ESTRO 2023 - Abstract Book

S2114

Digital Posters

ESTRO 2023

The dose accumulation proved that a daily ATS treatment is deviating minimally from the initial clinical intent of the reference plan.

Figure 1: MrgRT WorkFlow

Conclusion The daily ATS workflow for HeadNeck and GBM cases has proven its feasibility both in terms of daily accuracy as well as procedural times with Elekta Unity, showing that the initial clinical intent can be adapted and controlled day by day.

PO-2350 Evaluation of AI based synthetic CT generation solutions for T2w pelvis MRIs

K. Shreshtha 1 , F. Alongi 2 , M. Rigo 2 , R. Giuseppe Pellegrini 3 , T. Roque 4 , N. Paragios 5,6 , R. Ruggieri 2

1 Therapanacea, Artificial Intelligence, Paris, France; 2 Sacro cuore - don Calabria IRCCS, Radiation Oncology Department, Negrar della Valpolicella, Italy; 3 Elekta AB, Medical Affairs, Stockholm, Sweden; 4 Therapanacea, Research & Partnerships, Paris, France; 5 Therapanacea, CEO, Paris, France; 6 CentraleSupelec, University of Paris Saclay, Center for Visual Computing, Gif-sur-Yvette, France Purpose or Objective Recent advances in deep learning image to image translation techniques has seen a surge of AI based synthetic CT generation tools from MRIs for MR-guided and MR-only workflow. These techniques try to match the intensity distribution of planning CTs while maintaining the anatomical context from the MRIs. In this study we evaluate two such synthetic CT generation methods on multiple image similarity metrics to assess their generation quality. Materials and Methods Synthetic CTs were generated from T2w pelvis MRIs using two AI based synthetic CT generation tools for a test cohort of 29 patients. In order to compute the imaging metrics on these synthetic CTs the planning CT was deformably registered to the T2w MRIs using organ contours on the two images as control points for accurate organ level alignment between the CT and the MR. Mean absolute errors at the organ level and structural similarity (SSIM) and peak signal to noise ratio (PSNR) on the image level were then computed between the synthetic CTs and the deformed planning CT. Results Both algorithms show similar results on image similarity metrics, PSNR and SSIM. Organ level MAEs show that vendor B performed better on soft tissues while vendor A had superior performance on bones. Table 1 shows the results for the image similarity metrics over the test cohort for both algorithms and Figure 1 shows the distribution of mean absolute errors for each of the two models across the test dataset for multiple organs.

SSIM PSNR Vendor A Vendor B Vendor A Vendor B

Mean

0.825

0.833 0.018 0.779 0.825 0.834 0.848 0.860

23.857 24.839

Standard Deviation 0.018

1.264

1.490

Minimum

0.771 0.814 0.828 0.839 0.850

19.144 19.526 23.444 24.406 24.326 25.389 24.783 25.779

1st Quartile

Median

3rd Quartile

Maximum 25.013 26.382 Table 1: image similarity metrics for synthetic CTs generated from T2w MRIs with the two AI based algorithms, over the test cohort of 29 patients.