Diffusion-weighted magnetic resonance imaging (DW-MRI) is used to measure the mobility of water molecules in the microscopic environment of tissues. DW-MRI is very sensitive to cellular density, proliferation rate and cellular permeability (8, 9), and can reveal microscopic details of normal and diseased tissues. The sensitized signal is modelled by the amount of diffusion weights, called b-value. The amount of diffusion existing in the tissue, predominantly in the extracellular space (10, 11), is given by the apparent diffusion coefficient (ADC) (12). ADC map is an MRI image that shows better diffusion than conventional DWI (1). ADC map can be measured/defined by DW-MRI with at least two b-values (1). Darker areas in ADC images represent smaller magnitude of diffusion. Lower ADC values indicate slower water infusing where malignant tumour appears (13). Therefore, ADC and cellular density have an inverse correlation relationship. Several studies estimated the relationship between ADC and cellular density for different tumour sites. Ginat et al. performed a histological study and achieved the relationship between ADC and cellular density for chordoma (14), while Gupta et al. estimated the relationship for GBM (15).

The restricted spread of water has been considered as an indicator of solid tumours. However, in DW-MRI with b values of 0 and 800-1000s/mm², ADC values are usually unpredictable because of the mixture of micronecrosis, normal tissues, high-cell tumors and oedema, as shown in Figure 1 . ADC values of a tumor are sometimes higher than those in normal tissues. Researchers used the minimum ADC value to determine its prognostic value (16–19). However, the area of malignant tumours is usually larger than the area with the minimum ADC value. Therefore, high b-value DW-MRI (e.g., 3000-4000s/mm²) are used to eliminate the signal from edema (20–22). Pramanik et al. showed that the hypercellular subvolume (HCV) of GBM determined by high-b value (3000s/mm²) DWI can predict progression-free survival (PFS) (23). About 40% of HCV may exceed the area of the traditional high-dose volume, which suggests a potential biological target with more radiation dose (23). Cao et al. showed an example of DW-MRI with two different b values. The lesions were more obvious in the high b-value DW-MRI (b=3000s/mm²) than the conventional b-value DW-MRI (b=800-1000s/mm²) (24).

Hamstra et al. (25) and Moffat et al. (26) provided information for individual patient’s adaptation to radiotherapy. Patients usually have higher overall survival (OS), whose voxel ADC values are changed dramatically after 3 weeks of radiotherapy (25, 26). However, when the tumor grows or shrinks, the paired images obtained before and during radiotherapy must consistently produce high-quality registration (27). ADC can be used to calculate tumour control probability (TCP) to analyse patient-specific characteristics. A recent study showed that MRI-driven cellular density can enhance TCP value differences in patients (28). Buizza et al. evaluated DW-MRI for modelling TCP in skull-base chordomas, which has enrolled in CIRT protocol (29). The aforementioned methods facilitate personalized and optimized treatments. The main limitation of these studies lies in uncertainties that inherently affect the relationship between ADC and cellular density. Moreover, single-shot echo planar imaging (EPI) pulse sequence is used in DW-MRI, which is very sensitive to geometric distortion. To reduce the geometric distortion due to EPI, multi-shot EPI, read-out segmented multi-shot EPI, and high parallel imaging factor are used. To achieve the golden standard, tumor target defined by high b-value DWI and parameter response curves therefore requires pathological verification.

MR-Spectroscopic Imaging (MRS) uses radio-labelled glucose and methionine to identify high-risk regions in large tumours. In MRS, active tumors exhibit at areas with high resonance in the choline spectral peak and a low NAA (N-acetylaspartate). Figure 2 shows an example of MRS on post-operation MRI with choline/NAA. In other words, creatine resonance correlating with high choline/NAA, or choline/creatine ratios versus low ratios for areas of inactivity (30–37). Graves et al. analysed 36 patients with recurrent high-grade gliomas, who are treated with Gamma Knife stereotactic radiosurgery (SRS). Patients in high-risk regions of the SRS target had an improved survival rate versus those with MRS high-risk regions outside the SRS target (38). Croteau et al. studied 31 patients whose high-grade gliomas were resected after conventional MRI and MRS and found that MRS can accurately define the tumor boundary using histopathologic correlation (36). Moreover, Pirzkall et al. presented a pre-treatment analysis of 34 patients with high-grade gliomas (39) and found that high-risk regions defined by MRS were significantly smaller than regions in conventional T2-wighted imaging. Thanks to MRS, more normal areas in the brain can be distinguished from the tumour, thereby, reducing side effects (39).

Einstein et al. illustrated the procedure of Gamma Knife SRS incorporated with MRS functional imaging. Einstein et al. conducted the first prospective Phase II trial using MRS-targeted SRS for patients treated with GBM. In this work, the value of Cho/NAA> 2 was used as the area that is required to boost doses (40). However, the value of the Cho/NAA has no consensus to define the tumour area. The metabolic abnormality of Cho/NAA often exceeds the enhancement range of the lesion, and sometimes even exceeds the abnormal range of FLAIR (37, 41, 42). In some cases, the tumor recurs where the Cho/NAA is normal (41). Therefore, this technique has been discussed for several decades but not been transferred into clinical practice. Other obstacles include long acquisition, low spatial resolution, low robustness of spectral acquisition, and how to consistently obtain high-quality spectral images (1, 42, 43).

Perfusion MRI is also a widely used method for tumour characterisation and diagnosis (1), which includes dynamic susceptibility contrast (DSC) and dynamic contrast enhanced (DCE) MRI. When using Perfusion MRI, patients will get injection of gadolinium-based agent during continuous image acquisition. Gadolinium contrast agents can decrease the T1 relaxation time. Therefore, the distribution within the patient can be studied by continuously acquiring T1-weighted images.

In particular for brain tumor, modelling the distribution of contrast agent in tissue allows us to quantify vascular leakage, cerebral blood volume (CBV), mean transit time and cerebral blood flow (CBF) (43–45). As tumor keeps growing, new blood vessels will appear. CBV, CBF and vascular leakage will be increased. Thus, CBV and CBF can be used as prognostic biomarkers to predict OS and progression-free survival PFS for tumours (43, 44, 46–48). Law et al. suggested that the mean relative CBV > 1.75 can be used as the threshold of low progression time for low grade and high grade GBM (43). As another biomarker, Gd-diethylenetriamine pentaacetic acid (Ktrans) can be used to quantify the leakage of vascular for OS prediction (49). Using the mean or median of CBF and CBV in the entire tumor volume has certain limitations for GBM, since it is a highly heterogeneous tumor which reduces the sensitivity of indicators for evaluating efficacy. Several methods have been proposed to solve this problem, such as dividing the entire tumor volume into several different sub-volumes according to different CBV values or vascular leakage (44, 46), and comparing the CBV before and during RT. Prior to the use of elevated CBV for defining enhancement target in GBM, the amount of elevated CBV in GBM related to tumors is a key factor. This makes sense since tumors can exist at areas where CBV is not elevated (50).

Even though DSC and DCE MRI can be used to estimate CBV, there are serval limitations of perfusion MRI. Similar with DW-MRI, DSC MRI has serious geometric distortion and signal loss because it is obtained by EPI. Moreover, CBV can be achieved from T1-weighted DCE MRI (51), however determining reliable arterial input function is still non-trivial. Use of a large sagittal view is able to determine the arterial input function of aorta, thereby reducing uncertainties. To incorporate CBV calculations into the RT workflow, a validated software is also necessitated. Last but not least, consider the fact that longer scanning time for every picture empowers large spatial resolution, however a high temporal resolution restricts the spatial resolution (52–54). Therefore, temporal and spatial resolution needs to be balanced.

Diffusion tensor imaging (DTI)-MRI is one of the most popular technique to detect brain tumour and implement radiotherapy. Kelly et al. (55) and Price et al. (56) have shown that tumour cells preferably transmit along the white matter tracts (WMT) and have decreased infiltration in gray matter. Krishnan et al. investigated another retrospective study and showed the process from the original tumor to the recurrence using DTI-MRI for glioma patients. Their results confirmed that WMTs can be used as a route to indicate tumor spreading (57). These findings demonstrated that GBM grows along with WMTs. WMTs can be used as a biomarker to simulate tumor growth, and DTI-MRI can be used to define the corresponding CTV, as shown by serval retrospective studies (57–61). Moreover, DTI-MRI incorporated with mathematical model could define the RT target areas and evaluates the coverage rate of recurrence in patients with clinical target volume (CTV) derived from DTI.

Trip et al. implemented the first phase 0 study to deline CTVs using the DTI in post-operative chemo-/radiotherapy for GBM (61). Their results showed that CTVs derived by DTI-MRI performed worse for the central recurrences, however they are better for non-central and satellite recurrences. Unfortunately, this study lacks enough patient samples to conclude fair observations. In addition, the authors did not use the deformable registration, and measured the Hausdorff distance (HD) geometrically rather than anatomically. Nonetheless, due to the time-dependent migration of tumor cells, the effect of adopting a new target definition can only be truly evaluated in interventional studies.

Positron emission tomography (PET)-CT has been a valuable technique for RT in staging (62) and accurate target-volume delineation (63, 64). When imaging patients using PET-CT, a radioactive substance emitting positrons must be injected. Thereafter, the scanner detects the emitted photon pairs (511 keV) and quantifies their distribution throughout the patient after signal correction and normalization steps. A variety of PET radiotracers are available for in vivo imaging of biological processes. There are only five oncologic indications approved by the U.S. Food and Drug Administration, i.e. [¹⁸F]-fluoro deoxy-D-glucose (FDG), Na[¹⁸F], ¹⁸fluciclovine, [¹¹C]-choline, and [⁶⁸Ga]-DOTA-octreotate (DOTATATE). Many other candidates are being evaluated for clinical treatment.

As the most widely used PET radiotracer in the clinic, the details of FDG-PET are discussed in the following. FDG-PET depends on the relation between glucose metabolism and the upregulation of glucose transporters in cancer cells, and has played a significant role for patient staging, selection and RT target delineation (65–71). For example, mistreatment [¹⁸F]-FDG PET scan serves as a biomarker for adaptive dose painting. Kong et al. recently performed a phase II RT trial based on PET-CT for patients with NSCLC using interim [¹⁸F]-FDG PET to identify regions of poorly responding disease (72). An improved 2-year control rate was achieved with a factor of 82% versus 69% in the trial Radiation Therapy Oncology Group (RTOG) 1106 trial (73). Given the aforementioned results, FDG-PET is now the basis for RTOG 1106. Generally, the higher the uptake of FDG is, the more dose is needed, with a 10-30% increase to achieve the same control probability in the low FDG area (74).

Vogelius et al. derived dose–response functions for different structures that were distinguished by pre-treatment FDG-PET CT (75). Such dose response functions are used in the dose optimization process. Versus traditional treatments, they can maximize the patient’s TCP under the constraint of a constant average dose. Based on this study, Grönlund et al. implemented dose painting by numbers (DPBN) (76), and showed that TCP values were increased compared to uniformly delivered dose. Since uncertainties exist in the relationship between SUV and TCP, Grönlund et al. focused on the heterogeneity of SUV rather than absolute SUV, which provide a method to deal with the uncertain threshold of SUV.

In 2005, Ling et al. proposed the first DPBC technique (3). DPBC applies a dose boost by a certain threshold to a subvolume of the tumour. The regions of relatively lower and higher risk for recurrence are fixed with the threshold from the quantitative functional imaging. There are uncertainties related to the imaging modalities where the prescription function should be based, and when translating the image into a prescription function for dose painting (96, 97). Some major uncertainties, induced by PET imaging partial volume effect (PVE), tumour deformable image registration (DIR), and variation of the time interval between FDG injection and PET image acquisition have been investigated by Chen et al. (89).

In particular, interobserver variability in target volumes is a well-known factor for radiotherapy. For PET-CT based dose painting, detailed contouring guidelines can be referred to the PET-Plan (NCT00697333) clinical trial protocol (98). As a part of the PET-Plan quality assurance, a contouring dummy run (DR) was performed to analyse the interobserver variability (82). In addition, Korreman et al. investigated dose conformity by using the RapidArc optimizer and beam delivery technique (87). Optimisations for dose panting were performed in Eclipse by tuning the leaf width of multi-leaf collimator (MLC), the number of arc and collimator rotation. Positioning uncertainties are also considered for robust analysis and have been demonstrated with a low positional error (87).

A commonly used treatment planning approach for DPBC is the simultaneous integrated boost (SIB) technique (99) to increase the dose at a subvolume of the tumour. Patients treated by this approach are expected to achieve better dose tolerance (100). A homogeneous boost dose is assigned to the subvolume in many clinical trials (101). PET-guided DPBC was used to assess the feasibility of intensity-modulated radiotherapy (IMRT), where the maximum tolerated dose in head and neck cancer can be determined (88). The same purpose of increasing tolerated dose is also explored in non–small-cell lung cancer (NSCLC) treatments (85). Fleckenstein et al. proposed a source-to-background contouring algorithm for FDG-PET in the process of RT planning (85). There have been other advantages when using DPBC for RT planning. For example, Kong et al. demonstrated that adapting RT by boosting dose via DPBC to the FDG avid region improves 2-year local-regional tumour control and overall survival rate (84).

The term ‘‘dose painting by numbers” describes prescription dose on a voxel-by-voxel level (6). DPBN is a method to increase the additional dose gradually, adjusted by the local voxel intensities. Mathematical models are usually used to identify the relationship between the voxel values of the functional imaging and the risk of local recurrence (76). In particular, Chen et al. (89) analysed how the uncertainties in quantitative FDG-PET CT imaging impact intratumorally dose–response quantification, such as ones cause by PVE and tumour DIR (89). The negative effect arising from some of these uncertainties could be optimised by DPBN techniques (89).

There has been various technical feasibility and robustness of DPBN published recently (102–104). Dose prescription with steep gradients can be delivered by numerous subvolumes via a conventional linear accelerator (105). Rickhey et al. used the DPBN approach in brain tumours with [¹⁸F]-FET-PET, and achieved with high accuracy (106). Moreover, [¹⁸F]-FDG-PET-guided DPBN was proved to be feasible in phase I clinical trial by Berwouts et al. in head and neck RT (93). Recently, Grönlund et al. investigated the spatial relation between retrospectively observed recurrence volumes and pre-treatment SUV from FDG-PET (76). As were shown in the aforementioned studies, SUV driven dose–response functions have been presented to optimize ideal dose redistributions under the constraint of equal average dose of a tumour volume (76). A further analysis was proposed to investigate the feasibility of DPBN to increase the TCP in a clinical scenario (91).

Some DPBN approaches have been proposed by using sub-volumes as targets (76, 87, 91) or dose maps with prescription to the voxel as objective function (107), but these methods belong to dose-volume based optimization algorithms. Jiménez-Ortega et al. presented a new optimization algorithm to implement directly constraints to voxels instead of volumes, where Linear Programming (LP) is used to carry out DPBN approximation. This method is implemented in CARMEN, a Monte Carlo (MC) treatment planning system (81). Since proton therapy has been reported as potentially capable of decreasing toxicity, Håkansson et al. investigated DPBN in proton RT planning by comparing proton dose distributions with delivered photon plans from a phase-I trial of FDG-PET based dose-painting (90). Experimental results stay in line with the physical properties of the photon and proton beams, i.e. proton DPBN can be optimised with a quality comparable to photon DPBN (90).

DPBC mainly refers to the specific function image parameters to set the threshold for the replenishment area. Biomarkers in the high-risk area for recurrence have larger values over the defined threshold, while low-risk recurrence area corresponds to biomarkers having smaller values than the threshold. DPBN assumes that the recurrence risk of a certain pixel in the tumor area is positively correlated with the parameter intensity of its specific function image pixel, and the radiation dose of a certain pixel is directly related to its corresponding functional image pixel information. DPBN directly relies on theragnostic imaging (6). DPBN requires a customised software package to optimize the irradiation plan, but there is no commercial software directly implement optimization of DPBN (108, 109).

We remark that DPBC and DPBN have their own advantages and disadvantages (109). Advantages of DPBC are that sub volumes that need to boost dose can be pre-drawn before the treatment plan. Then the sub volumes can be set to add margins to supplement the geometric uncertainty, and the treatment plan can also be evaluated by conventional DVH. Disadvantages of DPBC mainly include the lack of consensus of the threshold for biomarkers. DPBN has more theoretical advantages than DPBC because it can deliver doses to voxel level. However, it cannot extend the margin of specific voxels and is more sensitive to uncertainty arising in image registration. Therefore, online image-guided treatment that can clearly show soft tissues is required.

Meijer et al. examined both DPBC and DPBN techniques for non-small cell lung cancer (NSCLC) patients’ treatment (95). In general, the amount of DPBC dose-boosting is limited whenever the GTV boost is close to any serial risk organ. However, DPBN shows significant higher dose values to high SUV voxels and are more distant from the organs at risk, since DPBN boosts work at a voxel-by-voxel basis (95).

Current quantitative imaging approaches have larger voxels and worse signal-to-noise-ratio (SNR) compared to other clinical imaging techniques, since quantitative imaging requires one or multiple quantitative parameters for each voxel (92). For dynamic contrast-enhanced MRI/CT, the acquisition of multiple images is required (1). Therefore, target delineation and dose painting on functional images rely more on the values of parameters of biomarkers compared to other clinical imaging. As was reviewed in Section 2, the main limitation of functional imaging lies in the uncertainties related to the relationship between biomarker and their corresponding parameters. Further studies involving more robust calibration are needed to propose a more robust relationship. One alternative way is to use the heterogeneity of the biomarkers rather than the absolute value of biomarker, which can reduce the impact the uncertainties of the biomarkers (76). For tumor types with high tumor heterogeneity, such as GBM, it makes more sense to consider the biomarker parameters of voxel level instead of the mean value of a whole tumor. Ideally, dose painting prescriptions should be based on dose response data which can be observed from multiple functional imaging methods (108).

At present, the interest of imaging biomarkers is not only due to the state-of-the-art treatment development and personalized medicine but also due to the recent application of artificial intelligence (AI) algorithms in image processing (140).

AI techniques can help construct image-based biomarkers which can provide a comprehensive view of the entire extent of the tumor and can capture regional tumor heterogeneity (141–143). Since biomarkers are used in a non-invasive (or mildly invasive) way, AI techniques, such as convolutional neural networks (CNN) can facilitate revealing tumor phenotype related to prognosis and monitoring the development and progression of the disease or its response to radiotherapy. Traditionally, CNN-based medical image processing has been used to segment the organ and tumor (140). In addition, it is necessary to explore other kinds of biomarkers (such as ADC, SUV, NAA) instead of tumour size and volume only. However, the quantification of these biomarkers has been difficult. Therefore, by defining biomarkers as a function of a large number of image descriptors (features) extracted from an image, we can employ machine learning or deep learning techniques to analyse these image features. The most discriminatory subset of features can be found to achieve the best performance for tasks such as result prediction (144–146). Savadjiev et al. presented a novel approach which firstly uses a segmentation-specific CNN to narrow the analysis only to the needed organ and tumor area in the image. Then, the relevant radiomic features will be transferred to another neural network model to classify tumor phenotype (140).

However, use of AI to analyse deep radiomics features has several concerns. Firstly, the theoretical analysis of deep radiomic features extraction are not easily interpretable due to the complexity of deep learning models. This is because a deep neural network requires a large number of complex but weak regularities in the data, which has highly complex and non-linear interactions between multiple network nodes and layers (147). In addition, a deep neural network usually contains millions of parameters, which is also a difficult task, involving sufficient compute power, patient data, and reproducible network models (140). Therefore, interests in building customised neural network models used for different medical imaging-based biomarkers have risen, through which a tremendous potential for biomarker qualification is expected.

Image registration is the process of geometrically aligning two or more images of the same scene taken at different times, from different viewpoints, and/or by different sensors (148).

Since current radiotherapy treatment planning systems have to create a treatment plan on the images produced by a radiation simulator. Most of the simulators are CT-based, therefore, they can only produce anatomical imaging. Properties of functional imaging as mentioned in this article can be employed by image registration to the anatomical imaging. Thereafter, the dose painting can be implemented to deliver more doses to the high-risk area of tumour defined by the functional imaging. As mentioned in Sections 2 and 3, the high quality of registration is important when using functional imaging to perform the dose painting. For DPBC, the uncertainties of registration can be reduced by adding margins to the high-risk area. However, DPBN features the voxel-level analysis and is more sensitive to uncertainty arising in image registration (92). Most treatment planning systems support image registration and fusion to allow the use of multimodality and time-series image data and even anatomical atlases to assist in target volume and normal tissue delineation. Image registration is also needed to assess dose variation during treatment course and enable adaptive radiotherapy, as will be discussed in Section 5.4.

Medical image registration is a broad topic which can be grouped from various perspectives. From an input image point of view, registration methods can be divided into unimodal, multimodal, interpatient, intra-patient (e.g., same- or different-day) registration. From a deformation model point of view, registration methods can be divided in to rigid, affine and deformable methods. From a region of interest (ROI) perspective, registration methods can be grouped according to anatomical sites such as brain, lung registration. From an image pair dimension perspective, registration methods can be divided into 3D to 3D, 3D to 2D and 2D to 2D/3D. There was a survey summarizing algorithms of conventional image registration (148). The Therapy Physics Committee of the American Association of Physicists in Medicine commissioned Task Group 132 to review current approaches and solutions for image registration (both rigid and deformable) in radiotherapy and to provide recommendations for quality assurance and quality control of these clinical processes (149). However, there is no standard mathematical formalism to perform registration, because noise, distortion and complex anatomical variations can occur. In addition, validation of the software systems performance is complicated due to the lack of documentation available from commercial systems. Recently, many researchers studied AI-based image registration. Compared with the conventional image registration, performing registration with AI can be more accurate and generally order of magnitude faster. Fu et al. summarized the details of the latest developments, challenges and trends in AI-based medical image registration methods (150). It is foreseeable that AI will continue to play an important role in this field.

Most RT plans remain “unchanged”. Once the plans are initialized, they are carried out until the end of treatment, along with minimal modifications on top of the original plans (151). RT treatments are usually delivered for the period of several weeks, which is a long and arduous process. In most cases, the tumour size, location and microscopic status (e.g., metabolism and hypoxia) will change, thereby resulting in long-term and even life-long consequences on patients’ life quality after treatment. To address such challenges, the adaptive RT has been developed recently, where interests in functional image-based dose painting have been growing.

Adaptive radiotherapy is a process to control for anatomical and functional variation over the treatment course (126). Figure 3 shows the workflow of the conventional and adaptive radiotherapy. Both conventional and adaptive radiotherapy require four key technologies: imaging, planning, quality assurance and treatment. Versus conventional radiotherapy workflow, adaptive radiotherapy additionally requires assessment (126). Assessment is the process by which imaging is used to decide whether to adapt the plan or not. This process can range from a manual evaluation to highly automated review of cumulative dose. Therefore, in this step we can embed functional imaging to evaluate the functional variation and re-prescribe dose, employ AI to automatically deline targets and organ at risk, and perform AI-based the image registration. Many studies have discussed the potential of functional images for a more personalized RT planning, however, how to incorporate such quantitative functional imaging into adaptative RT has been barely investigated. For example, since functional imaging can define easy-to-control and hard-to-control areas in the tumor before a treatment, assume that patients take the same functional imaging after a period (e.g., 2 weeks) of radiotherapy, changes in biomarkers can be measured, indicating radio sensitivity and radio resistance after serval fractions of radiotherapy. Adaptative RT can be therefore realisable.

The field of radiation oncology has been used to address different sources of uncertainties during patient planning, including range and setup uncertainties (152), organ movements (153) and clinical target volume (CTV) definition (154). Consider that the implementation of dose painting depends on the quality of data obtained via biomarkers. The interpretation of such data inevitably introduces some uncertainties. Therefore, it is foreseen that potential mathematical tools to solve uncertainties, e.g., partially observable Markov decision processes (POMDP), imperfect state information (ISI) and adjustable robust optimization (ARO) might be necessitated.