Imaging is an integral part of modern radiotherapy (RT), from treatment planning to image-guided and adaptive RT as well as response assessment. Although computed tomography (CT) and cone-beam CT (CBCT) are still most commonly used for RT planning and on-board imaging, magnetic resonance imaging (MRI) is rapidly gaining ground, particularly with the advent of the MR-guided RT (MRgRT) systems.1 Advantages of MRI include improved soft-tissue contrast compared to (CB)CT and its ability to generate a large set of image contrasts by adapting acquisition settings.2 The most straightforward implementation of MRI is using conventional anatomical MRI as a replacement for the (CB)CT images for treatment planning and online image-guidance before and during each RT fraction. However, besides anatomical imaging, MRI also allows for quantifying microstructural processes, such as perfusion, diffusion, elasticity and local spin-relaxation properties, using quantitative MRI (qMRI).3 These properties are closely related to the local tumour microstructure, such as tumour cell density (with diffusion-weighted MRI),4 vascularity (perfusion MRI),5 stiffness/stroma (with elastography) and hypoxia (with relaxometry).6 As many of those qMRI parameters were shown previously to correlate to outcome,7, 8, 9, 10, 11, 12, 13 future RT personalization strategies might strongly rely on qMRI-based interventions.

Hence, qMRI has many potential uses within the RT workflow.3 The improved tumour to surrounding tissue contrast obtained by qMRI can help improve tumour delineation for treatment planning.14,15 Furthermore, qMRI can be used for treatment stratification, including the determination of optimal dose16,17 or even optimal local dose in the form of dose painting.18 Finally, qMRI can be used for response assessment and treatment adaptation throughout RT.7,19,20

As many aspects of RT, also qMRI has seen a massive increase in the use of deep learning.21 Neural networks can, for instance, be used to help with contouring, either to retrieve qMRI parameters22 from the region of interest (ROI), or as a basis for an RT plan.23,24 They can further be used to enhance parameter map quality by improving parameter estimation.25 Furthermore, deep learning can help generate synthetic CT26 that can be used for MR-only planning or daily planning on MRgRT systems. Finally, deep learning can help with the online adaptive workflow, where an automated decision-making workflow is essential to include the daily qMRI information into the treatment plan.

In this paper, we discuss the current state of deep learningfor qMRI and highlight its potential. For this work, we limited our definition of qMRI to MRI techniques that result in parameter maps that quantify underlying microstructural properties. This definition excludes quantitative features obtained from images, such as tumour size, tumour motion and radiomics features. On the other hand, it includes synthetic CT, in which we use MRI to obtain quantitative maps of electron density. This article will first discuss the use of deep learning for contouring of qMRI data, secondly introduce current deep learning approaches to estimate qMRI parameters, thirdly provide an overview about deep learning techniques to estimate synthethic CT based on MRI data and finally summarize the current and potential future use of those aspects for RT planning and delivery.