ESTRO 2024 - Abstract Book

S5236

Radiobiology - Normal tissue radiobiology

ESTRO 2024

Results:

For training/validation of the acute toxicity model, 1859 image sets were used, and 655 for testing. Leg detection accuracy reached 98.5%, with some challenging unsharp images and obscured legs as factors for non-detection. For classification, the model demonstrated good performance with 85.3% accuracy on the test set along with an average misclassification distance of 0.6. Only 13 (2.0%) images were classified more than one category off. With Krippendorff’s Alpha coefficients ranging between 0.865 and 0.951, there was high agreement between observers. The DL model achieved a higher accuracy than the majority of human observers and displayed the lowest average misclassification distance (Figure 1). For late toxicity, we imaged both hind legs of five mice with RIF=4 and two with RIF=3. The left leg served as untreated control for the treated right leg (Figure 2). In 4/5 cases with RIF=4, no subcutaneous fat was visible on the treated leg. There were also substantial differences in the leg bones, for example tibia shape and diameter. The last case with RIF=4 looked like the cases with RIF=3, with reductions in subcutaneous fat of 43%, 65%, and 30% compared to the untreated leg, and only minor bone differences.

Conclusion:

We have developed an imaging setup to systematically record and assess acute skin toxicity in the leg model. The developed DL classification model had good performance, at least as good as human observers. For RIF, we have shown that microCT imaging has potential to serve as a quantitative tool. This is especially interesting, since microCT imaging can be performed repeatedly with live animals.

Keywords: mouseleg model, acute and late toxicity, imaging