A biomarker is a “defined characteristic that is measured as an indicator of normal biological processes, pathogenic processes or responses to an exposure or intervention, including therapeutic interventions” ^[1][2][1,2]. Biomarkers must be measurable, but can be numerical (quantitative) or categorical (either a quantitative value or qualitative data). To be applied in clinical practice they must be reproducible, linked to relevant clinical outcomes, and must demonstrate clinical utility. In healthcare settings (and in research), biomarker uses include screening for disease; diagnosing and staging cancer; surgical targeting and radiotherapy treatments; guiding patient stratification; and predicting and monitoring therapeutic efficacy and/or toxicity. Biomarkers can be also used as predictors of traditional endpoints (i.e., surrogate endpoints) such as treatment response and survival [2].

Imaging biomarkers (IB) are a special example of biomarkers where the indicator is derived from in vivo medical images providing an attractive choice for clinical use as they can be implemented and used as a real-time, non-invasive, cost-effective option. While extensive literature exists on the development of biological biomarkers, a roadmap for the definition of an IB is not adequately defined yet. Nonetheless, in particular settings, the question of how IB acquisition and analysis should be standardized, and how terminology should be harmonized have been addressed by numerous academic, clinical, industrial, and regulatory groups ^{[3][4][5][6][7][8][9][10][11][12][13][14][15][16][17]}[3,4,5,6,7,8,9,10,11,12,13,14,15,16,17].

Despite some IBs being used extensively and others showing great clinical potential, there is only a limited number of them currently guiding the clinical decision-making processes in hospitals, while many IBs of potential interest are often confined to academic literature without a clear clinical route for implementation ^[18][19][18,19]. Many IBs have been evaluated on a limited cohort of patients, and in retrospective and monocentric studies. To become routinely used in the management of patients, an IB should prove to be a reliable measure used to test the clinical hypotheses within a multicenter research clinical trial.

A characteristic of IBs is that the same biomarkers can be produced by imaging systems of different makes and models installed in different clinical sites. These devices are designed, approved, maintained, and operated to provide images that diagnostic radiologists and nuclear medicine physicians interpret. Usually, the quality assurance plans, carried out to optimize and assure adequate image quality, are performed for visual analysis in many cases (not all) with little focus on how the image quality will affect quantification, and little focus on the standardization of IBs though different sites.

The roadmap to carry the IB to clinical practice is required to follow different simultaneous validation routes, such as technical, clinical, and cost-effectiveness evaluations [18]. The technical evaluation requires that the IB is measurable precisely and accurately, and also that it is widely available in all geographical territories. In the clinical evaluation, the IB should demonstrate measurability of biologically relevant characteristics or be a predictor of clinical outcomes. Cost-effectiveness must also be considered because IBs must not only demonstrate an association with health benefits, but also demonstrate ‘value for money’ when compared with the use of other clinical information.

Researchers initially aimed to perform bibliographic research by searching on PubMed for various techniques (e.g., “SPECT”) together with the words “quantitative”, “quality assurance”, “MPE”, “clinical trials”, “standardization”, and “imaging biomarkers”. This, however, produced a large range of IBs that were not generalizable enough for a broad review (i.e., a new IB specific to a single trial only), and thus researchers gave priority in the selection to those proposed by the consensus of many researchers, peer-reviewed publications from professional societies and associations, published clinical guidelines, and also on the basis of the authors’ experience. Many international initiatives have dealt with the definition, description, and evaluation of the usability of QIBs. These include the European Imaging Biomarkers Alliance (EIBALL), the Quantitative Imaging Network (QIN), and the National Cancer Imaging Translational Accelerator (NCITA) ^{[19][20][21][22][23][24]}[19,20,21,22,23,24]. Biomarkers selected and described in this review are reported in many of the above initiatives. 

PET/CT reporting has historically been based more on the extent of the disease than on the affinity of the tumor itself to the tracer used [25]. Within the next paragraph researchers limit the description to the IB of glucose metabolism provided by 18F-fluoro-deoxyglucose (18F-FDG) PET/CT, because of its prevalence in PET imaging, and on the understanding that other IBs provided by different radiotracers can be rapidly extrapolated from this one.

Discrete qualitative scale, where the areas of uptake are compared to normal, disease-free organs and tissues are used often in response assessment. The most commonly used references are the liver and the blood pool usually measured in a large vessel such as the aorta as in Deauville [26] and in PERCIST [27] metabolic response criteria. The application of a discrete scale requires training for the readers and a good inter-reader variability was demonstrated in the literature for Hodgkin lymphoma (HL) ^[28][29][30][28,29,30] and Non-Hodgkin lymphoma (NHL) ^[31][32][31,32]. Moreover, it is easy to perform in clinical settings, can be done with standard-of-care software, and only requires that reference uptake of different scans of the same patient remain within a tolerable limit.

On the other hand, it does not overcome the problem of visual analysis when two areas with the same uptake could be misclassified if they are surrounded by different backgrounds [33].

Standardized Uptake Value (SUV) is the ratio of the decay-corrected concentration of activity in the tissue to the injected activity normalized to the patient’s body weight (SUV corrected for lean body mass (SUL) has been introduced recently [27] but has not been yet used extensively in clinical practice ^[3][34][3,34]) has been proposed to override this problem [31]. SUV moves PET image analysis from a discrete to a continuous scale, and SUV metrics are calculated inside a volume of interest (VOI) that could be:

• Maximum value within the VOI (SUVmax): simple to measure and provides information about the most active tumor foci. Drawbacks include a strong dependence on image noise as it corresponds to a single voxel measure.
• Average of all values in the VOI (SUVmean): less vulnerable to image noise, but depends heavily on the delineation method used for drawing the VOI [35].
• Maximum tumor activity within a 1 cm³ VOI in the most active part of the tumor volume ^[3][35][3,35] (SUVpeak): associated with the most active part of the tumor, but in a standard volume, being therefore less dependent from noise.
• Volume of the VOI itself, Metabolically Active Tumor Volume (MTV) measured in ml or the product of the MTV by SUVmean, Total Lesion Glycolysis (TLG): they evaluate the extent and aggressiveness of disease. MTV and TLG are marginally affected by noise but are dependent on the delineation method used for drawing the VOI [35].

MTV measurements are of crucial importance in pursuing a quantitative approach to PET; however, as of yet, the precision, accuracy, and repeatability of MTV and TGV have not been established, which are prerequisites of a good prognostic index.

Although it is possibly the most accurate metric, kinetic modeling, which describes the delivery, retention, and utilization of any radiotracer under investigation, is difficult to perform since it requires long dynamic PET scans, personnel with specific knowledge, and dedicated software for the analysis. It is usually used in single sites to perform pharmacodynamics studies, in phase I-II studies, or in novel non-18F-FDG tracers and has, to date, limited application in routine clinical practice. Recent developments such as the implementation of whole-body PET scanners have enabled kinetic modeling without increasing patient scan time (a large limitation with traditional PET systems), and thus may translate kinetic modeling to clinical routine for a variety of oncology and non-oncology radiotracers ^[36][37][36,37].

The sources of error in uptake evaluation can be attributed to specific factors such as the dependence on the PET/CT scanners, site procedures for patient preparation, data acquisition and post-processing. The former point may present issues with cross-calibration of PET/CT scanners whereas the latter are faced with standardization of procedures [38].

Cross-calibration of PET scanners and dose calibrators is important to minimize uptake variability. Cross-calibration consists of the acquisition, as detailed by the PET/CT manufacturer’s manual, of a cylindrical phantom in which an activity of ¹⁸F, measured with the dose calibrator, has been inserted. A calibration factor is set on the PET/CT scanner in such a way that the activity measured by the scanner is the same measured by the dose calibrator. Since the dose calibrators are usually calibrated to 3–10% precision and there could be some errors in the filling procedure of the phantom, the accepted difference in clinical practice between the activities measured by the scanner and the dose calibrator is usually set to 10%, although may be set to a tighter control limit in clinical trials or may be radiotracer-dependent depending on half-life (i.e., assay of ¹⁵O-H₂O).

Variability amongst different scanners, even in a controlled environment of a multicenter clinical trial can be large, and in some cases prove to be different up to 25% [39]. Efforts of cross-calibration through the measurements of different phantoms permit a variability of less than 10% in multicenter clinical trials ^[40][41][40,41] while 5% should be a good requirement to derive PET/CT-based QIBs ^[42][43][42,43].

A plethora of image reconstruction algorithms exist, and many differences arise when the measurements are carried out in small volumes, where loss of counts due to spill-out and partial volume effects occur [44]. The recovery coefficient curve is a figure of merit describing the ratio between real and measured activity varying the dimension of the volumes. Ideally, the ratio should be one with a 10% variation due to the scanner calibration as seen above. In reality, the ratio is one for larger volumes and decreases slowly when the dimensions of the volumes are smaller than 5–8 mL [45]. A NEMA/IEC phantom with hollow spheres, which can be filled with a known activity, can be used to tune the reconstruction algorithm. Several parameters of the reconstruction algorithm, such as the number of iterations, the number of subsets, and the width of the Gaussian smoothing filter can be adjusted to achieve the best recovery coefficients.

SUV measurement variation across PET/CT scanners in the range of 10–25% just as the consequence of the intrinsic variability of the instrument is commonly observed in the context of multicenter clinical trials [45]. Hence, the cross-calibration of PET/CT scanners and ancillary instrumentation is the first condition to achieve an accuracy in tracer uptake measurement of 5–10% [39]. Several programs for the cross-calibration of PET/CT scanners have been carried out in the recent years by imaging and oncology societies: the EANM (EANM) accreditation program for site of excellence carried out by EARL Ltd. (Vienna, Austria) [46], the UK NCRI PET Clinical Trial Network [40], the ACRIN program [43], the Clinical Trial Network of Society of Nuclear Medicine and Molecular Imaging (SNMMI) [47], the JSCT NHL10 trial in Japan [48] and the FIL Cuneo core lab in Italy [41].

Regarding the standardization of the acquisition parameters, various recommendations have been released by the US and European nuclear medicine associations ^[3][34][35][3,34,35] in order to provide minimal standards to guarantee an efficient comparison of PET/CT metrics acquired at different time points (intra-patient) and between different patients (inter-patient), either at a single site or across multiple sites ^[25][49][25,49].

Several IBs are currently used in nuclear medicine such as left ventricular ejection fraction, renogram radiotracer clearance, thyroid uptake, and myocardial perfusion. They all rely on dynamic imaging to measure the change in the number of detected counts per frame. Similar to PET scanners, gamma cameras are susceptible to the same issues with variability in QIBs across scanners employing different detection technology, acquisition protocols, and image reconstruction parameters in multicenter clinical trials.

Standardization using phantoms for many of the clinical indications that can be scanned may be implemented, for example, a custom-built 3D printed myocardial perfusion phantom aimed at providing ground-truth validation of multimodal, absolute MPI applications in the clinical setting [50]. The phantom enables tracer kinetic measurement, including time–activity curve and potentially compartmental myocardial blood flow analysis. Phantoms for MPI quantification of ejection fraction and left ventricular volume have also been utilized to evaluate the dependence on collimator and reconstruction parameter choice, demonstrating the need for standardization when evaluating these QIBs in nuclear imaging [51].

A recent workgroup has aimed to summarize normal QIB values for MPI stress/rest using 99mTc and 201Tl perfusion agents as well as 123I-MIBG sympathetic imaging and included correction values for sites performing CT attenuation correction, demonstrating dependency on gender, ejection fraction, number of ECG gates, and volumes [52]. The work shows that appropriate quantification based on common normal databases and standard technology plays a pivotal role in clinical practice and more so for research where multicenter studies require standardization between scanning and resulting IB values.

In radionuclide angiography (RNA) using radiolabeled red blood cells, the calculation of left ventricular ejection fraction (LVEF) is the primary IB for cardiotoxicity assessment in chemotherapy regimens. Work has shown that depending on the acquisition of SPECT or planar imaging data, the LVEF value may not be equivalent due to technical reasons [53], and also when compared to cardiac MR using thresholds of 50 and 55%, there was misclassification of 35 and 20% of cancer patients, respectively [54].

Recently, in particular for bone scintigraphy, an analog of PET SUV has been proposed in SPECT [18], based on the inter-calibration between the activity calibrator and the SPECT system through the acquisition of a uniform phantom with known activity. A new imaging gamma-camera that will be commercially available in a few years, i.e., based on integrated cadmium-zinc-telluride detectors, will help in having more reliable quantitative indices also in SPECT, which would also greatly help in dosimetry application (see next paragraph).

Phantom standardization provides a means to assess the differences not only between scanners, but also between acquisition techniques such as SPECT and planar, and in cross-modality imaging such as PET and MR ^[55][56][55,56]. Similar phantoms exist for many other clinical indications for gamma camera imaging such as dynamic renograms [57], SPECT renal imaging [58], head and neck/thyroid imaging [59], and excellent review works have been published summarizing phantom availability and types, more commonly involving the use of 3D printing ^[60][61][60,61]. The 3D printing of radioactive phantoms serves to address certain limitations with nuclear medicine phantoms in general, allowing phantoms without inactive walls (thus contributing to lower partial volume issues) and more patient-specific and complex shapes [62]. Other initiatives include radioactive cryogel phantoms which allow the study of motion correction strategies in realistic lesion-simulating shapes that deform with movement [63].

In the field of RNT, to achieve the desired prescribed dose and to estimate the absorbed radiation dose after administration of the radiopharmaceutical, accurate dosimetry is needed pre- and post-treatment. The quantification of radiation dose can be performed with the use of planar or SPECT imaging of post-therapy radioligands, in that patient images of variable radionuclide distribution should be ‘translated’ from three-dimensional count maps into maps of radioactivity. A careful quantitation procedure is required where a radioactive source (typically a point source, syringe, or large phantom) calibration with an accurately known activity is required to convert recorded counts to activity.

Currently, there is no consensus method for calibration; multicenter studies have aimed to harmonize quantitative imaging across a range of vendor systems with a specific focus to dosimetry applications for thyroid radioiodine therapies ^[64][65][66][64,65,66], 177Lu-based therapies [67] and 99mTc pretherapy for microsphere embolization therapies [68]. As image acquisition and reconstruction conditions can be radionuclide dependent in terms of the number of energy windows, appropriate scatter/attenuation correction, sensitivity to radionuclide, dead-time, and collimator blurring, societal guidelines have been produced for quantitative imaging of 177Lu [69], 131I ^[70][71][70,71], and 90Y [72] RNT radioligands. Advanced 3D voxel-based dosimetry cannot be achieved by planar imaging, yet planar imaging still serves as the main dosimetry method for RNT due to its ease of applicability. Uncertainty of the data due to the diverse range of variable physical and scanner parameters that can be adjusted remains complex, although recent EANM guidelines offer a framework to model the propagation of errors in the measurement process specific to the dosimetry chain of absorbed dose at the organ or tumor level, together with clinical examples of how this can be implemented [73].

At the time of writing, no guidelines for quantitative SPECT/CT systems harmonization have been published [74]. Many studies showed the need for harmonization of quantitative SPECT/CT scanners across centers ^[75][76][75,76], given the unavoidable differences in the calibration of these systems, in the reconstruction methods and in the image/data correction techniques being applied. Software from different vendors may also produce different quantitative results from the same SPECT system. Recent work has demonstrated that five clinically suitable parameters of image quality assessed by uniform phantoms (background calibration factor, total activity deviation, noise coefficient of variation, hot contrast, and recovery coefficient) may be useful for the assessment of system performance in terms of correct quantitative acquisitions of images (however, in a limited bi-center study) [77].

In RNT, many national and international efforts have been made towards scanner harmonization, with the overall aim of allowing repeatability and reproducibility in the calculation of absorbed radiation dose. Initiatives so far have even led to the setup of a quantitative imaging network [78]. A key reason for this is likely to be the legal implications of the European Directive 2013/59/Euratom, stating that radiation doses for therapy purposes should be “individually planned and their delivery appropriately verified” [79]. Thus, there remain many efforts for RNT that have not yet propagated to SPECT or planar imaging. Although the EANM have produced PET harmonization standards [80], a SPECT/CT harmonization pilot study is in progress but, at the time of writing, it has yet to be completed.

The definition and subsequent translation into the clinic of MR-based IBs is one of the important points of research in medical imaging of the last years. Considering the versatility of the MR systems, there are many strategies that allow us to estimate parameters that are promising IB candidates.

However, for the majority of them, standardization is still lacking because there are many confounding effects that influence the measurement and it is often difficult to establish accuracy a priori. It is important to remember what has been done so far by QIBA ^[15][81][15,81], which has promoted initiatives for the best usability of MR biomarkers. A recent American Association of Physicists in Medicine (AAPM) report illustrates the main biomarkers derived from MR data [82]. 

MR spectroscopy (MRS) allows obtaining signals from different molecules (after suppressing water and lipids) due to the chemical shift produced by shielding the molecular electronic orbitals. 

Routine MRS use has been hampered by the low concentrations of the molecules (metabolites) of interest, often of a few mM, which generate a signal-to-noise ratio (SNR) much lower than that of MR imaging. It is limited in the literature primarily to brain and prostate applications, where the primary metabolites of interest are N-Acetyl-Aspartate (NAA), Creatine (Cr), Choline (Cho), Lipids (Li), Lactate (Lac), followed by Glutamate, Glutamine, and Gaba in the brain, and Choline, Creatine, and Citrate in the prostate [83].

MRS acquisitions can be based on single-voxel or multi-voxel (Chemical Shift Imaging, CSI) sequences, both 2D and 3D. The most common pulse sequences are Point RESolved Spectroscopy (PRESS), which samples a spin echo, and STEAM, which samples stimulated echoes. The latter allows the use of very short TEs (<10 msec), which is useful for reducing unwanted T2 weighting effects, but has lower SNR. Recently a consensus paper has been published on the reporting of MRS methods and results, to allow an adequate assessment of MRS studies [84].

DCE perfusion studies employ fast T1-weighted gradient echo (GRE) acquisition sequences repeated many times during the injection of paramagnetic contrast medium. The obtained dynamic images show the signal variation due to the contrast bolus in the arterial and venous vessels and, in the case of pathology, in the extravasation from the intravascular to the extravascular and extracellular space ^[85][92]. DCE perfusion is used in the clinic in a range of oncology applications: among others prostate ^[86][93], brain ^[87][88][94,95], and tissue sarcomas ^[89][96].

In DSC perfusion imaging (primarily used in brain imaging) very fast MRI acquisition sequences are acquired while a bolus of intravenous paramagnetic contrast agent is injected. The DSC contrast is due to the passage of the paramagnetic contrast agent, which generates a loss of phase coherence of the spins on the transverse plane and, consequently, a signal loss. Once the contrast medium has passed from the arterial perfusion bed to the venous one, if the blood–brain barrier is intact, the phase coherence of the spins is recovered and, consequently, the signal amplitude of the tissues under examination ^[85][92].

By analyzing the wash-in/wash-out curve of the contrast agent, it is possible to estimate the Cerebral Blood Volume (CBV) on a pixel-wise basis. In practice, only relative estimates of the CBV, rCBV, are used, as the calibration in absolute units is strongly affected by subject-dependent parameters that are difficult to estimate (for example the hematocrit and local AIF) ^[83][90][83,102]. rCBV is a widely used biomarker in brain MRI, whose value is related to tumor angiogenesis ^[91][92][103,104].

Arterial Spin Labeling (ASL) uses the endogenous flux of water molecules to study perfusion. In ASL, two images are acquired: a control image and an image tagged via an additional selective radio frequency pulse positioned on a feeding artery for the volume of interest. This pulse, called the “tagging pulse”, labels the water molecules present in the arterial flow. Therefore, the difference between the tagged image and the standard one will contain only the signal of the water molecules present in the arterial vessels and corresponds to an arterial flow map ^[93][94][105,106].

Cerebral Blood Flow (CBF), which is a relative measure of arterial blood flow, is the most commonly used IB in ASL. ASL is used in the clinic in the diagnosis and follow up of various brain diseases, such as neurodegenerative diseases ^[95][107], multiple sclerosis ^[96][108], stroke ^[97][109], and in the diagnosis and follow-up of brain tumors ^[98][110].

Many IBs can be estimated from Diffusion Weighted Imaging/Diffusion Tensor Imaging (DWI/DTI) acquisitions, the most common being the Apparent Diffusion Coefficient (ADC-logarithm of the signal obtained using at least two b-values, usually a small and a large one, normalized to the b-value itself). The b value measures the degree of diffusion weighting applied and depends on the amplitude and on the time of applied gradients. If the diffusivity of the tissue water molecules is spatially isotropic, it is not necessary to change the direction of the diffusion gradient, as the estimated ADC will always be the same. However, this is not the case in almost any organ in the human body and when using multiple directions of the diffusion gradient, it is possible to estimate another IB, the fractional anisotropy (FA). For the estimation of FA, at least six directions of the diffusion gradient are required, i.e., one DTI acquisition ^[99][115].

Other biomarkers can be defined using more complex models. For example, the Intra-Voxel Incoherent Motion (IVIM) model, which allows the estimation of slow, fast diffusion coefficients and perfusion fraction ^{[100][101][102]}[116,117,118], the Kurtosis model, which allows the estimation of the corrected diffusion coefficient and of the kurtosis, or the stretched exponential model ^[103][119]. All these parameters are currently used as IBs in oncology on different organs: prostate, breast, liver, prostate, and uterus ^{[104][105][106][107][108]}[120,121,122,123,124].

MR-based IBs less used in oncology or those with limited clinical use or experimental/unproven IBs are briefly introduced here. An important IB group is that related to relaxometry measurements, which allows a voxel-based estimate of T1, T2, and T2* tissue relaxation times. These IBs can be estimated with good spatial resolution and have found extensive use for the diagnosis of many diseases ^[109][110][138,139]. There is a QIBA profile dedicated to the use of relaxometry-based biomarkers, with indications also on the QA protocols to be adopted ^[82][111][82,140]. IBs based on CEST (Chemical Exchange Saturation Transfer—in which off-resonance RF pulses are used to saturate the signal of different molecules) are also proving popular. In particular, in oncology, the use of amide proton transfer (APT) contrast developed, which allows for the investigation of the tissue concentration of amide protons. APT is widespread especially for the study of cerebral gliomas, on MR systems with static field ≥3 T ^[112][113][141,142], although recent studies showed applications in extra-brain oncology ^[114][143].

Of note is the development of the first total-body MR system for Fast Field Cycling (FFC) MRI ^[115][116][144,145], a method which involves the transformation of the “static” field into a “variable” field. Another promising area of research for MR-derived QIBs is that of MR-LINAC hybrid systems, which have different technical characteristics with different challenges, and some references are indicated for interested readers ^[117][118][146,147].

Measurements of relative tissue density can be calculated for each pixel and numerically represented as Hounsfield units (HU) for comparison with reference tissues. HU is calculated based on a linear transformation of the baseline linear attenuation coefficient of the X-ray beam, where distilled water (at standard temperature and pressure) is arbitrarily defined to be zero HU and air defined as −1000 HU.

Specific organs can be highlighted by variations in the resulting gray tone scale (windowing and leveling). As in MRI, CT scans are often performed after the administration of timed doses of intravascular contrast media to enhance the differences between adjacent structures. The use of the HU to measure tissue density has aided radiologists in the diagnosis of a wide range of diseases such as any incorporating bone mineral density changes, fatty liver diagnosis, evaluation of pulmonary nodules, bone quality before/after mechanical intervention, cyst/tumor differentiation, coronary artery calcification, and kidney/gall stones amongst a host of other uses.

HU accuracy is extremely important from the perspective of quantification, and even when used directly in differential diagnosis criteria. For example, in the management of adrenal masses, clinical guidelines detail that an adrenal mass with CT density >10 HU has 100% sensitivity for the detection of adrenal malignancy (confidence interval 91–100%), and that HU = 0 is likely to suggest a benign adenoma ^[119][148]. Such criteria involving HU thresholds are commonplace in the management of other diagnostic criteria such as breast cancer ^[120][149], intracranial hemorrhage ^[121][122][150,151] and Agatston calcium scoring ^[123][152] to name but a few, and demonstrate an innate dependence on the accuracy of HU as an imaging biomarker.