ESTRO 2025 - Abstract Book

S3782

Physics - Radiomics, functional and biological imaging and outcome prediction

ESTRO 2025

Conclusion: This study showed that the prediction of taste loss can be achieved by using deep learning models, particularly transformer-based models that account for global features within comprehensive 3D information of dose distribution, CT-images and organs-at-risk segmentations. With further optimization, this new 3D modelling approach has the potential to compete with conventional NTCP models in predicting the complex toxicity of radiation-induced taste loss in HNC patients.

Keywords: NTCP modelling, Deep learning, Head and neck

References: [1]

L. Van den Bosch et al., “Comprehensive toxicity risk profiling in radiation therapy for head and neck cancer: A new concept for individually optimised treatment,” Radiotherapy and Oncology, vol. 157, pp. 147–154, Apr. 2021, doi: 10.1016/j.radonc.2021.01.024.

2409

Digital Poster Exploring MR-Linac data using classic neuroimaging fMRI analysis Peter J Koopmans, Rene Monshouwer, Jeroen Findhammer, Robert Jan Smeenk, Marcel Verheij, Erik van der Bijl Radiation Oncology, Radboud University Medical Center, Nijmegen, Netherlands Purpose/Objective: Biomarker extraction is often performed on pre- and post-treatment MRIs aiming to predict treatment efficacy or toxicity. If daily MR-Linac imaging is to provide added value, it requires that changes can be observed during treatment. We investigate the merits of an exploratory method rooted in neuroscience fMRI, to complement traditional contour-based approaches to detect temporal trends in MRL. Material/Methods: Prostate patients (N=96) received 5x7.25 Gy over two weeks. We acquired four T 2 -weighted volumes per fraction. Statistical Parametric Mapping has been in use in PET and fMRI neuroimaging for over 35 years [1] , designed to detect possible neuronal activity and its unknown location. Our approach: 1. Deformable registration of all volumes to a single representative volume using Elastix [2] . 2. Linear fit to each patient’s timecourse, for each voxel. Note this is also sensitive to non-linear trends (e.g. exponential decay). 3. A t -test on the slope coefficients to discover in which voxels the across-patient distribution clearly deviates from a null distribution. Clusters of active voxels were selected manually and labelled. 4. Timecourse extraction for each label (i.e. no longer ‘imposing’ linear behaviour). Additionally, we applied the same registrations to patients’ dose maps and compare the average dose to the t -score pattern. Results: Figure 1 shows the average T 2 w volume, t -score map, label delineations and their timecourses. Most positive signal changes level off after the fourth fraction and are likely oedema related. Peripheral prostate intensity decreases after fraction #2. The negative bladder change is still under investigation.