ESTRO 2024 - Abstract Book

S5332

Radiobiology - Tumour biology

ESTRO 2024

Figure 2 shows the percentage of patches per slide predicted as radiosensitive for each tumor model, ordered from the most to the least radiosensitive according to the TCD50 values [2]. We observe a considerable negative correlation between these two values (Spearman’s correlation coefficient between median percentage and TCD50: -0.87, p-value: 0.001). The CNN provides highly accurate predictions for very radiosensitive and radioresistant models, while predictions are more uncertain for the tumor models that fall in the intermediate range between radioresistant and radiosensitive.

Conclusion:

Convolutional neural networks are able to learn from H&E images to distinguish between radioresistant and radiosensitive HNSCC xenograft tumor models. Future work will involve validation on independent datasets, evaluation of model interpretability and exploring the translation to patient data. Additionally, underlying biological factors that potentially influence CNN predictions will be investigated, to provide relevant insights for future hypothesis-driven experiments.

Keywords: radiosensitivity, CNN, histopathology

References:

[1] Gurtner K, Deuse Y, Bütof R, Schaal K, Eicheler W, Oertel R, Grenman R, Thames H, Yaromina A, Baumann M, Krause M. Diverse effects of combined radiotherapy and EGFR inhibition with antibodies or TK inhibitors on local tumour control and correlation with EGFR gene expression. Radiother Oncol. 2011 Jun ;99(3) :323-30. Doi: 10.1016/j.radonc.2011.05.035. Epub 2011 Jun 12. PMID: 21665304.