S2465
Physics - Radiomics, functional and biological imaging, and outcome prediction
ESTRO 2026
training consisting of a 1D convolutional layer followed by flattening and two fully connected layers, (2) slice- wise training consisting of a 2D U-net with skip connections, and (3) volume-wise training consisting of a 3D U-net with skip connections (Fig. 1). Concordance correlation coefficients (CCC) between the predicted cellularity maps and the ground truth were compared across the three network structures and a two-species mechanistic reaction-diffusion model of tumor growth following a similar data assimilation scheme in a previous study by our team [3] (Fig. 2).
Conclusion: Our proposed neural network prediction pipeline yielded accurate forecasts of tumor growth, outperforming a mechanistic model integrating phenomena of cell proliferation, diffusion and reaction to treatment, thus highlighting the strengths of data- based approaches in cancer modeling at the cost of model interpretability. These findings point to the potential of combining neural networks and biology- based processes to optimize predictions of tumor growth while maintaining interpretability. References: 1. La Rosa, A., et al. (2025). Temporospatial tumor dynamic changes in glioblastoma during radiotherapy. Journal of Neuro-Oncology, 174(4), 493–501. https://doi.org/10.1007/s11060-025-04993-92. Ocanto, A., et al. (2024). MR-LINAC, a new partner in radiation oncology: Current landscape. Cancers, 16(270). https://doi.org/10.3390/cancers1603. Miniere, H. J. M., et al. (2025). A data assimilation framework for predicting the spatiotemporal response of high-grade gliomas to chemoradiation. BMC Cancer, 25, 1239. https://doi.org/10.1186/s12885-025-14557-3 Keywords: Glioma, modeling, neural network
Results: A total of 21 patients of median age 60 (range: 26-75, M:F = 13:8 were included). The median CCC for the three network structures were: (1) 0.97, (2) 0.96, and (3) 0.94 compared to 0.91 for the mechanistic model. Network structure (1) either performed similarly (p > 0.05, Tukey’s HSD) or outperformed (p < 0.05) structures (2) and (3). Both network structures (1) and (2) either performed similarly or outperformed structure (3) and the mechanistic model.
Made with FlippingBook - Share PDF online