ESTRO 2025 - Abstract Book

S3788

Physics - Radiomics, functional and biological imaging and outcome prediction

ESTRO 2025

References: [1] L. Van Den Bosch et al. , ‘Key challenges in normal tissue complication probability model development and validation: towards a comprehensive strategy’, Radiotherapy and Oncology , vol. 148, pp. 151–156, Jul. 2020, doi: 10.1016/j.radonc.2020.04.012. [2] S. Marcott, K. Dewan, M. Kwan, F. Baik, Y.-J. Lee, and D. Sirjani, ‘Where Dysphagia Begins: Polypharmacy and Xerostomia’, Fed Pract , vol. 37, no. 5, pp. 234–241, May 2020. [3] M. Asif, A. Moore, N. Yarom, and A. Popovtzer, ‘The effect of radiotherapy on taste sensation in head and neck cancer patients – a prospective study’, Radiat Oncol , vol. 15, no. 1, p. 144, Dec. 2020, doi: 10.1186/s13014-020-01578 4.

2557

Digital Poster Deep learning enables accurate quantification of imaging biomarkers from intravoxel incoherent motion modelling with a clinical set of b-values Marte Kåstad Høiskar, Amalie Toftum Hop, René M. Winter, Kathrine Røe Redalen Department of Physics, Norwegian University of Science and Technology, Trondheim, Norway Purpose/Objective: With magnetic resonance imaging-guided radiotherapy (MRIgRT), quantitative imaging biomarkers can be extracted repeatedly during treatment. This opens the door for individualised response-adapted treatment based on longitudinal biomarkers during therapy. Intravoxel incoherent motion (IVIM) modelling of diffusion-weighted MRI (DW-MRI) quantifies tissue perfusion and diffusion through the parameters pure tissue diffusion coefficient D t , perfusion fraction f p , and pseudo-diffusion coefficient D p . Though IVIM parameters are promising biomarkers, long fitting time and poor accuracy because of few b-values available from clinical DW-MRI scans, currently prevent clinical use. Deep learning neural networks (DNNs) have the potential to accurately calculate IVIM parameters fast and may perform well with few b-values. Our main objective was to evaluate the performance of DNNs to predict IVIM parameters compared to conventional fitting models, with a clinical set of b-values. A second objective was to investigate longitudinal changes in IVIM parameters during radiotherapy for head and neck cancer (HNC) patients. SL ), for IVIM modelling were evaluated in silico and in patients. First, synthetic IVIM signals with SNR between 8 and 100 were generated using 4, 5 and 11 b-values, with IVIM parameters estimated with all four algorithms. Accuracy was evaluated in terms of normalised mean absolute error (NMAE). Second, for 11 patients, DW-MRI was acquired with 11 b-values before radiotherapy. IVIM parameters were estimated by all four algorithms using 4, 5 and 11 b-values. Additionally, DW-MRI with 4 b-values was acquired for 20 patients before, during and after radiotherapy. Investigation of an association between patterns of longitudinal changes in IVIM and treatment outcome is ongoing. Results: For synthetic data, the DNN SL and DNN SSL estimated more accurate IVIM parameters (lower NMAE) than the conventional algorithms with SNR < 50 (Figure 1). This was also true for SNR of 20, the SNR of the clinical HNC DW MRI data. With reduced number of b-values, the accuracy of conventional fitting algorithms was worse, while the accuracy of DNNs was maintained. For patient data, preliminary results showed that the parameter maps from DNNs stayed consistent when reducing number of b-values (Figure 2), hence IVIM analysis with 4 b-values seemed feasible for HNC patients when using DNNs. Material/Methods: Two conventional algorithms and two DNNs, one self-supervised (DNN SSL ) 1 and one supervised (DNN