ESTRO 2025 - Abstract Book

S60

Invited Speaker

ESTRO 2025

postprocessing steps, such as image registration, outlier removal or bias field correction could be enhanced by DL, either improving their performance or shortening the processing time. The latter becomes particularly relevant for online-adaptive MR-guided radiotherapy on MR-Linacs. DL could also accelerate calculating quantitative parameter maps from the source images: DL-based parameter estimation could avoid iterative pixel-by-pixel fitting algorithms, where the ability of convolutional neural networks to take signals at surrounding pixels into account can improve robustness of quantitative parameter estimation from the often-noisy source images. To ensure that the parameters are evaluated for the same tissue structure and to increase the robustness of longitudinal analyses, DL based auto-contouring and image segmentation techniques can mitigate the impact of intra- and inter-observer variability. DL could identify relevant features of the quantitative parameter maps and combine radiomic signatures with other clinical patient data to improve disease characterisation and prediction of outcomes for clinical decision making. One of the key motivations behind quantitative MRI is that the quantitative tissue parameters could be independent of the exact scanner model, field strength or system vendor. This is, however, not always the case: While the field strength dependency of MR relaxation times is well known, dependency of tissue diffusivity on the field strength was reported. Deep learning could provide an opportunity for harmonisation across MRI systems, either of the source data or the quantitative parameter maps, which is particularly relevant for international multi centre studies on quantitative imaging biomarkers. To outperform conventional methods, DL models typically require large high-quality training datasets. In the context of image reconstruction, fully sampled ground truth data from many different anatomies and diverse patient populations would be required, due to the risks of poor generalization to previously unseen anatomies, scanners and acquisition protocols. This can prove challenging for rare diseases and affect performance for patient populations that are under-represented in the training data. These concerns extend to quantitative parameter values that might not be reflected in the training data and overfitting to specific noise patterns and artefacts could degrade performance. Another obstacle to clinical acceptance is the “black box” nature of the most performant DL models, which could be mitigated by using hybrid approaches, such as unrolled networks that allow for integration of MRI physics. In the context of quantitative MRI it is sometimes not obvious, what the correct underlying model is, where model-free approaches and information criteria could be used. There is a risk of applying self-supervised DL techniques to anatomies where the trained signal model is incorrect. Another important challenge is that the commonly used criteria to assess the performance of DL models, such as mean-squared error or structural similarity might not correlate with clinical utility and ground truth values and gold standard measurements are not available in a routine clinical setting. Challenges:

Conclusion:

While deep learning excels in controlled research environments, there are many practical and scientific barriers to applications in the real-world, particularly regarding data scarcity, interpretability of deep learning models and their ongoing validation. Hybrid models, which combine deep learning with physics-based methods could be a step towards clinical implementation. In this context self-supervised learning and uncertainty quantification could overcome data scarcity and trust issues.

References:

1. Blackledge M, Messiou C. RAD Magazine 2020;46:23. URL: https://www.radmagazine.com/scientific article/using-artificial-intelligence-to-support-the-adoption-of-quantitative-mri-into-clinical-practice/ 2. Gurney-Champion OJ, Landry G, Redalen KR, Thorwarth D. Semin Radiat Oncol 2022;32:377. DOI: 10.1016/j.semradonc.2022.06.007