S2296
Physics - Machine learning and AI algorithms
ESTRO 2026
Floranna Mauro 5 , Sami Aburas 5 , Lana Smiljanic 5 , Antonio Piras 5 , Carmela Di Dio 5 , Lorenzo Placidi 9 , Luca Boldrini 7 , Maria Antonietta Gambacorta 7 , Gian Carlo Mattiucci 5 , Davide Cusumano 1 1 Medical Physics Unit, Mater Olbia Hospital, Olbia, Italy. 2 ICSC, CNR, Casalecchio di Reno (BO), Italy. 3 INFN, INFN, Sesto Fiorentino, Italy. 4 ICT, Fondazione Policlinico Gemelli IRCCS, Rome, Italy. 5 Radiotherapy Unit, Mater Olbia Hospital, Olbia, Italy. 6 Department of Sciences, Università di Chieti, Chieti, Italy. 7 Radiotherapy Unit, Fondazione Policlinico Gemelli IRCCS, Rome, Italy. 8 Radiotherapy Unit, AUSL Reggio Emilial, Reggio Emilia, Italy. 9 Medical Physics Unit, Fondazione Policlinico Gemelli IRCCS, Rome, Italy
Purpose/Objective: MRI-only workflows are gaining traction in
radiotherapy as a promising alternative to traditional CT-based planning, with artificial intelligence (AI) playing a pivotal role in generating synthetic CT (sCT) images from MRI. This approach is particularly challenging in the thorax due to significant heterogeneity in electron density (ED). This study explores the impact of several key variables on the accuracy of thoracic sCTs images. In particular a novel architecture, the Adaptive Fourier Neural Operator (AFNO), was introduced, with the aim of incorporating a learnable convolutional filter that operates in the Fourier domain, enabling the model to capture frequency-specific features in MRI that are not readily accessible through spatial domain analysis. Material/Methods: This retrospective study included 122 thoracic cancer patients treated with MRI-guided radiotherapy (MRIgRT). For each patient, both 0.35T MRI and CT simulation scans were acquired under breath-hold conditions to ensure consistency. The investigation focused on three variables: (i) the size of the training dataset (34,68, and 102 cases), (ii) the effect of MRI pre-processing (raw images versus those processed with N4 bias field correction), and (iii) the choice of generator architecture, comparing U-Net, ResNet and the new AFNO. Model performance was evaluated on a separate test set of 20 patients using Mean Absolute Error (MAE) calculated on body and bones. Different configurations (Figure 1) were compared using Wilcoxon-Mann Whitney (WMW) test for paired samples. The best-performing model underwent dosimetric analysis.IMRT plans were calculated using CT and sCT as ED map and dose differences were assessed using DVH metrics. Three dose parameters (D2%, D98%, D50%) were analysed for PTV and ipsilateral lung. Statistical equivalence was tested using the two one-sided test for paired samples (TOST- P). Results: Increasing the training set size significantly improved
Conclusion: GaMMA-Surv demonstrates effective VAE-based joint fusion for OPSCC RFS prediction, with imaging and clinical features providing complementary prognostic value. Robust HPV+ risk stratification supports its potential for identifying patients suitable for biomarker-driven de-escalation strategies in personalized OPSCC management. References: 1. Saeed N, Hassan S, Hardan S, Aly A, Taratynova D, Nawaz U, Khan U, Ridzuan M, Andrearczyk V, Depeursinge A, Xie Y. A Multimodal and Multi-centric Head and Neck Cancer Dataset for Segmentation, Diagnosis and Outcome Prediction. arXiv preprint arXiv:2509.00367.2025.2. Ma B, Guo J, van Dijk LV, Langendijk JA, van Ooijen PM, Both S, Sijtsema NM. PET and CT-based DenseNet outperforms advanced deep learning models for outcome prediction of oropharyngeal cancer. Radiotherapy and Oncology. 2025;207:110852.3. Duran J, Zou Y, Vallières M, Enger SA. Beyond single-run metrics with CP-fuse: A rigorous multi-cohort evaluation of clinico-pathological fusion for improved survival prediction in TCGA. Machine Learning with Applications. 2025:100789. Keywords: Head & Neck Cancers, Outcome Prediction
Digital Poster Highlight 2708
Enhancing thoracic synthetic CT generation for MRI-only radiotherapy: evaluation of AFNO architecture and other key parameters Luca Vellini 1 , Alessandro Bombini 2,3 , Flaviovincenzo Quaranta 1 , Jacopo Lenkowicz 4 , Sebastiano Menna 1 , Elisa Pilloni 1 , Francesco Catucci 5 , Andrea D'aviero 6 , Claudio Votta 7 , Giuditta Chiloiro 7 , Martina Iezzi 5 , Francesco Preziosi 5 , Alessia Re 8 , Althea Boschetti 5 ,
Made with FlippingBook - Share PDF online