ESTRO 2025 - Abstract Book

S3738

Physics - Radiomics, functional and biological imaging and outcome prediction

ESTRO 2025

1242

Digital Poster Attention-based vision classifier to predict late radiation toxicity from MR images acquired early after radiotherapy of a murine model Manish Kakar 1 , Bao Ngoc Huynh 2 , Olga Zlygosteva 3 , Inga Solgård Juvkam 1,4 , Nina Edin 3 , Oliver Tomic 2 , Cecilie Marie Futsaether 2 , Eirik Malinen 1,3 1 Department of Radiation Biology, Institute for Cancer Research, Oslo University Hospital, Oslo, Norway. 2 Faculty of Science and Technology, Norwegian University of Life Sciences, Ås, Norway. 3 Department of Physics, University of Oslo, Oslo, Norway. 4 Institute for Oral Biology, Faculty of Dentistry, University of Oslo, Oslo, Norway Purpose/Objective: Patients receiving RT for H&N cancer may develop toxicities at any stage during the course of the treatment. There is a critical need for rapid, non-invasive methods of evaluating individuals at an early stage who may suffer from toxicities. Since visual assessment of MR images acquired early after radiotherapy reveals no discernible patterns, the purpose of this study was to investigate if an attention-based vision classifier could be used to detect and predict toxicity in a murine model.

Figure 1: (A) Transmission X-ray image with irradiation field indicated by the blue line and the midline contour indicated by the green line, (B) T2-weighted MRI with the same contours outlined. The submandibular gland (C) is indicated.

Material/Methods: In this work, C57BL/6J mice (n=14) were given fractionated X-irradiation to a total dose of 66 Gy covering the salivary glands, swallowing muscles, and oral cavity (see figure 1). The animals that get this therapy experience severe toxicity both early and late[1]. In this study, we obtained T2-weighted MR images three to five days following irradiation. For comparison, a control group of mice (n = 15) got sham treatment. Using pretrained weights from torch-vision [2], we re-trained a vision transformer model on MR images to classify control and irradiated animals employing TL methodology[3]. Saliva was collected for each mouse at day 105 after irradiation. Results: With five-fold validation approach, the model achieved 69% accuracy rate on average for classification. Images with low class probabilities (misclassified) were often found towards the lateral part of the MR field-of-view (the peripheral regions of the mice). The mean classification for peripheral and central images was 0.50 and 0.71 for controls, respectively, while both probabilities were 0.66 for irradiated animals. To investigate the relation between our model predictions and late toxicity in terms reduced saliva production, we computed the mean class probability over all images (including misclassified ones) for each mouse. The prediction probability was highly correlated (r2=0.65) with late toxicity in terms of reduced saliva production in individual animals (see figure 2). In terms of explainability, one of the attention maps highlighted the irradiated region in the animals.