ESTRO 2023 - Abstract Book

S2104

Digital Posters

ESTRO 2023

nodal irradiation. For radiotherapy including axillary lymph nodes, doses to oesophagus were substantially higher with average Dmean/Dmax of 10.2 Gy/29.4 Gy. Conclusion This study showed higher oesophagus exposure in radiotherapy including sub-clavicular nodes when compared to breast/chest wall radiotherapy alone. The mean Dmean oesophagus was 9.3 Gy in sub-clavicular nodal radiotherapy, which may increase threefold the risk of oesophageal cancer according to recent evidence. Therefore, oesophagus contouring and sparing should be routinely considered in nodal radiotherapy planning as it may substantially reduce the risk of radiation induced oesophageal cancer.

PO-2337 Evaluation of AI based CT auto-contouring on synthetic CTs as an indicator for synthesis quality

K. Shreshtha 1 , M. Rigo 2 , F. Alongi 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 AI based auto contouring tools for OARs have become a norm in radiotherapy workflow owing to their high quality delineations at a fraction of time. However, these tools are known to have sub-optimal performance when subjected to out-of-distribution data. We leverage this property of AI systems to assess the quality of AI based synthetic CTs by subjecting the synthesised images to a commercially available AI based auto contouring tool. The auto contouring performance on the synthetic CTs show: • How well do these synthetic CTs match the distribution of real CTs? • Whether anatomical structures are preserved when translating from MRI to CT. 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. The generated images were then contoured using a commercially available CT auto-contouring tool which was trained on thousands of CT scans from multiple centres and scanners but not on synthetic CTs. The quality of organ contours was then analysed against the clinical contours on the T2w pelvis MRIs using four key metrics: dice coefficient (DSC), surface dice coefficient (surface DSC), average symmetric surface distance (ASSD) and 95% hausdorff distance (HD95). Results High agreement for both synthetic CT algorithms was observed for the femoral heads and bladder while the smaller organs such as the penile bulb and seminal vesicle show lower performance as shown in Table 1. Figure 1 presents a qualitative assessment of the results generated by the two algorithms.

DSC

Surface DSC

ASSD

HD95

Vendor A Vendor B Vendor A

Vendor B Vendor A Vendor B Vendor A Vendor B

Anal Canal

0.652 0.937

0.345 0.818 0.810 0.269 0.448 0.755 0.746 0.471

2.170 1.159 0.627 3.215 1.933 1.407 0.741 2.600

6.029 5.988 2.134 8.646 7.049

0.639 0.901 0.949 0.470 0.815 0.851 0.952 0.554

0.220 0.664 0.767 0.248 0.357 0.692 0.786 0.418

2.541 1.879 0.682 3.323 2.078 1.656 0.655 3.088

7.134 8.457 2.123 8.905 6.721

Bladder

Left Femoralhead 0.953

Penile Bulb

0.486 0.829 0.868

Prostate Rectum

11.358 12.382

Right Femoralhead 0.947

2.078

2.379

Seminal Vesicle 11.217 12.936 Table 1: Comparison of organ contours generated by AI based auto-contouring on synthetic images generated by two AI based synthetic ct generation algorithms from T2w MRIs on a test cohort of 29 patients. The contours are compared using mean dice coefficient (DSC), surface dice coefficient (surface DSC), average symmetric surface distance (ASSD) and 95% Hausdorff distance (HD95) metrics averaged over the entire test cohort. 0.639

Figure 1: Qualitative results on a randomly selected patient from the test cohort.

Conclusion