting the transitional process between MR to CT images (figure 1a). This nuanced approach ensures the preservation of the initial modality's (MR images) geometric information during synthesis. As a result, this method effectively mitigates the noise associated with modality registration while achieving high-quality synthesis.