ESTRO 2024 - Abstract Book

S4413

Physics - Machine learning models and clinical applications

ESTRO 2024

[6] Soares KC, Arnaoutakis DJ, Kamel I, Rastegar N, Anders R, Maithel S, et al. Choledochal Cysts: Presentation, Clinical Differentiation, and Management. Journal of the American College of Surgeons 2014;219:1167-1180. doi: 10.1016/j.jamcollsurg.2014.04.023.

[7] Ishibashi H, Shimada M, Kamisawa T, Fujii H, Hamada Y, Kubota M, et al. Japanese clinical practice guidelines for congenital biliary dilatation. J Hepatobiliary Pancreat Sci 2017;24:1-16. PMID:28111910. doi: 10.1002/jhbp.415.

238

Digital Poster

Validation of a cycleGAN deep learning algorithm for synthetic CT generation for female pelvis

Rachael Tulip 1 , Spyros Manolopoulos 1 , Sebastian Andersson 2 , Robert Chuter 3,4

1 Northern Centre for Cancer Care – North Cumbria, Radiotherapy Physics, Carlisle, United Kingdom. 2 RaySearch Laboratories, Research and Development, Stockholm, Sweden. 3 Christie Medical Physics and Engineering (CMPE), The Christie NHS Foundation Trust, Manchester, United Kingdom. 4 University of Manchester, Division of Cancer Sciences, Manchester, United Kingdom

Purpose/Objective:

In conventional radiotherapy planning pathways, the CT scan provides both structural anatomical information and electron density information required for accurate dose calculation in the treatment planning system. The poor soft tissue contrast associated with CT images is traditionally overcome through fusion of MRI images with the radiotherapy planning CT scan bringing its own uncertainties deriving from the CT-MR registration errors. To overcome this, there has been increased interest in generating synthetic CT (sCT) datasets directly from the MRI scans [1, 2]. The purpose of this work was to generate a female pelvic model using a deep learning solution developed by RaySearch Laboratories, Sweden [3], based on generative adversarial networks (cycleGAN) to provide a way of generating the sCT. The model is trained on standard T2 spin echo scans available on all MR scanners without the need for specialist sequences or increased scan time required for acquisition of extra specialist sequences such as mDixon or ultra-short echo.

Material/Methods:

Using a research version (V11) of RayStation (RaySearch Laboratories, Sweden), we have developed a deep learning model (cycleGAN) to produce sCT datasets for gynaecological patients [4]. The model was based on 30 MR-CT paired datasets from patients previously treated in our centre. A further 10 patients were used to validate the model. DVH in terms of dose volume histogram parameter comparisons, HU mean average error of HU, gamma analysis (MICE toolkit v 1.1.3) [5] and direct dosimetric end to end testing in a novel 3D printed anatomic phantom (RTSAFE, Athens, Greece) [6] using a PTW 0.125cc semiflex chamber was performed . The comparisons were all made between a CT deformed to the geometry of the MR dataset and the sCT generated from that MR dataset. All patient contours and