S2051
Physics - Image acquisition and processing
ESTRO 2026
Digital Poster 1093 Evaluating the Generalisability of a Deep Learning Synthetic CT Model for Prostate MR-only Radiotherapy Christopher Thomas 1,2 , Jonathan Wyatt 3 , Andrew King 2 , Isabel Dregely 4 , Sally F Barrington 2 , Teresa Guerrero Urbano 5 1 Medical Physics, Guy's and St. Thomas' NHS Foundation Trust, London, United Kingdom. 2 BMEIS, King's College London, London, United Kingdom. 3 Medical Physics, Northern Centre for Cancer Care, Newcastle, United Kingdom. 4 Department Computer Science, Fachhochschule Technikum Wien, Veinna, Austria. 5 Clinical Oncology, Guy's and St. Thomas' NHS Foundation Trust, London, United Kingdom Purpose/Objective: Prostate radiotherapy (RT) planning using Magnetic Resonance (MR) imaging offers increased accuracy and healthy tissue sparing over conventional CT planning, but requires synthetic CT (sCT) for dose calculation. A deep learning (DL) model was trained for synthetic CT generation from 2D T2-weighted images from a single scanner and validated using further images from the same scanner and imaging protocol. DL models can be susceptible to over-fitting to the training data and so it is important to assess the generalisability of the model using data from another scanner and imaging protocol. The aim of this study was to externally evaluate the dose calculation accuracy of the DL model when using 3D T2- weighted images from a different scanner and imaging protocol. Material/Methods: All data use and transfer was covered by ethics approval. The DL model was trained using 24 patients from Guy’s and St Thomas NHS Foundation Trust, London [1][2]. Images from 10 patients from the Northern Centre for Cancer Care, Newcastle, were used to generate sCTs. The images differed significantly from the images used for model training, by scanner (Siemens MAGNETOM Sola vs Siemens Aera), sequence (3D SPACE vs 2D TSE), voxel size (1.1 x 1.1 x 2.0mm3 vs 0.4 x 0.4 x 3.0mm3) and echo time (205 ms vs 104 ms). Dose calculation accuracy was assessed by deformably registering the sCT to the planning CT and recalculating the clinical treatment plan. Dose distributions were compared with differences in dose to planning target volume (PTV) and organs at risk (OARs) and global gamma analysis with criteria 2%/2mm and 1%/1mm. Results: Mean T2w pixel value for bladder, muscle, fat, femoral heads and prostate for [local, external] datasets were [307.0, 116.5], [41.5, 8.4], [206.6, 45.9], [135.7, 40.3], and [74.5, 14.4] respectively.
the image without implant was obtained on four ROIs at different locations close to the implants. Results: Regarding artefact reduction, PCCT images with implants at 70keV and 190keV VMI demonstrated superior performance compared to EID CT for all scenarios as lower AI and ρ -index represent less severe metal artefact, as shown in Figure 1. In the presence of titanium implant, the ρ -index of PCCT 70keV VMI is better than EID CT for 46HU at 70keV VMI and 32HU at 190keV VMI for vertebral body ROI as shown in Figure 1(a). The T3D performance of metal artefact reduction was comparable to EID CT, with only marginal differences.
For CT number accuracy, PCCT T3D and 70keV VMI showed minimal deviation in intervertebral and vertebral body ROIs for titanium and steel implants scenarios, as shown in Figure 2.
Conclusion: PCCT, particularly with 70keV and 190keV VMI images, provides superior artefact reduction compared to EID CT for titanium and steel spinal implants. For 70keV VMI, the CT number difference in the presence of implant materials was minimal. This offers a potential clinical advantage for improving the accuracy of radiotherapy treatment planning. References: 1Cammin, J. (2024). A robust index for metal artifact quantification in computed tomography. Journal of Applied Clinical Medical Physics, 25(8), e14453. doi:10.1002/acm2.14453. Keywords: Radiotherapy planning, quantitative imaging, iMAR
Made with FlippingBook - Share PDF online