ESTRO 2024 - Abstract Book

S3763

Physics - Image acquisition and processing

ESTRO 2024

the need for CT during patient simulation can be desirable but not straightforward, given that MRI does not provide electron density maps [1].

These challenges have driven the development of synthetic CT (sCT) generation techniques, leveraging machine learning and image synthesis to predict CT from MRI and CBCT.

Material/Methods:

We organized the SynthRAD2023 Grand Challenge, hosted at https://synthrad2023.grand-challenge.org/, to promote advancements in sCT generation. Participants were tasked with developing fully automated sCT algorithms running within 20 min for each case. The challenge included task 1, MRI-to-sCT for MRI-only radiotherapy, and task 2, CBCT to-sCT for adaptive radiotherapy. The challenge dataset consisted of images from 1080 pelvis and brain cancer patients treated at three Dutch University Medical Centers, featuring paired MRI/CT or CBCT/CT [2]. The dataset was divided for each task into training (360), validation (60), and test phases (120), ensuring equal division of the three centers in each subset, and is publicly available [3]. Training and input of the validation data were shared with the participants, allowing participants to refine their algorithms and submit the obtained sCT, evaluating the image similarity to CT. In the test phases, the participating teams submit their algorithm to be run on the hosting platform, obtaining sCT. The evaluation metrics included image similarity and dose-based assessments for proton and photon plans. Participants' sCT images were evaluated using image similarity metrics: mean absolute error (MAE), peak signal-to noise ratio (PSNR), and structural similarity index measure (SSIM) during the validation and test phases. During the test phase, dose assessments were conducted, recalculating photon and proton treatment plans previously optimized on CT with matRad [4] (Fig. 1). The dose accuracy was evaluated using the dose MAE, a DVH-based metric, and 2 mm/2% gamma pass rate analysis on regions with >10% of the prescribed dose. Participants were ranked based on all the normalized metrics combined after excluding submissions with at least one of the image similarity metrics inferior to sCT comprising bulk water-filled body contour. To identify trends between architectures and performances, submissions were analyzed according to architecture features, e.g., model and spatial configuration (2D, 3D, models, etc.). Moreover, Kendall’s tau correlation was calculated between image similarity and dose metrics to identify whether image similarity may forecast dose accuracy.