ESTRO 2024 - Abstract Book

S5234 ESTRO 2024 We selected a radiation dose of 75Gy using a 5mm size of collimator to simulate radiation induced pulmonary fibrosis. The study is the first to compare two different size of collimators and different doses of radiation in a small animal model. The level of cytokine and immunoglobulin imply that evaluation of novel anti-fibrotic drugs for radiation induced lung fibrosis should be performed using corresponding experimental models. Radiobiology - Normal tissue radiobiology

Keywords: Radiation induced pulmonary fibrosis, mice model

113

Digital Poster

Image and deep-learning based toxicity assessment of an established radiobiology mouse-leg model

Morten Sahlertz 1,2 , Line Kristensen 2,3 , Navrit Bal 1,2 , Danny Mortensen 4 , Folefac Charlemagne Asonganyi 3 , Priyanshu Sinha 3 , Maria Bech Arnoldus 3 , Dorthe Grand 3 , Trine Tramm 1,4 , Per Rugaard Poulsen 1,2 , Brita Singers Sørensen 1,2,3 , Jasper Nijkamp 1,2 1 Aarhus University, Clinical Medicine, Aarhus, Denmark. 2 Aarhus University Hospital, Danish Center for Particle Therapy, Aarhus, Denmark. 3 Aarhus University Hospital, Experimental Clinical Oncology, Aarhus, Denmark. 4 Aarhus University Hospital, Pathology, Aarhus, Denmark

Purpose/Objective:

Since the 1970´s, a range of radiobiological studies have used a model system, where mice are irradiated on one hindleg, and both short- and long-term toxicity can be studied under varying conditions. To date, this model has been used in dozens of studies, investigating hyperthermia, fractionation, FLASH, etc. For acute toxicity assessment, visual inspection of skin damage is performed up to 30 days after irradiation. Late radiation-induced fibrosis (RIF) can be assessed with a leg flexibility assay up to one year after irradiation. These toxicity assessments are observer based, needing trained staff, have observer variation challenges, and do not allow for retrospective inspection of data. The goal of this study was to develop imaging based objective scoring systems for acute- and late-toxicity assessment for the mouse-leg model.

Material/Methods:

For acute toxicity assessment, a camera setup was developed with three 8-megapixel cameras which can capture a 360-degree view of the mouse leg. The setup was prospectively used during follow-up of four mouse cohorts (40 mice each) investigating proton and electron FLASH treatment. Data from three cohorts were used to develop a 2-step deep-learning (DL) classification model. The first step was to detect the mouse leg in the images using YOLOv8, and subsequently crop the images. Secondly, a ConvNeXt-based DL classifier was trained to categorize the skin toxicity, using the clinical scores as ground truth. Clinical scores were on a scale of 0 (no toxicity) to 3.5 (severe toxicity) in steps of 0.5. Data augmentation was employed in combination with a class-distance loss function to add penalty when the prediction is further from the ground truth. Final scores were based on a 5-fold cross-validation majority vote. The fourth cohort of mice was used as test-set, reporting accuracy, and average misclassification distance. Subsequently, we conducted an observer study, to investigate inter-observer variation when using the imaging data for toxicity scoring, instead of direct visual mouse inspection. For this part, 80 image-sets were selected (10 in each toxicity category) from the four mouse cohorts (images were excluded from the DL training sets). Five