ESTRO 2024 - Abstract Book

S3794

Physics - Image acquisition and processing

ESTRO 2024

737

Proffered Paper

Towards AI-enabled minimum dose CBCT-based synthetic CT: dose calculation and contouring accuracy

Yan Chi Ivy Chan 1 , Minglun Li 1,2 , Adrian Thummerer 1 , Katia Parodi 3 , Claus Belka 1,4,5 , Christopher Kurz 1 , Guillaume Landry 1 1 LMU University Hospital, LMU Munich, Department of Radiation Oncology, Munich, Germany. 2 Lueneburg Hospital, Department of Radiation Oncology, Lueneburg, Germany. 3 Ludwig-Maximilians-Universität München, Faculty of Physics, Department of Medical Physics, Garching b. München, Germany. 4 German Cancer Consortium (DKTK), partner site Munich, a partnership between DKFZ and LMU University Hospital Munich, Munich, Germany. 5 Bavarian Cancer Research Center, (BZKF), Munich, Germany

Purpose/Objective:

Daily cone-beam computed tomography (CBCT) scanning in image-guided radiotherapy exposes patients to additional radiation and secondary cancer risk. Reducing imaging dose remains challenging due to image quality deterioration. Despite the success of artificial intelligence (AI)-enabled full dose CBCT-to-CT translation, only few studies explored the possibility of using low dose CBCT. There is a scarcity of systematic investigation of the maximum dose reduction that could be offered by AI. This study aims at identifying the minimum achievable dose without loss of accuracy using two AI algorithms: 1) cycle-consistent generative adversarial network (cycleGAN) and 2) contrastive unpaired translation (CUT) [1], considering all metrics pertinent to CBCT-guided adaptive radiotherapy: patient positioning, image quality, organs-at-risk (OAR) contouring accuracy and volumetric modulated arc therapy (VMAT) dose calculation.

Material/Methods:

CBCT data of 41 prostate cancer patients (120 kV tube voltage, exposure time 20 ms, X-ray tube current 20 mA per projection) acquired at a linear accelerator were under-sampled from the full dose CBCT FS (350 projections) to create low dose CBCTs of 25% (90 projections), 15% (50 projections) and 10% (35 projections). A virtual CT (vCT) mapping the corresponding planning CT (pCT) onto the CBCT FS using a dedicated deformable image registration algorithm was used to evaluate the efficacy of an additional paired loss term in the trainings. In the cycleGAN implementation, a residual skip connection was added for both generators to attain higher anatomical fidelity. For CUT, cycle-consistency loss is replaced by patchwise contrastive loss. Image patches of corresponding locations are compared to maximize mutual information. Images are passed through the encoder network of the generator and a two-layer multilayer perceptron which allows the model to project patches to a shared feature space. Both models were trained using 4 fold cross-validation (33 patients) to allow using the median of the 4 models as output. Once trainings were finished, the generators for every dose level were applied to convert CBCTs of 8 test patients slice-by-slice into synthetic CTs (sCT). To infer the dosimetric accuracy, VMAT plans were optimized on a reference intensity-corrected full dose CBCT cor [2] and recalculated on sCTs in a treatment planning system. Mean absolute error (MAE), structural similarity index measure (SSIM) and peak signal-to-noise ratio (PSNR) were computed with reference to CBCT cor . For positioning accuracy, sCT images were rigidly registered to the pCT and compared to CBCT FS -to-pCT transformation. To determine