S589
Clinical – Head & neck
ESTRO 2026
2 Combination Innovation Department, Shanghai United Imaging Healthcare Co., Ltd., Shanghai, China
(EBV − /TLG − ), High-Low (EBV+/TLG − ), Low-High (EBV − /TLG+), and High-High (EBV+/TLG+). The primary endpoint was PFS; OS was secondary. Univariate analysis (UVA) was performed using the log-rank test, and the factors found significant in UVA were included in multivariate analysis (MVA) using complete-case Cox regression. Stage, TLG, EBV-DNA, gender, age, LDH level, combined EBV-DNA/TLG status and RT duration were analyzed. Results: Cohort was predominantly male (67/86, 77.9%), with a median age of 49 years (12–76); 80.2% had stage III–IV disease. Median follow-up was 65 months (13–170). Median PFS was 59.5 months; median OS was not reached. In UVA, EBV-DNA ( ≥ 3,500), TLG ( ≥ 200), female gender, and combined EBV-DNA/TLG status (High-High group compared to others) predicted PFS (p=0.033, p=0.010, p=0.013, and p=0.006, respectively). In UVA for OS, only TLG ( ≥ 200), female gender, and the High- High group remained significant (p=0.020, p=0.018, and p=0.007, respectively). In MVA for PFS (complete cases, n=59), only the High-High group was significant (HR=4.30 (95% CI, 1.40–13.00; p=0.010)). Similarly, in MVA for OS, only the High-High group remained significant (HR=4.90 (95% CI, 1.20–20.60; p=0.031)). In a 36-patient subgroup (Low–Low vs High–High), UVA showed stronger significance for PFS (p=0.005) and OS (p=0.004) Conclusion: A combined stratification using TLG ≥ 200 and EBV- DNA ≥ 3,500 identified the High-High subgroup as having significantly inferior PFS and OS, yielding a simple, clinically implementable pre-treatment risk tool; however, external multicenter validation is warranted. References: 1.Lin HC, Chan SC, Cheng NM, Liao CT, Hsu CL, Wang HM, Lin CY, Chang JT, Ng SH, Yang LY, Yen TC. Pretreatment 18F-FDG PET/CT texture parameters provide complementary information to Epstein-Barr virus DNA titers in patients with metastatic nasopharyngeal carcinoma. Oral Oncol. 2020 May;104:104628. doi: 10.1016/j.oraloncology.2020.104628. Epub 2020 Mar 9. PMID: 32163890. Keywords: NPC, Total lesion glycolysis (TLG), EBV-DNA
Purpose/Objective: The clinical implementation of MR-only radiotherapy is contingent upon the accuracy of synthetic CT (sCT) generated from magnetic resonance images (MRI). This study aims to evaluate a novel sCT generation method for head and neck cancer (HNC) radiotherapy that utilizes a text-guided latent diffusion model, with a primary focus on assessing its performance in terms of image quality and assessment of organs at risk (OARs) delineation consistency. Material/Methods: A retrospective dataset of 352 HNC patients was used. For each patient, data included tri-sequence MRI simulation images (T1-weighted, contrast-enhanced T1-weighted and T2-weighted MRI-Sims), a planning CT (pCT), and clinical text data (age, sex, weight, TNM stage). We developed a latent diffusion model (LDM) that integrates the tri-sequence MRI and patient- specific text to generate sCT. The model architecture comprises: (1) four separate autoencoders to extract latent features from the three MRI-Sims and the CT; (2) a conditioning control module to process and fuse the multi-sequence MRI latent features; and (3) a foundational diffusion model, conditioned on both the fused MRI latent features and a normalized 1 × 5 text data matrix, to generate the CT latent features. We conducted assessments of image quality and OARs contouring in 24 test datasets.The generated sCT was evaluated against the real pCT (rCT) using mean absolute error (MAE), peak signal-to-noise ratio (PSNR), and structural similarity index (SSIM). Further analysis involved contouring OARs on the sCT to calculate Dice similarity coefficients and the 95th percentile Hausdorff Distance (HD95). Results: On the 24 test datasets, the autoencoder achieved a full-window MAE of 9.96 1.95, SSIM of 0.98 0.01, and PSNR of 33.85 0.96.The latent diffusion model achieved a full-window MAE of 17.19 2.26, SSIM of 0.95 0.01, and PSNR of 28.09 1.12. A representative synthesis results for a 39-year-old patient weighing 61 kg with a T3N2M0 stage are shown in Figure 1. Furthermore, the metrics for automatically segmented OARs on both the sCT and rCT as shown in Table 1. The Dice of brain stem and temporal lobe were 0.89 and 0.91, which can meet clinical needs.
Digital Poster 2187
Text-Guided Latent Diffusion Model for Synthetic CT Generation from Multi-Sequence MRI of Head and Neck Cancer Guanqun Zhou 1 , Suman Zhang 1 , Ziquan Wei 2 , Zijie Mo 2 , Lecheng Jia 2 , Ying Sun 1 1 The department of Radiation Oncology, Sun Yat-sen University Cancer Center, Guangzhou, China.
Made with FlippingBook - Share PDF online