WO2019241266A1 - Black-box assessment of disease and response-to-therapy by t1 or t2* weighted dce-mri - Google Patents

Abstract

A black box magnetic resonance system includes a magnetic resonance imaging subsystem and a controller subsystem. The magnetic resonance imaging subsystem includes a magnetic assembly having a polarizing magnetic coil and gradient coil assembly and receiving coils that collect magnetic resonance imaging data from a subject being treated for a tumor from which dynamic contrast magnetic resonance images are constructed. The control subsystem is operable to perform calibration scans and determine values for MRI biomarkers from magnetic resonance images obtained from the subject. Characteristically, the MRI biomarkers includes tracer kinetic parameters obtained from a tracer kinetic model.

Description

BLACK-BOX ASSESSMENT OF DISEASE AND RESPONSE-TO-THERAPY BY T1 OR 12*

WEIGHTED DCE-MRI

CROSS-REFERENCE TO RELATED APPLICATIONS

[0001] This application claims the benefit of U.S. provisional application Serial No.

62/683,271 filed June 11, 2018, the disclosure of which is hereby incorporated in its entirety by reference herein.

TECHNICAL FIELD

)0002) In at least one aspect, the present invention is related to magnetic resonance systems for assessing diseases, and in particular, brain tumors.

BACKGROUND

[0003] Recent advances in chemotherapy and immunotherapy agents have resulted in significant improvements in patient outcomes but have caused significant confusion/challenges in the resultant accompanying imaging. Historically, increases in tumor size and contrast enhancement have signified tumor progression, and decreases have signified treatment response. With novel agents (particularly with existing and emerging immunotherapy), simple changes in size and enhancement are no longer sufficient to make treatment decisions. Pre -clinical studies have shown us that quantitative DCE-MRI provides the most powerful MRI-based biomarkers to detect early response to therapy and predict outcome. These biomaricent, once they can be reliably and routinely obtained, are expected to enhance our ability to evaluate therapies and prolong survival in a higher proportion of malignant brain tumor patients, especially as imaging biomarkers and progression-free survival are becoming accepted surrogate endpoints in clinical trials.

[0004] In 2010, the Response Assessment in Neuro-oncology (RANG) criteria (15) were introduced as a comprehensive reform of the Macdonald criteria and included evaluation of nonenhancing tumor progression, definition of measurable and non-meas arable disease and progression for patients potentially enrolled into clinical trials, consideration of pseudoprogression and pseudoresponse, requirement of confirmatory scans for response, and recommendations for handling equivocal imaging changes. Details of the RANG criteria and other modifications to RANG are well documented (15). In 2015, new immunotherapeutic agents ipilimumab, nivolumab, pembrolizumab and other immune checkpoint inhibitors prompted the RANG group to re-evaluate the RANG criteria for progressive disease, new lesions in patients who are clinically not worsening with new recommendations for steroid use and the role of advanced MRI such as DWI, DSC-MRJ and DCE- MRI, the so called iRANO criteria (16).

[0005] Pseudoprogression is also associated with higher rates of methylated (inactivated) MGMT promoter. MGMT is a DNA repair protein that provides resistance to TMZ, When methylated, chemotherapy causes more cytotoxicity and apoptosis (17-19). The ability for DCE-MRI to provide greater sensitivity to distinguish pseudoprogression from tumor progression is highly significant for patient outcome. Both methylated MGMT and pseudoprogression have been linked to improved overall survival (20-22). Further, with advances research and FDA-approval of immunotherapy drugs for multiple solid tumors, the incidence of pseudoprogression has also increased and the challenges associated with discriminating pseudoprogression from true tumor progression has become a major challenge. Current RANG criteria require subsequent confirmatory radiographic scanning in order to distinguish between these two. However, multiple studies has shown that MRI perfusion, particularly DCE-MRI, is helpful in differentiating whether an increase in enhancement or new enhancement represents pseudoprogression or true progression at the time of the clinical question.

[0006] Current brain DCE-MRI is qualitative, and limited in application for patient management. Attempts at quantification have shown poor test-retesl reproducibility (13,14), and the causes have been identified as inadequate spatial resolution, difficulty in identifying the arterial input function, and application of an appropriate tracer kinetic model. Tools and substantial preliminary data has recently been developed to show that that these challenges are resolved with specific calibration, scan prescription, and data processing techniques.

[0007] T₂*-wcighted dynamic susceptibility contrast (DSC) MRI using relative cerebral blood volume is the dominant perfusion-weighted MR] technique to assess brain tumors. However, it has several issues related to the need for contrast agent leakage correction and susceptibility artifacts (hemorrhage, air, necrosis, bone, etc.) that confound measurement (34,35). DSC-MRI metrics such as percent signal recovery can provide semi-quantitative assessment of microvascu!ar permeability, however, Ti -weighted DCE-MRI is able to provide a different perspective of micro vascular permeability, with susceptibility being a non-issue (36).

[0008] Among the many candidate biomarkers, pre-clinical studies have shown that DCE-MRI provides the most powerful and sensitive markers of early treatment response (10,11), superior even to diffusion weighted imaging and dynamic susceptibility contrast MKI (11). This is because DCE- MRI provides direct assessment of local neurovascular parameters that are altered by both cancer and cancer therapies. Among them are

a volume transfer constant between blood plasma and extracellular extravascular space, e.g., flow and blood-brain barrier permeability; v_p, fractional plasma volume; and v_c, fractional extracellular extravascular volume.

[0009] Several studies using conventional DCE-MRI have shown its potential as an imaging biomarker of HOG. As a potential prognostic biomarker, elevated

before bevacizumab treatment has been associated with decreased PFS and OS (37). Another study which used both DCE- and DSC- MRI found that both

and rCBV were correlated with OS in newly diagnosed GBM (38). However and iCB V may be reflecting different aspects of tumor biology as they were not highly correlated with each other (39). As a potential predictive biomarker, a recent retrospective study found that PFS and OS were predicted by baseline

before bevacizumab treatment, while this was not seen in a control group that did not receive bevacizumab (40).

[0010] Although many of the prior art studies are promising, there currently is no automated standardized quantitative method for DCE-MRI post-processing and analysis. Automation and standardization in stroke imaging, triage and management has changed the entire approach and outcome in stroke patients.

[0011] Accordingly, there is a need for automated standardized quantitative methods for DCE-

MRI post-processing and analysis. SUMMARY

[0012] In at least one aspect, the present invention standardizes, automates and translates quantitative dynamic contrast-enhanced (DCE) MRI biomarkers for the assessment of brain tumor response to therapy in clinical neuro-oncology trials. In particular, the invention provides a novel quantitative DCE-MRI capability to the clinical setting, including standardization of automated analysis tools needed for multi-center neuro-oncology clinical trials, and automated threshold values to differentiate pseudoprogression from true progression in individual lesions.

[0013] In another aspect, an automated thin/zero client“black box” DCE-MRI approach provides measures predictive of therapeutic outcome. This strategy is analogous to stroke imaging, where non -standardized, non-automated software failed to demonstrate clinical utility of advanced CT and MRI for imaging of ischemic penumbra and informed triage decision, but standardization and automation have begun to show clinical utility (e.g. DEFUSE and DEFUSE 2 trials (1-8)).

[0014] In another aspect, the MRI system and related methods provide a new high-resolution whole-brain quantitative DCE-MRI scan protocol and standardized, automated data processing tool that is well-suited for multi-center clinical neuro-oncology trials. This will overcome a major challenge in the evaluation of next-generation brain tumor therapies including chemotherapy and immunotherapy agents.

[0015] In still another aspect, the present invention provides a complete and coherent translation of new standardized, automated, quanti tati ve DCE-MRI capability to the clinical setting, including automated analysis tools needed for multi-center neuro-oncology clinical trials, and automated threshold values to differentiate pseudo progression from true progression in individual lesions. It is innovative to deliver a new capability to end users. Current tumor assessment criteria do not capture responses such as pseudoprogression and require follow-up scans at a later timepoint in order to retrospectively diagnose pseudoprogression or true tumor progression. In this regard, the present invention addresses the knowledge gap by developing new automated, quantitative DCE-MRI capability and automated threshold values for use in the standard patient care setting as well as in clinical trials. [0016] Advantageously, the standardization of neuroimaging endpoint biomarkers as provided by the methods and systems described herein is recognized as an unmet need for brain tumor drag development (9). Therefore, the provided standardization and automation of the DCE-MRI can be incorporated into the standardized response criteria such as RANO.

BRIEF DESCRIPTION OF THE DRAWINGS

[0017] For a further understanding of the nature, objects, and advantages of the present disclosure, reference should be had to the following detailed description, read in conjunction with the following drawings, wherein like reference numerals denote like elements and wherein:

[0018] FIGURE 1: Schematic illustration of a Blackbox MR! imaging system.

[0019] FIGURE 2: Sparse pre-contrast Ti mapping reconstruction flowchart. Note that phase information was directly copied from Magnitude phase separation to VFA simulation.

[0020] FIGURE 3: Network architecture: Input filters with temporal convolutions of different kernel lengths, encoder with dense layers.

[0021] FIGURE 4: Patients determined to have pseudoprogression demonstrate improved survival (R Young el al. 2011).

[0022] FIGURE 5 : The determination of Pseudoprogression is critical in the therapeutic triage of patients into continuation of Temozolomide versus consideration for other therapies or clinical trials. The challenges with making this determination has prompted the consideration of advanced MRI techniques such as DSC-MRI and DCE-MRI in the RANG and more recent iRANO criteria.

[0023] FIGURES 6A and 6B: Images and Bloch-Siegert B1 + maps from (A) an ISMRM system phantom (20.1 cm diameter, T i and conductivity matched to healthy brain tissue) (B) a healthy volunteer, at 3T. The phantom represents the range and spatial variation.

[0024] FIGURE 7: Olea Software Development Kit (OSDK) architecture. Reconstruction plug-ins will be implemented in OSDK through collaboration between engineers at USC and Olea Medical. This enables rapid prototyping and ease of deployment to the three clinical sites. [0025] FIGURE 8: Brain tumor K*”” maps from 5 (out of 15) representative cases for an under-sampling rate of 40. Top row contains reference maps obtained from fully-sampled data, followed by GLR, error between reference and GLR, MOCCO, error between reference and MOCCO, DL, and error between reference and DL. In general, GLR reconstructions appear excessively smooth (red arrows) and MOCCO reconstructions appear noisy (yellow arrows).

[0026] FIGURE 9: Brain tumor vp maps from 5 (out of 15) representative cases for an under- sampling rate of 40. Top row contains reference maps obtained from fully-sampled data, followed by GLR, error between reference and GLR, MOCCO, error between reference and MOCCO, DL, and error between reference and DL In general, DL results were comparable to MOCCO, and sharper than GLR (red arrows).

[0027] F1GURES 10 A, 10B, 10C, 10D and 10E: RMSE for the tumor ROI as a function of under-sampling rate for the 5 tumor cases (each column) shown in Fig. 1 and 2, (top) Anatomic images; (middle)

(bottom) vp. The DL method with temporal dictionary has reduced NRMSE compared to GLR and MOCCO at all under-sampling factors varying from 20 to 100. The NRMSE due to thermal noise in the fully sampled reference is shown as the shaded gray region. This indicates that the improvement with respect to GLR is substantial, but the improvement with respect to MOCCO is minor.

[0028] FIGURE 11A, 1 IB, and 1 1 C: Average RMSE across 15 cases as a function of undcr- sampling rate. (A) Anatomic images; (B) K trail* ; (C) vp. The NRMSE due to thermal noise in the fully sampled reference is shown as the shaded gray region. This indicates that the improvement with respect to GLR is substantial, but the improvement with respect to MOCCO is minor.

[0029] FIGURE 12: Data trimming illustration. Data from A repetitions for each of the first six flip angles and data from first B repetitions for the DCE angle were preserved, in which A ranges from 2000 to 5000 with a step size of 100 and B ranges from 8000 to 20000 with a step size of 1000. This results in a total of 403 datasets each corresponding to a unique (A, B) pair. [0030] FIGURES 13A and 13B: While matter R01 (A) and Ti, Mo, and anatomical image difference maps among 403 datasets (B). (5000,20000) was chosen as reference, so that its errors are

0.

[0031] FIGURES 14A and 14B: Statistics analysis of white matter Ti using different (A, B) settings. Choice of (A, B) had only subtle impact on the Ti (A) mean and (B) standard deviation. All histograms were approximately Gaussian with ~l000 ms mean, and ~103 ms standard deviation.

[0032] FIGURES 15 A, 15B, and 15C: Three different convergence curve during reconstruction. The red line in the left-most plot indicates the noise level in k-space. Convergence can be observed in all curves, however, there were still significant k-space residuals.

[0033] FIGURES 16A and 16B: Evaluation in a brain tumor digital reference object. (A):

Comparison of PK parameter maps. (B): Correlation of standard deviation as predicted by proposed method and conventional Monte Carlo method.

[0034] FIGURES 17A, 17B, 17C, and 17D: Contour plots of cost functions for PK parameter planes. The dots are the true value. Dark cloud shows samples from posterior distribution which are used to generate concentration time curves in the blue area.

[0035] FIGURE 18: Results from 6 (out of 17) representative brain tumor cases, (top 4 rows) maps, (bottom 4 rows) Vp maps. Each block contains reference maps, under-sampled with direct reconstruction, direct reconstruction followed by DCRN dcnoising, and error between the reference and DCRN results.

[0036] FIGURE 19A and 19B: Average NRMSE across 17 brain tumor test data, as a function of under-sampling factor. The proposed denoiser (DCRN) outperformed no

denoising (Direct), and two conventional denoisers (NLM & BM3D) at all under-sampling factor

[0037] FIGURE 20: Digital reference object. Panel a) shows the discrete brain model, the simulated magnitude image at peak intensity (prior to coil sensitivity encoding), and the pharmacokinetic parameter maps for vp and K^l. The tumor ROI is marked in red. Panel b) contains DRO parameters. [0038] FIGURE 21: Sampling schemes under investigation. Shown are representative patterns for under-sampling factor R— 5. All sampling patterns are isotropic 2D patterns. For all sampling patterns the first time frame is fully sampled. For lattice based sampling different time frames are rotated with a deterministic scheme, for random sampling different time frames have new realizations.

[0039] FIGURE 22: Cram6r-Rao lower bounds for pharmaco-kinetic parameters vp (top row) and K* (bottom row) for under-sampling rates R~2,5,10,l5. Bounds are computed for pre-contrast white matter SNR=10, flip angle of 24°, and 50 time frames at 5s temporal resolution. First column shows true PK parameters, all other columns illustrate the bounds for the whole brain slice in Figure 1. vp is measured in percent, K^l in [min ¹]. Horizontal axes show pixels in (stacked) image array. Vertical axes show standard deviations of PK parameters. There is no difference in the performances of the different sampling patterns.

[0040] FIGURE 23: Cram6r-Rao lower bounds for pharmaco-kinetic parameters vp (top row) and K* (bottom row) for undersampling rates R=2,5 , 10, 15. Bounds are computed for pre-contrast whi te matter SNR=10, flip angle of 24°, and 50 time frames at 5s temporal resolution. First column plots true PK parameters, all other columns illustrate the bounds in the tumor ROI as specified in Figure 20. vp is measured in percent, K* in [min^-1]. Horizontal axes show pixels in image array. Vertical axes show standard deviations of PK parameters. Confirming the results of Figure 22, there is no difference in the minimum achievable variance for pharmaco-kinetic parameter estimation between the sampling schemes.

DETAILED DESCRIPTION

[0041] Reference will now be made in detail to presently preferred compositions, embodiments and methods of the present invention, which constitute the best modes of practicing the invention presently known to the in ventors. The Figures are not necessarily to scale. However, it is to be understood that the disclosed embodiments are merely exemplary of the invention that may be embodied in various and alternative forms. Therefore, specific details disclosed herein are not to be interpreted as limiting, but merely as a representative basis for any aspect of the inventi on and/or as a representative basis for teaching one skilled in the art to variously employ the present invention. [0042] It is also to be understood dial this invention is not limited to the specific embodiments and methods described below, as specific components and/or conditions may, of course, vary. Furthermore, the terminology used herein is used only for the purpose of describing particular embodiments of the present invention and is not intended to be limiting in any way.

[0043] It must also be noted that, as used in the specification ami the appended claims, the singular form“a,”“an,” and“the” comprise plural referents unless the context clearly indicates otherwise. For example, reference to a component in the singular is intended to comprise a plurality of components.

[0044] The term“comprising" is synonymous with“including,”“having,”“containing," or “characterized by.” These terms are inclusive and open-ended and do not exclude additional, unrecited elements or method steps.

[0045] The phrase“consisting of* excludes any element, step, or ingredient not specified in the claim. When this phrase appears in a clause of the body of a claim, rather than immediately following the preamble, it limits only the element set forth in that clause; other elements are not excluded from the claim as a whole.

[0046] The phrase“consisting essentially of’ limits the scope of a claim to the specified materials or steps, plus those that do not materially affect the basic and novel characteristic's) of the claimed subject matter.

[0047] With respect to the terms“comprising,"“consisting of," and“consisting essentially of," where one of these three terms is used herein, the presently disclosed and claimed subject matter can include the use of either of the other two terms.

[0048] It should also be appreciated that integer ranges explicitly include all intervening integers. For example, the integer range 1-10 explicitly includes 1, 2, 3, 4, 5, 6, 7, 8, 9, and 10. Similarly, the range 1 to 100 includes 1 , 2, 3, 4. . . .97, 98, 99, 100. Similarly, when any range is called for, intervening numbers that are increments of the difference between the upper limit and the lower limit divided by 10 can be taken as alternative upper or lower limits. For example, if the range is 1.1. to 2.1 the following numbers 1.2, 1.3, 1.4, 1.5, 1.6, 1.7, 1.8, 1.9, and 2.0 can be selected as lower or upper limits. In the specific examples set forth herein, concentrations, temperature, and reaction conditions (e.g., magnetic field stress, RF frequency, pressure, pH, etc.) can be practiced with plus or minus 50 percent of the values indicated rounded to three significant figures. In a refinement, concentrations, temperature, and reaction conditions (e.g., magnetic field stress, RF frequency, pressure, pH, etc.) can be practiced with plus or minus 30 percent of the values indicated rounded to three significant figures of the value provided in the examples. In another refinement, concentrations, temperature, and reaction conditions (e.g., magnetic field stress, RF frequency, pressure, pH, etc.) can be practiced with plus or minus 10 percent of the values indicated rounded to three significant figures of the value provided in the examples.

[0049] Throughout this application, where publications are referenced, the disclosures of these publications in their entireties are hereby incorporated by reference into this application to more fully describe the state of the art to which this invention pertains.

[0050] The term "server" refers to any computer, computing device, mobile phone, desktop computer, notebook computer or laptop computer, distributed system, blade, gateway, switch, processing device, or combination thereof adapted to perform the methods and functions set forth herein.

[0051] When a computer or computing device is described as performing an action or method step, it is understood that the computing devices is operable to perform the action or method step typically by executing one or more line of source code. The actions or method steps can be encoded onto non-transitory memory (e.g., hard drives, optical drive, flash drives, and the like).

[0052] The term“computer or computing device” refers generally to any device that can perform at least one function, including communicating with another computer or computing device.

[0053] The term“subject” refers to a human or other animal, including birds and fish as well as all mammals such as primates (particularly higher primates), horses, birds, fish sheep, dogs, rodents, guinea pigs, pig, cat, rabbits, and cows.

[0054] Abbreviations: [0055] “AIF” means arterial input function.

[0056] “Bi^*” means the RF excitation field.

[0057] “DCE-MRI” means dynamic contrast enhanced magnetic resonance imaging.

[0058] “DSC” means dynamic susceptibility contrast.

[0059] “DWI” means diffusion-weighted imaging.

[0060] “FA” means flip angle.

[0061] “MRT means magnetic resonance imaging.

]0062] “Mo” means the magnitude of the magnetization, e.g., the longitudinal thermal equilibrium state of the net magnetization.

[0063] “Mo” means equilibrium magnetization.

]0064] ‘OS” means overall survival

[0065] ‘OSDK” means Olea Software Development Kit.

[0066] “PFS" means progression-free survival.

[0067] “rCBV” means relative cerebral blood volume.

[0068] “RF” means radiofrequency.

[0069] “ROl” means region of interest

[0070] “Ti” is the longitudinal relaxation time (i.e., spin-lattice relaxation time).

[0071] “T₂” is the transverse relaxation time (i.e., spin-spin relaxation time).

[0072] “TMZ” means temozoiomide.

[0073] “TR” means repetition time. [0074] “TE” means echo time.

[0075] “VFA” means variable flip angle.

[0076] In an embodiment, an MRI system for delivering standardized, automated, quantitative “black box” high-resolution DCE-MRI acquisition and processing for use in the clinic and in oncology clinical treatment is provided. In general, the black box magnetic resonance system includes a magnetic resonance imaging subsystem and a controller subsystem. In one variation, the system is operable to perform T₂* -weighted dynamic susceptibility contrast MRI. In another variation, the system is operable to perform Ti-weighted DCE-MRI. The magnetic resonance imaging subsystem includes a magnetic assembly having a polarizing magnetic coil and gradient coil assembly and receiving coils that collect magnetic resonance imaging data from a subject being treated for a tumor (e.g., cancer) from which dynamic contrast magnetic resonance images are constructed. The subject maybe treated with a chemotherapeutic agent and/or radiation. The control subsystem is operable to perform calibration scans and determine values fee· MRI biomarkers from magnetic resonance images obtained from the subject In a refinement, calibration scans include B1+ mapping, and pre-contrast Ti mapping with spatial resolution and coverage that match an area from which the magnetic resonance imaging data is obtained. In this regard, the control subsystem includes one or more computers running algorithms to construct the magnetic resonance images and½ calculating biomarker values and/or determining prognosis or optimal treatment protocols. Characteristically, the MRI biomarkers includes tracer kinetic parameters obtained from a tracer kinetic model. In a refinement, the control subsystem is operable to denoise the magnetic resonance images.

[0077] With reference to Figure 1 , imaging system 10 includes magnetic resonance imaging subsystem 14. Magnetic resonance imaging subsystem 14 includes coils 16 from which the (k, t) space data is collected, Pulse sequencer 20, data acquisition subsystem 22, gradient system 24, magnetic assembly 26, and RF subsystem 28. Magnetic assembly 26 includes polarizing magnetic coil 30 and gradient coil assembly 32. Imaging system 10 also includes a control subsystem 36 includes programmable computer 38. Control subsystem 36 can control the implementation of calibration scans and determine values for MRI biomarkers from magnetic resonance images obtained from the subject [0078] As set forth above, the MRI biomarkers typically include tracer kinetic parameters obtained from a tracer kinetic model. Computer 38 can also be operable to compare values or time trends of MRI biomarker values to standard biomarkers values or trends determined from prior patient trials. In a variation, the control subsystem 36 jointly estimates contrast concentration versus time images and tracer parameter maps from under-sampled (k,t) space data. Specific methods that computer 38 can implement are found in Y. Guo el al., Joint Arterial Input Function and Tracker Kinetic Parameter Estimation from Undersampled Dynamic Contrast-Enhanced MRI Using a Model Consistency Constraint , Magn Reson Med. 2018 May;79(5):2804-2815. doi: 10.1002/mrm.26904. Epub 2017 Sep 14; Y. Zhu et al., COCAR.T: GOlden-angle CArtesian randomized time-resolved 3D MRI, Magnetic Resonance Imaging 34 (2016) 940-950; and Y. Guo et al., Direct Estimation of Tracer-Kinetic Parameter Maps From Highly Undersampled Brain Dynamic Contrast Enhanced MRI, Magn Reson Med. 2017 Oct; 78(4): 1566-1578; the entire disclosures of which are hereby incorporated by reference. It should be appreciated that computer 38 can implement any combination of the methods set forth in these references including all of these methods. In addition, computer 38 can implement the standardization and denoising operations described herein. The combination of all this functionality allows for the“black box” aspect in which a subject’s MRI data can be inputted and useful measurements of the biomarkers outputted. In a refinement, computer 38 provides an interface 61 to the user that presents a listing of selections that can be performed such as whether or not a calibration is to be performed and what calibration method is to be used; to apply a previously measured calibration; whether or not biological markers are to be determined and which marker are to be calculated; whether or not to denoise the MRI image data; and what kind of sampling is be performed (e.g., sparse sampling, under sampling parameters, lull sampling and the like); which image reconstruction technique is to be used; and the like. Once selections are made, imaging system 10 can automatically implement the selected methods without further user input to ideally output the biomarkers.

[0079] In one refinement, the tracer kinetic model is a Patlak model and q includes

(sometimes herein

is abbreviated as and where is a transfer constant from blood

plasma into extracellular extravascular space (EES) and is fractional plasma volume. In another

refinement, the tracer kinetic model is an extended Tofts model and Q includes and v_e

where K¹™” is a transfer constant from blood plasma into extracellular extravascuiar space (EES), v_p is fractional plasma volume, K^l3p is a transfer constant from EES back to the blood plasma, and v_e is a fractional EES volume. In still other variations, control subsystem 36 is operable to automatically select the tracer kinetic model by specifying a col lection of possible models (nested or not nested) from which a model identification method is applied to select a model for each voxel or region. In some variations, control subsystem 36 is operable to construct dynamic images using a consistency constraint is applied to construct dynamic images from the magnetic resonance imaging data, the consistency constraint including a sum of a data consistency component and a model consistency component. In additional variation, control subsystem 36 performs one or more or all of: providing automatic arterial input function selection; performing motion correction; performing hierarchical TK modeling; and measuring biomarkers in a region-of-interest Control subsystem 36 can also be operable to determine an arterial input function or vascular input fiinction from the magnetic resonance imaging data, wherein the arterial input function includes a time variation of a magnetic resonance contrast agent at one or more predetermined locations in an artery of the subject. This feature can also be offered as a user selection in interface 61.

[0080] Details of the Patlak and extended Tofts-Kety (ETK) model are set forth in P. S. Tofts,

G. Brix, D. L Buckley, J. L Evelhoch, E. Henderson, M. V Knopp, H. B. Larsson, T. Y. Lee, N. a Mayr, G. J. Parker, R. E. Port, J. Taylor, and R. M. Weisskoff, "Estimating kinetic parameters from dynamic contrast-enhanced T(l)-weighted MR] of a diffusablc tracer standardized quantities and symbols. ,"J. Magn. Reson. Imaging, vol. 10, no. 3, pp. 223— 32, Sep. 1999; and Parker, G.J. and Buckley, D.L., 2005. Tracer kinetic modelling for T1 -weighted DCE-MRI. In Dynamic contrast- enhanced magnetic resonance imaging in oncology (pp. 81-92). Springer Berlin Heidelberg; the entire disclosures of these publications is hereby incorporated by reference. In general, the Pailak and ETK having kinetic parameters: which is a transfer constant from blood plasma into extracellular

extravascuiar space (EES) and v_p which is fractional plasma volume. The ETK model also has kinetic parameters which is a transfer constant from EES back to the blood plasma and v_e which is a fractional EES volume The concentration for the contrast agent in this model is

calculated from:

where t is time;

C_t is the equilibrium concentration of contrast agent in whole tissue;

C_P is the equilibrium concentration of contrast agent in plasma;

C_e is the equilibrium concentration of contrast agent in extracellular extravascular space.

In the simple Toft model, the contrast concentration in the whole tissue can be determined from:

When v_p is considered such as in the ETK model, the contrast concentration in whole tissue can be determined from:

When the time dependent concentrations, i.e., Ci and Cp, are known, these equations can be inverted to estimate the kinetic parameters. Additional details for converting TK parameter (e.g., K trans

> Vp) maps to contrast concentration over time using the Patlak model is provided by P. S. Tofts, G. Brix, D. L. Buckley, J. L. Evelhoch, E. Henderson, M. V Knopp, H. B. Larsson, T. Y. Lee, N. a Mayr, G. J. Parker, R. E. Port, J. Taylor, and R. M. WeisskofT, "Estimating kinetic parameters from dynamic contrast-enhanced T(l)-weighted MRI of a diftusable tracer, standardized quantities and symbols. ,"J. Magn. Resort Imaging, vol. 10, no. 3, pp. 223-32, Sep. 1999 and Patlak C.S., Blasberg R.G., Fenstermacher J.D. Graphical evaluation of blood-to-brain transfer constants from multiple-time uptake data. J. Cereb. Blood Flow Metab. 1983;3:1-7; the entire disclosures of these publications is hereby incorporated by reference. [0081] In a variation, the dynamic images and the tracer kinetic parameter maps can be estimated jointly by enforcing a consistency constraint. Details of this approach are set forth in U.S. Pat. Serial No. 16384795 filed April 15, 2019; the entire disclosure of which is hereby incorporated by reference. The consistency constraint includes a weighted sum of a data consistency component and a model consistency component. The data consistency component assesses how well the acquired magnetic resonance data (e.g., raw or Fourier transformed data) for each voxel in a Field Of View is approximated by the dynamic images calculated from the estimate of concentration-time curves of the contrast agent for each voxel in a Field Of View. Therefore, a first difference can be between the acquired magnetic resonance data for each voxel in a Field Of View and data calculated from the estimated concentration-time curves for each voxel in a Field Of View. Similarly, the model consistency component assesses how well the concentration-time curves calculated from the estimate of the tracer kinetic model and its plurality of parameters approximate measured concentration-time curves of the contrast agent for each voxel in a Field Of View. Therefore, a second difference can be determined between the measured concentration-time curves for each voxel in a Field Of View and concentration-time curves calculated from the tracer kinetic model for each voxel in a Field Of View. The consistency constraint seeks to minimize the combination of both differences (e.g., see the b parameter below).

[0082] In a variation, the consistency constraint is provided by the minimization represented by Equation 1:

where:

y is a signal operator which converts concentration-time-curves (per voxel) C of a contrast agent to an image intensity time series;

U is an under-sampling mask;

F is a Fourier transform, and in particular, the discrete Fourier transform matrix or a Fast-Fourier- Transform (FFT) algorithm;

E is a sensitivity encoding matrix providing the spatial relative sensi tivities of the pickup coils.

So is image intensity for each voxel prior to contrast agent arrival; y is the under-sampled magnetic resonance imaging data:

C is a measured concentration-time curves of the contrast agent for each voxel in the Field Of View; R(q) is a predicted concentration distribution of the contrast agent from the selected tracer kinetic model;

P is a penalty or weight factor for the model consistency component; and

Q are tracer kinetic parameters such

' and for the Patiak and ETK models and K^ep and v_e for the

ETK as described herein.

It should be noted that U, F, and E are linear operators which can be expressed as matrices, while y can be either a linear or non-linear operator. So is expressed as matrix with each matrix entry being a value for a spatial point or voxel. C is expressed as matrix with each matrix entry being a value for a spatial point or voxel and time point. In equation 1, the first hnorm represents the data consistency component and the second L norm represents model consistency component. Additional details of this variation are found in US. Pat. Appl. No. 16/384795 and in Y. Guo et al„ Joint Arterial Input Function and Tracker Kinetic Parameter Estimation from Undersampled Dynamic Contrast- Enhanced MRl Using a Model Consistency Constraint, Magn Reson Med. 2018 May,79(5):2804- 2815. doi: 10.l002/mrm.26904. Epub 2017 Sep 14; the entire disclosures of which are hereby incorporated by reference.

[0083] With reference to Figure 1 , computer 38 includes central processing unit (CPU) 40, mcmoiy 42, display 44 and input/output interface 46 which communicate over buses 48. Computer

38 communicates with magnetic resonance imaging subsystem 14 and input devices such as a keyboard and mouse via input/oulput interlace 46. In one variation, memory 42 includes one or more of the following: random access memory (RAM), read only memory (ROM), CDROM, DVD, disk drive, tape drive. The methods set forth above (i.e., the feature for which control subsystem 36 is operable for) are implemented by routines that is stored (i.e., encoded) in non-transitoiy memory 50 and executed by the CPU 40. Computer 38 is electrically coupled to pulse sequencer 20 and data acquisition subsystem 36 of magnetic resonance system 14. Pulse sequencer 20 is also in electrical communication with data acquisition subsystem 36. Pulse sequencer 20 receives instruction from computer 38 to operate a gradient system 24 and a radiofrequency (RF) system 48. RF system 48 includes RF transmitters for this putpose in order to generate the prescribed pulses and one or more receiver channels for receiving signal from coils 16. Gradient waveforms necessary to perform the magnetic pulses set forth above are produced and applied to the gradient system 38, which excites gradient coils in coil assembly 50 to produce the magnetic field gradients and used for position encoding magnetic resonance signals. RF waveforms are applied by the RF system 48 to the RF coil 16, or a separate local coil (not shown), in order to perform the prescribed magnetic resonance pulse sequence, Responsive magnetic resonance signals detected by the RF coil 16 or a separate local coil (not shown), arc received by the RF system 48 where they are processed accordingly.

[0084] Computer 38, and in particular, CPU 40 in conjunction with magnetic resonance imaging system 14 implements the methods set forth above as follows. Computer 38 send control signals to pulse sequencer 20 to control gradient system 24 to apply a gradient magnetic field pulse from polarizing magnetic coil 42 to a subject 56 along a first direct. A subject who has been administered a contract agent is placed in this gradient magnetic field. Computer 38 also sends control signals to pulse sequencer 20 to RF system 28 to apply an excitation radioftequency pulse to the subject during the first gradient magnetic field pulse where the excitation radiofrequency pulse is resonant with a region in the subject. Computer 38 also sends control signals to pulse sequencer 20 to control gradient system 24 to apply gradient magnetic field pulses to the subject after the first gradient magnetic field pulse in order to provide spatial encoding. RF system 28 receives an output signal from the subject 56 during the second gradient magnetic field pulse such that magnetic resonance imaging data is collected from the subject for a tissue or organ. This output signal is ultimately transferred to computer 38 for processing. Computer 38, and in particular CPU, applies a selected tracer kinetic model to the magnetic resonance imaging data to estimate tracer kinetic parameter maps; and applying the tracer kinetic model to the magnetic resonance imaging data to estimate tracer kinetic parameter maps; and reconstructing tracer kinetic maps and dynamic images from the tracer kinetic parameter maps. The tracer kinetic model can be user selected in advance or automatically selected or determined by computer 38.

[0085] In a variation, system 10 applies dictionary learning (DL) to assist in determining the tracer kinetic parameters. DL involves finding a basis set that is optimal for a specific set of signals to provide the sparsest possible representation for that particular signal (Aharon M, Elad M, Bruckstein A. K-SVD: An algorithm for designing overcomplete dictionaries for sparse representation. IEEE Trans. Signal Process. 2006;54(l 1 ):4311 -4322). The reconstruction problem can be expressed as the following optimization:

where i is a label for the summation (e.g. an integer), c is a label for the summation (e.g an integer), s(r,t) are dynamic anatomic images, C_c(r) are the coil-sensitivities estimated using sum-of-squares coil combination, S_c(k,t) is die under-sampled (k,t) space data with kk representing the kk-spaee coordinates, r denoting the image domain spatial coordinates, c and t are the coil and time dimensions, Ftt is the tinder-sampling Fourier transform, D is the learned dictionary, a is the sparse representation, j|.jjo counts the number of non-zero elements, and Pj is an operator that extracts 3D patches as a column vector. The optimization problem is non-convex and can be solved by splitting into two simpler subproblems. Iteratively solving these two subproblems will yield an approximate solution. Results were compared against GLR, MOCCO, and the fully-sampled reference. A fixed number of iterations was empirically chosen for both DL and GLR based on convergence behavior in retrospective under-sampling studies. Dictionary Training is performed as follows. First, we assume s(r,t) is fixed while D and aa are free variables. This step finds a dictionary that can sparsely represent the dataset using k-SVD algorithm (Aharon et al.). The trained dictionary is later used to sparsely code the dataset using orthogonal matching pursuit (OMRc Tropp JA, Anna, and Gilbert C. Signal recovery from random measurements via orthogonal matching pursuit. IEEE Trans. Inf. Theory. 2007;33(12):46S3-4666). Data consistency is determined as follows. The second step is data consistency, in which s(r,t) is the only free variable. In this step, we insert the measured data at the sampling locations. Experimental verification of the present variation is set forth below in the Sparse DCE-MRI using a Temporal Constraint Learned from Clinical Data Experiments section. Moreover, additional details of the present variation is found in S. Gutta et al., Sparse DCE-MRI using a Temporal Constraint Learned from Clinical Data. Proc. Inti. Soc. Mag. Reson. Med. 27 (2019), 4749; the entire disclosure of which is hereby incorporated by reference. [0086] Another reconstruction method that can be implemented by system 10 is set forth in

Guo Y, Lebel RM, Zhu Y, Lingala SG, Shiroishi MS, Law M, et al. High-resolution whole-brain DCE-MRI using constrained reconstruction: Prospective clinical evaluation in brain tumor patients. Med Phys. 2016;43(5):2013~23; the entire disclosure of which is hereby incorporated by reference. Similarly, another reconstruction method and/or method of determining tracer kinetic parameters that can be implemented by system 10 is set forth in U.S. Pat Pub. No. 20170325709; the entire disclosure of which is hereby incorporated by reference.

[0087] In a variation, system 10 applies a sparse Ti mapping method and characterizes a tradeoff between data acquisition and T statistics, using a variable flip angle (VFA) approach and sparse Cartesian spiral sampling pattern, with image domain wavelet sparsity constraint. This method provides the necessary high-resolution whole-brain T t/ Mo maps for DCE-MRI tracer kinetic analysis. In this variation, a pre-contrast Ti mapping is applied and die trade-off between data acquisition time and T i mapping accuracy is determined. In accordance with this method, pre-contrast Ti / Mo mapping is performed by solving the following constrained inverse problem:

where the estimated anatomic image magnitude is

TR means repetition time, p concatenates Ti and Mo into a vector, Bi is the radiofrequency field strength, Fu is the under-sampled Fourier transform, S is the coil sensitivity, m is VFA images (e.g., a label), D converts Ti and Mo values into VFA images, Y is a sparsifying transform (e.g,, wavelet), a is the flip angle, d is measured k-space data, and l is a regularization parameter. This problem is solved using POCS iterations, which alternate between thresholding wavelet coefficients and forcing data consistency. A flowchart of this workflow is shown in Figure 2. Experimental verification of the present variation are set forth below in the Sparse Pre-Contrast T i Mapping for DCE-MRI Calibration Experiments section. Moreover, additional details are provided in A. Zhu et al., Sparse Pre-Contrast TJ Mapping for DCE-MRJ Calibration, Proc. Inti. Soc. Mag. Reson. Med. 27 (2019) 4544; the entire disclosure of which is hereby incorporated by reference.

[0088] In a variation, system 10 applies machine learning (ML) to simultaneously estimate

DCE- MRI pharmacokinetic (PK) parameters and uncertainty. Dynamic contrast enhanced (DCE)

MRI aims to estimate sub-voxel parameters of pathology pharmacokinetics through fitting of pharmacokinetic models to contrast agent concentration-time curves. Many of the involved cost functions from non-linear models are not strongly convex or not even convex which introduces ambiguity in the parameters and deteriorates accuracy and precision due to susceptibility to noise and initializations. Development of such estimators is furthermore challenged by lack of reference methods and ground truth, and measures of precision can only be obtained through time-consuming Monte- Carlo simulations (MCS) with multiple noise realizations and initializations, or variance estimation through linear error propagation. The present variation explores the use of re-enforcement learning of neural networks to estimate posterior distributions from which ranges of possible pharmaco-kinetic parameter as well as various metrics of certainty of the estimation can be derived. The neural network (Figure 3 ) consists of two separate input filter stages for measured concentration-time curves C" and arterial input functions (AIF) whose output is concatenated and fed into an encoder network of four dense layers. Each layer except for the last is followed by a LeakyReLU layer; the last is activated by sigmoids. The output of the last layer are 9 parameters Y per pixel for location, scale, and rotation of a uniform ellipsoidal distribution qy which serves as approximation to the posterior distribution. The cost function for training is given by:

The first term enforces data consistency of the PK parameters Q drawn from the posterior q% . The second term, a negative entropy, enforces spread of the posterior to cover all possible PK parameters that could explain the data. Training data consisted of 1572 patches of 20x20 pixels and 50 time points whose noisy concentration time curves are generated by random maps of pathologically realistic parameters for the extended Tofts model with 61 AIFs with synchronized bolus arrival measured from clinical exams at our institution. Test data was taken from a pathologically and anatomically realistic digital reference object. Experimental verification is found below in the Posterior Approximation For Simultaneous DCE-MRI Pharmaco-Kinetic Parameter And Uncertainty Estimation Experimental section.

[0089] In another variation, computer 38 is operable to apply machine learning-based denoising to improve the quality of brain tumor DCE-MRI TK maps, specifically those reconstructed from sparsely sampled (k, t) space data. Brain DCE-MRI is a powerful imaging technique to measure neurovascular parameters and can be used for determining tumor response to therapy. Tumor characteristics are determined using quantitative parameters, such as vascular permeability and

fractional plasma volume (Vp), that are estimated using TK modeling. Accurate and precise TK estimation is a prerequisite for clinical adoption. Recently, sparse DCE-MRI with constrained reconstruction has provided substantially improved spatial resolution and coverage. However, the TK maps estimated from these approaches suffer from increased variance at higher under-sampling rates. In this variation, computer 10 applies deep convolutional residual network (DCRN) for denoising these maps.

[0090] In another variation, system 10 can apply a number of sampling strategies such as lattice and random under-sampling, or between their uniform and variable density variants. Experiments showing the effects of the different sampling techniques are set forth below in the Influence Of Whole-Brain DCE-MRI (K,T) Sampling Strategies On Variance Of Pharmaco-Kinetic Parameter Estimates Experiments section. Additional details of this variation are found in Y. Bliesener cl al., Influence Of Whole-Brain DCE-MRI (K T) Sampling Strategies On Variance Of Pharmaco-Kinetic Parameter Estimates, Proc. Inti. Soc. Mag. Reson. Med. 26 (2018) 0555; the entire disclosure of which is hereby incorporated by reference. In another example, a 3D Cartesian sampling scheme as set forth in U.S. Pat. No. US1020339 and Y. Zhu et al., 3D Cartesian sampling scheme which is well suited for time-resolved 3D MR] using parallel imaging and compressed sensing time- resolved 3D MRJ using parallel imaging and compressed sensing ; Magnetic Resonance Imaging 34

[2016] 940-950; the entire disclosures of which is hereby incorporated by reference. In this latter sampling scheme, termed GOIden-angle CArtesian Randomized Time-resolved (GOCART) 3D MRI, golden angle (GA) Cartesian sampling is applied with random sampling of the ky-kz phase encode locations along each Cartesian radial spoke. This method can be applied in conjunction with constrained reconstruction of retrospectively and prospectively undersampled in-vivo dynamic contrast enhanced (DCE) MR1 data and simulated phantom data. Moreover, system controller of Figure 1 controls the magnet, gradient coils, RF pulse transmitter, and RF receiver so as to generate data representative of at least a portion of a composition of an object, including controlling the gradient coils and RF receiver so as to cause MRI data to be acquired that includes information about at least one attribute of the object at different points in lime and that represents an incomplete sample of a portion of k-space that is a Fourier transform of the object. The MRI data is acquired by foil Cartesian sampling along a frequency encoding direction kx with subsampled and/or reordered phase encoded sampling in a ky-kz plane wherein the system controller causes k-space data to be acquired along golden angle radial spokes in the ky-kz plane.

[0091] In another embodiment, a method for determining a patient prognosis or monitoring patient treatment with the system set forth above is provided. The method includes steps of administering a magnetic resonance contrast agent to a subject, collecting magnetic resonance imaging data from the subject for a tissue or organ, selecting a tracer kinetic model to be applied to the magnetic resonance imaging data, reconstructing dynamic images from the magnetic resonance imaging data, applying the tracer kinetic model to the reconstructed dynamic images to estimate tracer kinetic parameter maps, determining a patient prognosis and/or monitoring effectiveness of treatment; and adjusting patient treatment if necessary. The method of the present embodiment is not limited by the type of tumor which can be selected from the group consisting of brain, breast, prostate, liver, kidney, lung, heart, thyroid, pancreas, spleen, intestine, uterus, ovary, limbs, spine, bones, and eyes. The method is found to be particularly useful for monitor brain tumor. Typically, the steps of collecting magnetic resonance imaging data from the subject for a tissue or organ, selecting a tracer kinetic model to be applied to the magnetic resonance imaging data, the tracer kinetic model being defined by a plurality of tracer kinetic parameters, reconstructing dynamic images from the magnetic resonance imaging data, and applying the tracer kinetic model to the reconstructed dynamic images to estimate tracer kinetic parameter maps.

[0092] The MR! systems set forth above build upon an existing qualitative DCE-MRI approaches that provides high isotropic spatial resolution and whole brain spatial coverage. The method employed translate several previously validated component technologies to enable extraction of quantitative parameters. A single protocol and automated analysis package is provided that 1) integrates calibration scans (e.g., B1+ mapping, and pre-contrast Ti mapping) with spatial resolution and coverage that match the DCE-MRI; 2) translate data processing methods: automatic arterial input function selection, motion correction, hierarchical TK modeling, and measurement of region-of- interest biomarkers; and 3) clinically evaluate the quantitative DCE-MRI package in ongoing multi - center neuro-oncology trials, to determine equivalence across sites, diagnostic cut-points for, and advantages in characterizing treatment response relative conventional biomarkers including bidirectional anatomic assessment and DWI, and relative to qualitative and semi-quantitative DCE-MRI.

[0093] In another aspect, the present invention provides novel reproducible, automated outcome markers, addressing one of the greatest challenges the field of neuro-oncology. Some of the other potential applications in neuro-oncology would be for the determination of tumor biology (using automated DCE MR] to predict 1DH wild-type vs mutation in high grade gliomas), differentiation between radiation necrosis and recurrent tumor This automated black box approach can also be translated to the DCE-MRI of prostate, breast and other organs.

[0094] In another aspect, the methods and systems set forth herein standardize, automate and translate quantitative DCE-MRI biomarkers for the assessment of brain tumor response to therapy in clinical neuro-oncology multi-center trials, and in the radiology clinic. Increases in brain tumor size and contrast enhancement have historically signified tumor progression, and decreases signified treatment response. However, with new and novel chemotherapy and immunotherapy agents, these simple changes no longer suffice for treatment decisions (9) because of pseudoprogression, pseudoresponse, and post chemotherapeutic/radiation changes. This has produced a major unmet need for improved biomarkers of response to therapy, especially for brain tumor patients ofien regarded as the direst prognostic group, especially as imaging biomarkers and progression-free survival are becoming accepted surrogate endpoints in clinical trials.

[0095] MRI system of Figure 1 standardizes the prescription of a whole-brain DCE-MRI scan with respect to such parameters as resolution, coverage, time duration, encoding directions, flip angle, and the like. The system also standardizes the acquisition of calibration scans including patient- specific RF transmit inhomogeneity maps and pre-contrast Ti maps with spatial resolution and coverage that match the DCE-MRI acquisition. The scan protocol, can be implemented on many of the currently used MRI systems (e.g., GE MRI scanners).

[0096] In another variation, automated and standardized data processing techniques for quantitative brain DCE-MRI are deployed. Previously validated DCE-MR1 data processing methods are integrated into a thin/zero client solution. These include 1) patient-specific A1F selection and de- noising by principal component analysis, 2) option of nested tracer kinetic model selection, 3) ROl- based metrics, and 4) automated detection of longitudinal changes when using a single tracer kinetic model.

[0097] In a variation, parameters are intelligently tuned and optimal threshold values determined for discriminating disease progression from pseudoprogression in individual lesions. Equivalent classification accuracy can be tested by testing receiver operating curves for each site. Once validated, we will deploy them for multi-cento- clinical trial validation at our three institutions, and thereafter to larger multi-center clinical trials. These thresholds enable the automated, user independent, auto-calibrated DCE-MRI approach to produce predictive percentage maps of pscudo- and true-progression for each patient, for automatic presentation, for example to the neuroradiologist, the neuro-oncologist, the neurosurgeon, the radiation oncologists and anyone else in the treating team for the patient.

[0098] The system and methods described herein allow for improved outcomes for patients with primary brain tumor. High-grade gliomas (HGG) are the most common and aggressive primary brain tumors in adults (24). The evaluation of anti-angiogenic therapies in recurrent HGG is important given the routine use of bevacizumab as well as clinical trials for other agents such as cediranib (25- 29). Bi-directional assessment of conventional contrast-enhanced MR] (i.e., the only validated and FDA-qualified imaging biomarker) is limited as these therapies“normalize" the BBB, resulting in decreased leakage of gadolinium-based contrast agent and contrast enhancement This decrease in enhancement occurs independent of any actual anti-tumor effect and is referred to as “pseudoresponse.” This phenomenon may underly the high radiographic response rate and longer progression free survival without better overall survival (OS) in glioblastoma patients treated with bevacizumab (21 ,30,31). Furthermore, these limitations of contrast-enhancement are also apparent for determination of early response assessment after administration of anti-angiogenic treatments in terms of pseudoresponse (15,32,33). The determination of response has important implications for clinical outcome as shown in Figure 4, and also plays a very significant role in making treatment decisions to continue current therapy, change therapy, or consideration for novel therapies in clinical trials for patients with either definite recurrent disease or known pseudoprogression, as shown in Figure 5.

[0099] In still other variations, the systems and method described herein allow improved outcomes for patients with metastatic brain tumor. Metastatic brain tumors are the most common adult malignant brain tumors overall (44). Treatment with immunotherapy is the front line standard of care for many metastatic brain tumor patients, as supported by national guidelines and FDA-approved prescribing indications. Immunotherapy treatment of brain metastasis in non-immunosuppressed asymptomatic patients results in survival similar to those patients without brain metastatic disease. However, assessing response in patients treated with immunotherapy remains very challenging (45). Traditional measures focusing on size reduction of contrast-enhancing lesions on imaging are unreliable since there is now a plethora of data showing that individual lesions may increase in size yet show dramatic responses at a later date. This is analogous to pseudoprogression in HGG where some melanoma metastases will increase in the volume of enhancement (whereas some within the same patient will decrease in volume) but on further fol low up, all the lesions demonstrate a favorable decrease in volume. In addition, stable disease as a response end point, has proved to be valuable since it is associated with long-term survival, even in the absence of tumor size regression. Several recent clinical studies have shown the potential of DCE-MRI derived parameters including to

differentiate pseudoprogression from true progression in HGG (46-48) and we have shown preliminary evidence that this also applies to metastatic lesions (49).

[0100] In still another variation, the systems and method described herein can be used to evaluate response in immunotherapy. The Immune Response Criteria (IRC) is one such attempt to more accurately detect responders (50). These have arisen because of the inability of imaging modalities, currently focused mostly on changes in tumor contrast enhancement/size, to accurately segregate responders from non-responders. This is important given the potential development of high- grade toxicity, economic impact of continued therapy, and lost opportunity for follow-up therapy. The etiology for the morphologic changes in melanoma lesions treated with immunotherapy is also unknown. Hypotheses include infiltration of antitumor immune cells and increased vascular permeability secondary to immune-related cytokine effects. Quantitative DCE-MRI may help in not only assessment of treatment response but also potentially help in determining the mechanisms of action for immunotherapy.

[0101] Embodiments of the present invention provide a unique opportunity to do the same for

DCE-MRt in the therapeutic triage of brain tumor, by implementing a standardized, automated, and reproducible method for DCE-MRI and by facilitating its translation to a commercial product. Variations of the invention provide an automated DCE-MRI processing pipeline which will produce DCE-MRt maps automatically on a zero-client server and then electronically transmit (email) the results to the end user, clinical neuroradiologists, oncologist, neurosurgeon with negligible human interaction. Resulting maps can demonstrate the area of tumor that fulfills the threshold for true progression vs pseudoprogression based on thresholds determined from ROC analysis of data from the three sites; and could include probability maps, showing the likelihood of true progression in individual voxels. Summary statistics can, among other things, include the volume of tumor (cc) exceeding the threshold. Advantageously, the invention can provide a wealth of standardized multicenter data, and lay the groundwork for development of a“pseudo-progression score”.

[0102] DCE-MRI calibration scans for whole brain radiofrequency transmit inhomogeneity

B1+ and pre-contrast Ti mapping are integrated into the brain DCE-MRI protocol. The spatial resolution and coverage of the DCE-MRI acquisition is then matched. In the quantitative DCE-MRI pipeline, conversion of signal intensity to contrast agent concentration requires precise knowledge of the B1+ and pre-contrast Tt maps. In addition, Tj mapping in itself relies on the knowledge of B1+ map. Literature has documented that errors or inaccurate assumptions in either of these maps will result in substantive errors in the estimated tracer-kinetic parameter maps. For example, a 20% lower flip angle will result in an erroneously 20% lower vp estimate and 20% lower K*™⁸ estimate, assuming that the same arterial input function (AIF) is used. In 3T brain imaging, B1+ spatial variations are on the order of 30-50% (55,56). Several in-vivo DCE-MRI studies of the breast, abdomen, prostate, and brain have documented that correction of B1+ variations provides substantive reduction in the estimation uncertainty of the pre-contrast T i maps (57-62), and tracer kinetic parameter maps; as well as improves test-retest reproducibility of Ti mapping (63). This improvement is extremely critical for clinical trials where the biomarker coefficient-of-variation is a major determinant of the required sample size and the cost of the trial. The objective of this aim is therefore to integrate a robust whole- brain B1+ and Ti mapping protocol and quality assurance procedure for the proposed quantitative brain DCE-MRI.

[0103] In a refinement, whole-brain high-resolution Ti/ Mo mapping can be performed of the system of Figure 1 as follows. In this regard, imaging system 10 is utilizes the 3D spoiled gradient echo based variable flip angle (VFA) method as this has been shown to be the most time efficient method to produce high resolution Ti/Mo mapping of the whole-brain (57,65). The resolution of T i/Mo mapping is matched to the high-resolution DCE-MRI acquisition since the Ti/Mo maps contain sharp features distinguishing different tissues as evidenced by the contrast on any standard Ti -weighted brain scan (e.g., MP-RAGE). Measurements from a predetermined number (e.g., 3, 4, 5, 6, 7, 8, 9, 10) of flip angles can be used to ensure well-posedness of the estimation problem. A few candidate choices of flip angles between 2 to 20 degrees are then evaluated via noise based Monte Carlo simulations in estimating varied Ti values in the range of normal human brain tissue and brain tumors. The choice of using large number of flip angle measurements is motivated by studies that have reported substantial estimation errors in DCE-parameters by using fewer flip angles (e.g., use of two flip angles as opposed to three resulted in estimation errors of 32% and 16% respectively in the

and vp maps ofprimary tumors in the head and neck) (66). Approximate scan time will be 2.5 minutes.

[0104] Another aspect of the system of Figure 1 is the implementation of quality control procedures. In this regard, experiments using multiple (e.g., 2, 3, 4, etc.) human adult head-sized spherical phantoms are performed. The B 1+ phantom, generated as part of the QIBA Groundwork project, will have a single interior tillable volume. This phantom is filled with a solution that has a conductivity (c.g., electrical conductivity) matched to that of human tissue (e.g., 30 mM NaCl). For quality control of Ti/Mo mapping, the quantitative ISMRM system phantom (Figure 6) (High Precision Devices Inc., USA) is used. This phantom includes a plurality of objects (e.g., spheres) filled with solutions mimicking different brain tissues. Fee· example, such a phantom can include 14 spheres Ti range (20 msec, to 2 sec), 14 spheres T2 range (8 msec -800 msec.), and 14 proton density spheres. As part of the quality control process, B1+ mapping, and Ti/Mo mapping sequences on both these phantoms are performed at distinct time separated by a predetermined time period, e.g„ on two separate occasions separated by roughly 2 months. In a further refinement, gold-standard B1+, Mo/ Ti maps are established from slower sequences, respectively from the Double Angle Method (67,68), and spin echo based inversion recovery. In a further refinement, B1+, Mo/Ti mapping from the phantoms are performed at a plurality of MRI facilities to determine the amount and spatial pattern of B1+ variation for all scanners and RF transmit geometries. Any (unexpected) variation in the B1+ pattern due to tesl-retest at individual sites and across sites are noted. For each MRI system utilizing the methods of the invention last calibration methods are compared to slow gold-standard methods.

[0105] Another aspect of the system is Automated Data Processing for DCE-MRI. in this regard, the Olea Software Development Kit (OSDK) architecture can be utilized. Reconstruction plugins can be implemented in OSDK. New features can also be integrated into the thin client DCE-MRI processing solution that have been previously validated in the literature and set forth herein. For example, image calibration including B1+ mapping and Ti/Mo mapping will be integrated. Automatic A1F identification can be integrated using principal component analysis applied to concentration time curves. Similarly, nested TK model selection (plasma only, Patlak, and extended Tofts) so that a minimal model is used for each voxel can also be integrated. Several ROI-based metrics, and automated detection of changes for longitudinal comparisons can also be included. The integration of the new features and new algorithms within the Olea Sphere platform provides an application dedicated to automatic, thin client, post-processing, analysis, follow-up and automatic reporting.

[0106] In another variation, post processing of calibration data generated by the methods set forth above are performed using Bayesian method to accurately estimate B1+ map, pre-contrast Ti map, and proton density (Mo) maps. This information is used to accurately convert signal into concentration lime curves.

[0107] In another variation, automatic patient-specific A.1F identification is performed using principal component analysis applied to concentration-time curves (69). To improve the robustness and reproducibility of the method, anatomical information and/or cluster-based analysis (70) can be added. The value of phase-based and complex-signal-based AIFs for better detection of the peak (71,72) can also be determined. [0108] In another variation, quantitative DCE-MR1 metrics are integrated into a novel plug- in. Multi-compartment models with increased complexity are implemented. Based on the Bayesian Information Criterion (BIC) (73), a minimal model can be selected for each voxel. In a refinement, once integrated, the plug-in is optimized in terms of computation time and memory requirements to allow response within the constraints of clinical use. A validation of the quantitative algorithms can be performed using anatomically-realistic digital reference objects phantoms developed by the academic sites (74,75), i.e. in silico data for which the ground truth value is known and can be directly compared to estimated values. Following the integration of algorithms, adjustments are performed to have a post-processing solution tailored to neuro-oncology for analysis and follow-up. Therefore, modification of the existing multi-parametric plug-ins is done with integration of co-registration and specific tools and reports that facilitate longitudinal analysis. These tools (including registration) have been thoroughly tested with acid tests and have performed very well.

[0109] In some variation, the performance and stability of the solution set forth above are audited. The software will be analyzed following a clinical logical process to verify its consistency and the calculation time. In parallel to this performance and safety study, the qualification team of Olea Medical will work on a click-by-cliclc testing of features and functionalities. The objective is to verify the software resilience.

[0110] These algorithms set forth herein can be included into a new Olea Sphere 3.0 module using the Olea Software Development Kit (OSDK). Olea Medical has recently developed OSDK in order to drastically simplify the development of new applications and/or plug-ins. OSDK is a "lego" box, containing all developed libraries and designed to easily integrate new ones. From a practical point of view, OSDK allows one to build a new application or plug-in quickly. OSDK comes with a user-friendly interface allowing basically any user to integrate a new algorithm or library and to build his or her own new module, directly within Olea Sphere environment and interface, regardless of level of programming skills (see schematic description, Figure 7).

[0111] In another variation, the ability to standardize quantitative brain DCE-MRI in a multicenter setting is evaluated as follows. A physical phantom is used evaluate the outcome of standardization. However, anatomically and physiologically realistic brain tumor DCE-MRI phantoms do not exist. Note that realistic in silico digital reference objects are available (75) and have been used to characterize several of the component technologies in this proposal. As an alternative approach, we will test the equivalence in terms of accuracy when predicting progression from pseudoprogression across different clinical sites. The standardization of clinical classification of progression and pseudoprogression can be achieved and validated using inter-rater agreement as the conventional approach. If the classification accuracy (AUC from the ROC curve) are equivalent across sites, we can claim the DCE-MRI procedure were standardized. The information for adjudicating progression and pseudoprogression will be entered into our database. This web -based database will allow rater to rate the progression vs pseudo-progression using a web interface. The rating result will be blinded from raters. We will use the first 20 patients from each of our three clinical sites to validate the standardization procedure of clinical classification.

[0112] In another variation, a diagnostic cut point in brain DCE-MRI for discriminating disease progression from pseudoprogression can be identified by collecting and pooling MRI data from multiple sites. A prediction model is then derived using supervised machine learning. Multivariate Adaptive Regression Splines (MARS) is used to select important predictors. The final predictors are entered into a logistic regression model to derive the predicted probability. The prediction accuracy is evaluated using ROCs based on the predicted probability generated by the multi-variate logistic regression model.

[0113] In a refinement, the threshold is applied to an automated, user independent analysis, auto-calibrated DISCO DCE-MRI approach to produce maps with the predictive percentage for PP vs TP for that patient, and electronically transmit (emailed) them to the neuroradiologist. Once confirmed, we would then deploy this in multi-center clinical trials at our 3 institutions, and thereafter to larger multi-center clinical trials. In essence, we will provide the end user with a validated automated thin/zero client black box approach that provides pixel -based biomarker visualization. Operator- interaction will only be needed to define regions-of-interest around individual lesions for lesion-based biomarker evaluation (including histograms).

[0114] In a variation, patients are imaged using an institutional standard-of-care brain tumor protocol that follows the recent consensus recommendation by neuro-oncologists and brain tumor imaging experts (76). For example, DISCO DCE-MRI can be implemented with a 3D spoiled gradient echo pulse sequence covering the entire brain with a coronal or sagittal orientation. The calibration set forth above is incorporated into this scan. In a refinement, the frequency-encoding direction is superior-inferior. In a refinement, a flip angle of 25^* to 30^* and a repetition time of approximately 4.5 ms (e.g., 3.5 to 5.5 ms) (minimum TR) can be utilized. The acquisition matrix will be 256x256x128 or higher, corresponding to a spatial resolution of 0.95x0.95x1.9 mm³ or better. Temporal resolution will be set to 3.5 seconds. Gadobulrol (Gadavist, Bayer Healthcare) or Gadobenate dimeghunine (Multihance, Bracco Diagnostics) (0.1 mmol/kg) will be administered into the left antecubital vein using a power injector (AC1ST EmpowerMR Injector, Bracco), at a rate of 3 ml/sec, followed by a 20- ml saline flush. For standardization, we are acquiring data from 3T but the product will ultimately be vendor and field strength agnostic.

(OllSj For prospective clinical evaluation, data collected from multiple sites that has been standardized and analyzed as set forth above is utilized. DCE MRI metrics such as

can be plotted on a receiver operating curve allowing determination of the most optimal DCE MRI metrics and thresholds that provide optimal predictive value, sensitivity and specificity for making diagnoses of pseudoprogression. These threshold values are then be implemented a priori into an automated thin client server (Olea implementation of the Olea Sphere 3.0 platform). Data is sent from the magnetic resonance system to the server which will then apply the processing algorithm set forth above. The multiple clinical sites can then test this automated standardized software in the clinical setting.

[0116] Sample size and power analysis is evaluated as follows. Inter-rater (from a group of experts) agreement for clinically adjudicated pseudoprogression vs. true progression is validated using Kappa coefficient. After clinical rating standardized, it is used as the anchor to validate the standardization of DCE-MRI. Assuming the patient population is similar across the multiple clinical sites, the association between DCE-MRI and clinical rating should be the same. The equivalency of discrimination power (the AUC from ROC curve) can be tested across the multiple clinical sites when using DCE-MRI to predict clinically adjudicated pseudoprogression vs. true progression. The procedure introduced by Liu etal (77) of testing the equivalency AUC across the multiple sites. With pre-specified clinically meaningful limitsd=0.05, the evaluating equivalence is to perform the two onesided tests (TOST) based on the difference in paired areas under ROC curves. The bootstrap technique can then be used to empirically obtain the sampling distributions of test statistics. However, the approach of Liu el al is based on paired ROC curve (two tests apply to the sample patient). Since our patients arc from different clinical sites, it is necessary to estimate the correlation of DCE-MRI if we have patients repealed the measurement across multiple sites. According to Liu et al, a higher the correlation will associate with the stronger power. The strongest correlation can be approximated by pairing patient according to the rank of DCE-MRI result (highest in site 1 pair with the highest in site 2) within clinical rating category. The bootstrap method for the pairs with highest correlation is conducted using 2000 samples. The lowest correlation can be approximated by a permutation method (randomly reassign the matching pairs) within the clinical rating category. Such permutation method will be combined with the original bootstrap method, as the result, first 5000 samples will be generated in estimating the empirical confidence interval. Then additional sample will be generated at 1000 incrimination until the empirical confidence interval becoming static. By using this approximation approach, the highest equivalence and lowest equivalence of DCE-MRI in predicting pseudoprogression vs. true progression can be estimated. If we obtain a poor equivalence (beyond the limit of ±0.1) under highest correlation scenario, we will discourage further study. However, if a good equivalence (within the limit of ±0.05) under lowest correlation is achieved, it can be confidently concluded that a DCE-MRI protocol across different clinical sites is established.

[0117] The sample size for standardizing clinical rating can be estimated based on Kappa coefficient For example, a plurality of common subjects (e.g., 25 subjects) can be used for the standardizing clinical rating across the multiple sites, since the clinical data can put posted in a secure research database and shared with different sites. The gold standard rating result will need to be established by an expert panel. The study rater will need to rate al l cases and compare to the result of gold standard. With 25 or more subjects, we can derive a half-width of 95% confidence interval of Kappa=0.15 when the observed agreement is Kappa=0.85, so that we will be confident to conclude the true Kappa agreement is at least 0.7.

[0118] The sample size estimation for the AUC equivalence test is based on the simulation study result reported by Liu et.al. Administratively we can recruit 150 patients per site. Since it is impractical for 150 patients to have repeated measurement across three sites, we will estimate the highest and the lowest power using the following assumption: 1) The AUC for the three site is around 0.75 but will vary by 0.05. in other word, if the variation is within 0.05 we will claim equivalent. 2) The highest correlation coefficient of DCE measurement assuming repealed measurement across sites can reach 0.9, and the lowest correlation coefficient can be 0.5. With the highest correlation, we will have 80% power, while the lowest correlation will only result in a power of 29% in testing the equivalency of prediction accuracy of DCE-MRI in predicting pseudoprogression vs. true progression.

[0119] In another variation, multivariate adaptive regression splines (MARS) are used to select important predictors. The final predictors can be entered into a logistic regression model to derive the predicted probability. The prediction accuracy can be evaluated using ROCs based on the predicted probability generated by the multi-variate logistic regression model. The discrimination power of DCE-MRI in predicting pseudoprogression vs. true progression can be derived using the area under the curve (AUC). By using the combined patients (e.g., 450 patients) from multiple sites (e.g., 3 sites), it is possible achieve 90% power to detect a difference of 0.08 between die area under the ROC curve (AUC) under the null hypothesis of 0.7 and an AUC under the alternative hypothesis of 0.78 using a one-sided z-test at a significance level of 0.05. This sample size estimation is based on a conservative assumption of only 30% patient with true progression. PASS 14 was «toed for all power estimations.

[0120] The following examples illustrate the various embodiments of the present invention.

Those skilled in the art will recognize many variations that are within the spirit of the present invention and scope of the claims.

[0121] 1. SPARSE DCE-MRI USING A TEMPORAL CONSTRAINT LEARNED

FROM CLINICAL DATA EXPERIMENTS

[0122] The datasets used in this study were from brain tumor patients receiving a routine brain

MRI with contrast on a clinical 3 Tesla scanner at our institution. A 3D cartesian fast SPGR sequence with field of view (FOV): 22x22*4.2cm3; spatial resolution: 0.9x1.3x7.0mm3; temporal resolution: 5 seconds; 50-time frames; 8 receiver coils; 15° flip angle; 1.3 ms echo time; and 6 ms TR was used. A dictionary of 800 elements was trained using 768000 patches from 20 fully-sampled datasets. This dictionary was then used to reconstruct 15 test datasets (retrospectively down-sampled).

[0123] 1.1 Results and Discussion [0124] Reconstruction time was: GLR ~0.55 minutes for 50 iterations; MOCCO ~5.33 minutes; DL ~23.74 minutes for training and ~7.15 minutes for reconstruction with 20 iterations. The major computational burden is the sparse coding step, which takes ~ 6.67 minutes for 20 iterations.

[0125] Figure 8 contains tumor

ROl maps for 5 representative tumors (out of 15) for an under-sampling rate of 40 in the testing dataset. Figure 9 shows matching tumor vp ROI maps. TK maps reconstructed from GLR appear spatially smooth and those from MOCCO show a higher noise level.

[0126] Figure 10 compares the performance as a function of under-sampling. Each column corresponds to the tumors shown in Figs. 8 and 9. DL consistently provided lower NRMSE compared to the alternatives that were studied. Figure 11 shows the average KMSE for all 15 tumors in the test dataset. The improvement of DL with respect to GLR is significant, whereas the improvement with respect to MOCCO is minimal.

[0127] We also explored the use of spatio-temporal dictionaries (not shown here), and these provided significant denoising. This warrants further exploration along with a comparison to previous spatio-temporal constraints.

[0128] 1.2 Conclusion

[0129] Sparse DCE-MRI reconstruction using a learned temporal constraint from clinical data may outperform current temporally constrained methods. The proposed technique was evaluated on 15 brain tumor cases and provided superior performance in every case.

]0130j 2. SPARSE PRE-CONTRAST T1 MAPPING FOR DCE-MRI

CALIBRATION EXPERIMENTS

[0131] Data were acquired MI a glioblastoma subject using a GE MR750 scanner with a 12- channel head coil The vendee- provided 3D spoiled gradient echo sequence was modified to include sparse undersampling. Six increasing flip angles each with 5000 TRs and a final flip angle with 20000 TRs were performed. Note that the last flip angle has a longer duration since it is used in the DCE scan and precedes the contrast injection. The flip angle increased from 1.5° to 15° logarithmically2. Other settings: 5 ms TR, 1.9 ms TE, 240x240x240 mm3 FOV, 2 mm slice thickness, and 256x240x120 matrix size.

[0132] Pre-contrast Ti/Mo mapping is performed by solving the constrained inverse problem set forth above in equations 3 and 4: We used wavelet transform as the spatial sparsifier as it preserves subtle features as well as denoises data, l was empirically chosen as 0.3 and remained constant in all experiments. To avoid model failure impacting the estimation process, we impose both non-negative and non-infinite constraints on each pixel's Ti and Mo values. We explored Ti mapping accuracy as a function of the number of TR periods included at each flip angle. Faster acquisitions were simulated by discarding samples at each flip angle, as illustrated in Figure 12. We measure and report the error in Ti and Mo maps as functions of the amount of retained data.

[0133] 2.1 Results

[0134] Figure 13 shows the ROI outline and difference maps for Ti, Mo, and anatomic images within the ROI. These maps were obtained by computing differences between results of (5000, 20000) setting and results of any (A, B) setting. Note that Mo difference decreases as B decreases when A is small. In addition, there was at most 85 ms Ti difference per pixel, and the differences in quantity of magnetization and image energy were 1.873% and 3.290% at most.

[0135] Figure 14 shows means and standard deviations of estimated T i in normal white matter. Ti values were selected from the ROI in Figure 13. All results show similar mean Ti. Interestingly, increasing A or decreasing B both result in a subtle decrease in standard deviation of Ti.

[0136] Figure 15 illustrates a kk-space error curve, a ROI Ti absolute change curve and a ROI image increment curve. Both Ti change and image increment became trivial after 80 iterations, while residuals in kk-space were quite significant.

[0137] 23 Discussion

[0138] This approach enables adequate pre-contrast Ti estimation matched to the resolution and coverage of modem high-resolution whole-brain DCE-MRI in a reasonable scan time (<3 minutes), in stark contrast to a fully sampled VFA acquisi tion that would require more than 20 minutes. Despite different subsample settings, both Ti values mean and standard deviation were stable, and no substantial bias was observed. This indicates that the proposed approach is flexible and allows a trade- offbetween data acquisition lime and Ti statistics.

[0139] Another important observation is significant kk-space residuals. One possible cause is that wavelet sparsity regularization can over-smooth images. One way to compensate for these residuals is to impose TGV constraints on vascular parameters directly¹, or to enforce data consistency after each iteration (used in this work).

]014Q] 2.4 Conclusion

[0141] Sparse pre-contrast Ti mapping matched to high-resolution whole-brain DCE-MRI acquisition is feasible. There is a trade-offbetween data acquisition time and Tiaccuracy.

[0142] 3. POSTERIOR APPROXIMATION FOR SIMULTANEOUS DCE-MRI

PHARMACO-KINETIC PARAMETER AND UNCERTAINTY ESTIMATION

EXPERIMENTAL

[0143] 3.1 Results

[0144] As shown in Figure 16, the proposed method is able to estimate PK parameter maps that are comparable to conventional model fitting with multiple initializations. Scatter plots of the standard deviations of K^l predicted by the proposed method ami MCS of conventional methods show good correlation between the two predictions, yet no absolute agreement. Figure 17 shows concentration lime curves, and contour plots of the data consistency cost function for various projections into the PK parameter plane. The true value (red) is well contained inside the ellipsoid (blue cloud) and samples from the ellipsoid (blue cloud) result in very similar concentration time curves (blue shaded area).

[0145] 3.2 Discussion

[0146] The results indicate that the proposed method can simultaneously predict PK parameters in tumor tissue as well as provide metrics for uncertainty of the estimation due to noise as well as the general structure of the cost function landscape. Since minimum variance unbiased estimation is impossible for non-linear models in the Gaussian noise case a posterior distribution was chosen that does not give preference to a single value (like the mode). Convex posterior distributions like the uniform ellipsoid break down at low values of K* when ve is impossible to estimate and the cost function is non- convex. Possible solutions are the use of two super-imposed posterior distributions, or sparsity constraints on K* that enforce the posterior to expand into the desired trench of the cost function landscape.

[014h 33 Conclusion.

[0148] Neural networks can be used to simultaneously provide PK parameter estimates and measures of uncertainty without relying on sampling methods, linearizations, or ground truth labels during training.

[0149] 4. DEEP CONVOLUTIONAL RESIDUAL NETWORK FOR DENOISING

BRAIN TUMOR DCE-MRI TRACER-KINETIC MAPS

[0150] 4.1 Methods

[0151] Network: A DCRN was designed and trained to denoise TK. maps (K¹™* and Vp) directly estimated from under-sampled (k, l) space data. The DCRN architecture was 7 layers with each layer containing a convolutional layer (Conv) followed by batch normalization (BN ) and rectified linear unit (RcLU) activation (Conv+BN+ReLU), except the last layer which was a convolutional layer only. The network employs multilevel decomposition with the filter sizes of the middle layers larger than the early and late layers and the number of filters in each layer being 128. Training the network with the residue mitigates the problem of vanishing gradient. Utilization of BN layer between Conv and ReLU makes the network less susceptible to the learning rate and careful about the initialization. Training: Inputs were patches of TK maps from under-sampled data and outputs were the corresponding residue between maps from under-sampled and fully-sampled data. Mean squared error was considered as a loss function and stochastic gradient descent with momentum of 0.9 and learning rate of le-5 was utilized to minimize the loss function. Clinical Data: This study utilized raw data from 80 brain tumor DCE-MRI datasets from our institution for training, from which 105600 overlapping patches (patch size of 80u80 with stride 10) were extracted. An additional 17 datasets were used for testing. The DCRN was applied to TK-maps from under- sampled data with rales varying between 20 to 200 using direct- reconstruction. Data Analysis: TK maps were de-noised using the proposed network, and also with non-local means (NLM) and block matching three-dimensional (BM3D) denoising algorithms. All tuning parameters were empirically chosen. Normalized root mean squared error (NRMSE) was computed relative to reference TK maps, NVIDIA K40 GPU was used for training. Total training time was approximately 6 hours for 50 iterations.

[0152] 4.2 Results & Discussion

[0153] Figure 18 contains results from 6 (of 17) representative tumors in the testing dataset The most valued TK map features are preserved, including the depicti on of the narrow enhancing rim, whereas noise level is substantially reduced. The lack of anatomic features in the error plots suggest that the DCRN is not introducing bias. Figure 19 summarizes denoising performance as a function of under-sampling factor and compares results with BM3D and NLM. Across all under-sampling rates, the DCRN produced lower NRMSE, with the largest improvement in Vp maps. It remains future work to apply more sophisticated tumor-specific evaluation criteria including histogram and/or texture analysis.

[0154] 4.3 Conclusion

[0155] ML-based denoising is able to improve brain tumor DCE-MRI TK maps by at least

1.5-fold. This could be an effective complement to sparse DCE-MRI methods that suffer from increased TK map variance at high under-sampling factors.

[0156] 5. INFLUENCE OF WHOLE-BRAIN DCE-MRI (K,T) SAMPLING

STRATEGIES ON VARIANCE OF PH ARM ACO-KI NETIC PARAMETER ESTIMATES

EXPERIMENTS

[0157] We investigate the influence of 2D (ky,kz,t) sampling strategies on the minimum achievable variance without bias for pharmaco-kinetic parameter estimation in 3D whole-brain DCE- MRI (equivalent to the best possible precision without bias). Cramer-Rao analysis is combined with a pathologically- and anatomically-realistic digital reference object to objectively compare measurement procedures independent of any estimator. This study did not identify any significant difference between lattice and random undersampling, or between their uniform and variable density variants.

[0158] Dynamic contrast enhanced MRI (DCE-MRI) of the brain provides a powerful tool to non-invasively assess neurovascular parameters, such as permeability of the brain-blood barrier (K^l), plasma (vp), and interstitial volume (ve). DCE-MRI has potential to monitor early treatment response in brain tumor therapy. Yet in order to make reliable predications the entire measurement and estimation procedure has to become accurate, precise, and reproducible 1. In this work, we focus on one specific aspect of this measurement chain by objectively comparing (k,t) sampling patterns in a framework that works independent of any specific estimator. Undersampling in each time frame is inevitable to achieve desired temporal resolution to reliably capture bolus arrival, accumulation, and wash-out, while still maintaining a clinically relevant FOV and high spatial resolution 2. This raises the question which under sampling strategy allows for estimation of pharmaco-kinetic (PK) parameter with lowest variance.

[0159] 5.1 Methods

[0160] 5.1.1 Framework

[0161] The Cramer-Rao bound (CRB) gives a lower bound on variances of any unbiased estimator, and has been widely used to optimize MRI experiment design 3-5. Evaluation of the CRB requires the derivatives of the DCE-MRI forward model with respect to the parameters being estimated. The forward model was simulated by PK modeling based on the Patlak model, an SPGR sequence, sensitivity encoding, and Fourier undersampling. A population based arterial input function (AIF) was chosen for this simulation 6. Coil sensitivities and noise covariance matrix for an eight- channel head array were taken from measurements. The derivative of the forward model needs to be evaluated at the parameter being estimated. Hence, a pathologically- and anatomically-realistic digital reference object (DRO) was taken to be the ground-truth (Figure 20 ) 7. CRBs were computed for pre- contrast white matter SNR=10, flip angle of 24°, and 50 time frames at 5s temporal resolution. As a linear system model allows the existence of the minimum variance unbiased estimator in the Gaussian noise case, pharmaco-kinetic model and flip angle were chosen to operate the signal equation in the linear regime 5.

[0162] 5.1.2 Sampling schemes

[0163] Cartesian sampling schemes can broadly be classified into k-space region sampling

(e.g., Keyhole, TRICKS), lattice-based, and random sampling with uniform or variable density variants (e.g., kt-SENSE, DISCO, TWIST, kl-SPARSE). Early in this study, we found that k-space region sampling schemes often lead to very ill-conditioned or even singular Fisher Information matrices, which indicate that certain PK parameters cannot be estimated with finite variance 8. Hence, zone-based sampling is omitted a-priori from the comparison of candidate sampling patterns, shown in Figure 21. The undersampling factors for each sampling schemes shown in Figure 21 were R=2,5,10,15.

[0164] 5.2 Results

[0165] Figure 22 illustrates the best achievable precision for the PK parameters vp and K^l as predicted by the CRB for the whole DRO, while Figure 24 contains the bounds for the tumor ROI only. Monte-Carlo simulations using direct reconstruction of PK parameters 9 have previously confirmed lightness of the bound (not shown). We observed no difference in the minimum achievable variance for pharmaco-kinetic parameter estimation, between lattice and random sampling, nor between uniform and variable density variants. Pairwise differences in vp standard deviation were all at least 3-fold lower than the tumor average vp value, and 50-fold lower for the clinically relevant K^l parameter.

[0166] 5J Discussion

[0167] In this work, we used Cramer-Rao analysis to objectively compare lattice-based sampling with random sampling both in their uniform and variable density variants. Extending on the results in a previous study 3 which analyzed common ID undersampling strategies, the sampling patterns in this study had no discernible impact on the minimum achievable variance for PK parameter estimation. Note that the actual performance of a DCE measurement also depends on the availability of a numerically stable post processing procedure to estimate PK parameters with the predicted variance.

[0168] An important limitation of this study is that we assumed perfect knowledge of the AIF, bolus arrival time, and coil sensitivities, and no motion. During a clinical scan these assumptions do not hold. Sampling pattern may indirectly influence PK measurement precision through the ability to more precisely estimate AIF, bolus arrival time, and coil sensitivities, and to correct for motion artifacts.

[0169] While exemplary embodiments are described above, it is not intended that these embodiments describe all possible forms of the invention. Rather, the words used in the specification are words of description rather than limitation, and it is understood that various changes may be made without departing from the spirit and scope of the invention. Additionally, the features of various implementing embodiments may be combined to form further embodiments of the invention.

[0170] REFERENCES CITED

[0171] 1. Wouters A, Lemmens R, Christensen S, Wilms G, Dupont P, Mlynash M, et al.

Magnetic resonance imaging-based endovascular versus medical stroke treatment for symptom onset up to 12 h. bit J Stroke. 2016 Jan;l 1(1): 127-33.

[0172] 2. Lansbeig MG, Cereda CW, Mlynash M, Mishra NK, Inoue M, Kemp S, et al.

Response to endovascular reperfusion is not time-dependent in patients with salvageable tissue. Neurology. 2015 Aug 25;85(8):708-14.

[0173] 3. Mishra NK, Albers GW, Christensen S, Marks M, Hamilton S, Straka M, el al.

Comparison of magnetic resonance imaging mismatch criteria to select patients for endovascular stroke therapy. Stroke. 2014 May;45(5): 1369-74.

[0174] 4. Olivet JM, Mlynash M, Inoue M, Marks MP, Wheeler HM, Kemp S, et al.

Hypoperfusion intensity ratio predicts infarct progression and functional outcome in the DEFUSE 2 Cohort. Stroke. 2014 Apr, 45(4): 1018-23. [0175] 5. Liggins JTP, Yoo AJ, MishraNK, Wheeler HM, Straka M, Leslie-Mazwi TM, et al. A score based on age and DW1 volume predicts poor outcome following endovascular treatment for acute ischemic stroke. Int J Stroke. 2015 Jul; 10(5): 705-9.

[0176] 6. Inoue M, Mlynash M, Straka M, Kemp S, Jovin TG, Tipimeni A, et al. Clinical outcomes strongly associated with the degree of reperfusion achieved in target mismatch patients: pooled data from the Diffusion and Perfusion Imaging Evaluation for Understanding Stroke Evolution studies. Stroke. 2013 Jul;44(7): 1885-90.

[0177] 7. Wheeler HM, Mlynash M, Inoue M, Tipimeni A, Liggins J, Zaharchuk G, et al.

Early diffusion-weighted imaging and perfiision-weightcd imaging lesion volumes forecast final infarct size in DEFUSE 2. Stroke. 20.13 Mar;44(3):681-5.

[0178] 8. Lansberg MG, Straka M, Kemp S, Mlynash M, Wechsler LR, Jovin TG, et al.

MRI profile and response to endovascular reperfusion after stroke (DEFUSE 2): a prospective cohort study. Lancet Neurol. 2012 Oct;l l(!0):860-7.

[0179] 9. Wen PY, Cloughesy IF, Ellingson BM, Reardon DA, Fine HA, Abrey L, et al.

Report of the Jumpstarting Brain Tumor Drug Development Coalition and FDA clinical trials neuroimaging endpoint workshop (January 30, 2014, Bethesda MD). Meuro Oncol. 2014;16(July):vii36-vii47.

[0180] 10. Gossmann A, Hclbich TH, Kuriyama N, Ostrowitzki S, Roberts TPL, Shames DM, et al. Dynamic contrast-enhanced magnetic resonance imaging as a surrogate marker of tumor response to anti-angiogenic therapy in a xenograft model of glioblastoma multiforme. J Magn Reson Imaging. 2002;15:233-40.

[0181] 11. Jalali S, Chung C, Foltz W, Burrell K, Singh S, Hill R, et al. MRI biomarkers identify the differential response of glioblastoma multiforme to anti-angiogenic therapy. Ncuro Oncol. 2014; 16( April):868-79. [0182] 12. Kalpaihy-Cramer J, Gerstner ER, Emblem KE, Andronesi OC, Rosen B.

Advanced magnetic resonance imaging of the physical processes in human Glioblastoma. Cancer Res. 2014;74(17):4622-37.

[0183] 13. Hcye T, Davenport MS, Horvath JJ, Feuerlein S, Brcault SR, Bashir MR, et al.

Reproducibility of Dynamic Contrast-Enhanced MR Imaging Part 1 . Perfusion Characteristics in the

Female Pelvis by Using Multiple Computer-aided Diagnosis Perfiision Analysis Solutions. 2013;266(3):801~11.

[0184] 14. Hcye T, Meride EM, Reiner CS, Davenport MS, Horvath JJ, Feuerlein S, et al.

Reproducibility of dynamic contrast-enhanced MR imaging. Part P. Comparison of intra- and interobserver variability with manual region of interest placement versus semiautomatic lesion segmentation and histogram analysis. Radiology. 2013;266(3):812-21.

[0185] 15. Wen PY, Macdonald DR, Reardon DA, Cloughesy TF, Sorensen AG, Galanis

E, et al. Updated response assessment criteria for high-grade gliomas: response assessment in neuro- oncology working group. J Clin Oncol. 2010 Apr 10;28(11):1963-72.

[0186] 16. Okada H, Weller M, Huang R, Finocchiaro G, Gilbert MR, Wick W, et al.

Immunotherapy response assessment in neuro-oncology: a report of the RANG working group. Lancet Oncol. Elsevier Ltd; 2015;16(15):e534-42.

[0187] 17. Hcgi ME, Diserens A-C, Goriia T, Hamou M-F, dc Tribolct N, Weller M, et al. MCMT Gene Silencing and Benefit from Temozolomide in Glioblastoma. N Engl J Med. Massachusetts Medical Society ; 2005 Mar 10;352(10):997-1003.

[0188] 18. Esteller M, Garcia-Foncillas J, Andion E, Goodman SN, Hidalgo OF,

Vanaclocha V, et al. Inactivation of the DNA-Repair Gene MGMT and the Clinical Response of Gliomas to Alkylating Agents. N Engl J Med. Massachusetts Medical Society ; 2000 Nov 9;343(19):1350-4.

[0189] 19. Hegi ME, Diserens A-C, Godard S, Dietrich P-Y, Regli L, Ostermann S, et al.

Clinical Trial Substantiates the Predictive Value of O-6-Methylguanine-DNA Methyltransferase Promoter Methylation in Glioblastoma Patients Treated with Temozolomide. Clin Cancer Res. 2004;! 0(6).

[0190] 20. Brandsma D, van den Beni MJ. Pseudoprogression and pseudo response in the treatment of gliomas. Curr Op in Neurol. 2009 Dec>22(6):633-8.

[0191] 21. Brandes AA, Franceschi E, Tosoni A, Blalt V, Pession A, Tallini G, et al. MCMT promoter methylation status can predict the incidence and outcome of pseudoprogression after concomitant radiochemotherapy in newly diagnosed glioblastoma patients. J Clin Oncol. American Society of Clinical Oncology, 2008 May l;26(13):2192-7.

[0192] 22. Hegi ME, Liu L, Herman JG, Stupp R, Wick W, Weller M, et al. Correlation of

06-methylguanine methyl transferase (MGMT) promoter methylation with clinical outcomes in glioblastoma and clinical strategies to modulate MGMT activity. J Clin Oncol. American Society of Clinical Oncology; 2008 Sep 1;26(25);4189-99.

[0193] 23. Motegi H, Kamoshima Y, Terasaka S, Kobayashi H, Yamaguchi S, Tanino M, et al. IDH1 mutation as a potential novel biomarker for distinguishing pseudoprogression from true progression in patients with glioblastoma treated with temozolomide and radiotherapy. Brain Tumor Pathol. Springer Japan; 2013 Apr 3;30(2):67-72.

[0194] 24. Telles BA, D’Amore F, Lemer A, Law M, Shiroishi MS. Imaging of the

Postthcrapeutic Brain. Top Magn Rcson Imaging. 2015 Jun;24(3): 147-54.

[0195] 25. Friedman HS, Prados MD, Wen PY, Mikkelsen T, SchifF D, Abrey LE, et al.

Bevacizumab alone and in combination with irinotecan in recurrent glioblastoma. J Clin Oncol. 2009 Oct 1 ;27(28):4733-40.

[0196] 26. Kreisl TN, Kim L, Moore K, Duic P, Royce C, Stroud I, et al. Phase II trial of single-agent bevacizumab followed by bevacizumab phis irinotecan at tumor progression in recurrent glioblastoma. J Clin Oncol. 2009 Feb 10;27(5):740-5. [0197] 27. Zhang G, Huang S, Wang Z. A meta-analysis of bevacizumab alone and in combination with irinotecan in the treatment of patients with recurrent glioblastoma multiforme. J Clin Neurosci. 2012 Dec;19(12): 1636-40.

[0198] 28. Batchelor TT, Muiholland P, Neyns B, Nabors LB, Campone M, Wick A, et al.

Phase HI randomized trial comparing the efficacy of cediranib as monotherapy, and in combination with lomustine, versus lomustine alone in patients with recurrent glioblastoma. J Clin Oncol. 2013 Sep 10;31(26):3212-8.

[0199] 29. Sorensen AG, Batchelor TT, Zhang W-T, Chen P-J, Yeo P, Wang M, et al. A

“vascular normalization index^** as potential mechanistic biomarker to predict survival after a single dose of cediranib in recurrent glioblastoma patients. Cancer Res. 2009 Jui l;69(13):5296-300.

[0200] 30. Young RJ, Gupta A, Shah AD, Graber JJ, Zhang Z, Shi W, et al. Potential utility of con ven tional MRI signs in diagnosing pseudoprogression in glioblastoma. Neurology. Lippincott Williams & Wilkins; 2011 May 31;76(22):1918-24.

[0201] 31. Gilbert MR, Dignam JJ, Armstrong TS, Wefel JS, Blumenthal DT, Vogelbaum

MA, et al. A randomized trial of bevacizumab for newly diagnosed glioblastoma. N Engl J Med. 2014 Feb 20;370(8):699-708.

[0202] 32. Norden AD, Drappatz J, Muzikansky A, David K, Gerard M, McNamara MB, et al. An exploratory survival analysis of anti -angiogenic therapy for recurrent malignant glioma. J Neurooncol. 2009 Apr, 92(2): 149-55.

[0203] 33. Ellingson BM, Cloughesy TF, Lai A, Nghiemphu PL, Mischel PS, Pope WB.

Quantitative volumetric analysis of conventional MRI response in recurrent glioblastoma treated with bevacizumab. Neuro Oncol. 2011 Apr;13(4):401-9.

[0204] 34. Shiroishi MS, Castellazzi G, Boxerman JL, D’Amore F, Essig M, Nguyen IB, et al. Principles of T2 ^-weighted dynamic susceptibility contrast MRI technique in brain tumor imaging. J Magn Reson Imaging. 2014 May 12;(online). [0205] 35. Essig M, Shiroishi MS, Nguyen TB, Saake M, Provenzale JM, Enterline D, et al. Perfusion MRI: the five most frequently asked technical questions. AJR Am J Roentgenol. 2013 Jan;200(l):24-34.

[0206] 36. Artzi M, Liberman G, Nadav G, Vitinshtein F, Blumenlhal DT, Bokstein F, et al. Human cerebral blood volume measurements using dynamic contrast enhancement in comparison to dynamic susceptibility contrast MRL Neuroradiology. 2015;57(7):671-8.

[0207] 37. Zhang W, Kriesl T, Solomon J, Reynolds R, Glen D, Cox R, et al. Acute effects of bevacizumab cm glioblastoma vasculatiry assessed with DCE-MR1 and relation to patient survival. In: 1SMRM 2009. 2009. p. 282.

[0208] 38. Bonekamp D, Deike K, Wiestler B, Wick W, Bendszus M, Radbruch A, et al.

Association of overall survival in patients with newly diagnosed glioblastoma with contrast-enhanced perfusion MRI: Comparison of intraindividually matched T1 - and T2 (*) -based bolus techniques. J Magn Reson Imaging. 2015 JuI;42(i):87-96.

[0209] 39. Law M, Yang S, Babb JS, Knopp EA, GolfinosJG,Zagzag D, etal. Comparison of cerebral blood volume and vascular permeability from dynamic susceptibility contrast-enhanced perfusion MR imaging with glioma grade. AJNR Am J Neuroradiol. 2004 May£5(S):746-55.

[0210] 40. Kickingereder P, Wiestler B, Graf M, Heiland S, Schlemmer HP, Wick W, et al. Evaluation of dynamic contrast-enhanced MRI derived microvascular permeability in recurrent glioblastoma treated with bevacizumab. J Neurooncol. 2015 Jan;l 21(2):373-80.

[0211] 41. Piludu F, Marzi S, Pace A, Vil!ani V, Fabi A, Carapeila CM, et al. Early biomarkers from dynamic contrast-enhanced magnetic resonance imaging to predict the response to antiangtogenic therapy in high-grade gliomas. Neuroradiology. 2015; 1269-80.

[0212] 42. Artzi M, Bokstein F, Blumenlhal DT, Aizenstein O, Liberman G, Com BW, et al. Differentiation between vasogenic-edema versus tumor-infiltrative area in patients with glioblastoma during bevacizumab therapy: a longitudinal MRI study. Eur J Radiol. 2014 Jul;83(7):1250-6. [0213] 43. Artzi M, Blumentha! DT, Bokstein F, Nadav G, Liberman G, Aizenstein O, et al. Classification of tumor area using combined DCE and DSC MRI in patients with glioblastoma. J NeurooncoL 2015 Jan;121(2):349~57.

[0214] 44, Lin NU, Lee EQ, Aoyama H, Barani U, Barboriak DP, Baumert BG, ct al.

Response assessment criteria for brain metastases: proposal from the RANG group. Lancet Oncol. 2015;16(6):e270-8.

[0215] 45. Hoos A, Eggcrmont AMM, Janetzki S, Hodi FS, Ibrahim R, Anderson A, el aL

Improved endpoints for cancer immunotherapy trials. J Natl Cancer Inst. 2010 Sep 22; 102(18): 1388-

97.

[0216] 46. Thomas AA, Arevalo-Perez J, Kaley T, Lyo J, Peck KK, Shi W, et al. Dynamic contrast enhanced T1 MRI perfusion differentiates pseudoprogression from recurrent glioblastoma. J NeurooncoL Springer US; 2015;125(1):183-90.

[0217] 47. Yun TJ, Park CK, Kim TM, Lee SH, Kim JH, Sohn CH, et al. Glioblastoma treated with concurrent radiation therapy and tcmozolomide chemotherapy: differentiation of true progression from pseudoprogression with quantitative dynamic contrast-enhanced MR imaging. Radiology. 20l5;274(3):83(M0.

[0218] 48. Hatzoglou V, Yang TJ, Omuro A, Gavrilovic I, Ulaner G, Rubel J, et al. A

Prospective Trial of Dynamic Contrast Enhanced (DCE) MRI Perfusion and Fluorine-18 FDG PET- CT in Differentiating Brain Tumor Progression from Radiation Injury after Cranial Irradiation. Neuro Oncol, 2016; 18(6): 873-80.

[0219] 49. Kammen A, Morkin B, Cen S, Lingala SG, Arevalo-Perez J, Thomas AA, et al.

Multi-Center Study demonstrates Dynamic Contrast Enhanced Permeability MR] differentiates pseudoprogression from true disease progression in primary high-grade gliomas and metastatic melanoma. In: ASNR. 2016. [0220] 50. Wolchok JD, Hoos A, O’Day S, Weber JS, Hamid O, Lebbe C, et al. G uidelines for the evaluation of immune therapy activity in solid tumors: immune-related response criteria. Clin Cancer Res. 2009 Dec l;15(23):7412-20.

[0221] 51. Lee HL, Shankaranarayanan A, Pohost GM, Nayak K. Improved 3-Tesla cardiac cine imaging using wideband SSFP. Magn Reson Med. 2010 Jun 1 ;63(6):1716-22.

[0222] 52. Lee HL, Shankaranarayanan A, Pohost GM , Nayak K. Improved coronary MR angiography using wideband steady state free precession at 3 tesla with sub-millimeter resolution. J Magn Reson Imaging. Ming Hsieh Department of Electrical Engineering, University of Southern California, Los Angeles, California, USA. hsu-lei.lee@unikUmk-iTeiburg.de; 2010 Apr 30;31(5): 1224-9.

[0223] 53. Guo Y, Lebei RM, Zhu Y, Lingala SG, Shiroishi MS, Law M, et al. High- resolution whole-brain DCE-MRI using constrained reconstruction: Prospective clinical evaluation in brain tumor patients. Med Phys. 2016;43(5):2013-23.

[0224] 54. Best Hospitals for Cancer [Internet]. US News and World Report. 2016 [cited

2017 Jan 20], Available from: http://health.uynews.com/best-hospital8/rankings/cancer

[0225] 55. Lutti A, Hutton C, Finsterbusch J, Helms G, Weiskopf N. Optimization and validation of methods for mapping of the radiofrequency transmit field at 3T. Magn Reson Med. 2010 Jul;64(l):229-38.

[0226] 56. Cunningham C, Pauly J, Nayak K. Saturated double-angle method for rapid

BH· mapping. Magn Reson Med. 2006;55(6): 1326-33.

[0227] 57. Deoni SCL. High-resolution T1 mapping of the brain at 3T with driven equilibrium single pulse observation of T1 with high-speed incorporation ofRF field inhomogeneities (DESPOT1 -HIFI). J Magn Reson Imaging. 2007 Oc426(4): 1106-11.

[0228] 58. Sung K, Daniel BL, Hargreaves BA. Transmit B1+ field inhomogeneity and T1 estimation errors in breast DCE-MRI at 3 tesla. J Magn Reson Imaging. 2013 Aug;38(2}:454-9. [0239] 59. Di Giovanni P, Azlan C A, Ahcam TS, Semple SI, Gilbert FJ, Redpath TW. The accuracy of pharmacokinetic parameter measurement in DCE-MR1 of the breast at 3 T. Phys Med Biol. IOP Publishing; 2010 Jan 7;55(1):121~32.

[0230] 60, Azlan CA, Ahcam TS, Di Giovanni P, Semple SBC, Gilbert FJ, Redpath TW. Quantification techniques to minimize the effects of native T1 variation and B1 inhomogeneily in dynamic contrast-enhanced MRI of the breast at 3 T. Magn ResonMed. 2012 Feb;67(2):531-40.

[0231] 61. Fennessy FM, Fedorov A, Gupta SN, Schmidt EJ, Tempany CM, Mulkem R

V. Practical considerations in Tl mapping of prostate for dynamic contrast enhancement pharmacokinetic analyses. Magn Reson Imaging. 2012 Nov;30(9): 1224-33.

[0232] 62. Treier R, Steingoetter A, Fried M, Schwizer W, Boesiger P. Optimized and combined Tl and B1 mapping technique for fast and accurate Tl quantification in contrast-enhanced abdominal MRI. Magn Reson Med. 2007 Mar;57(3):568-76.

[0233] 63. Liberman G, Louzoun Y, Ben Bashat D. Ti mapping using variable flip angle

SPGR data with flip angle correction. J Magn Reson Imaging. 2014 Jul;40(l):171~80.

[0234] 64. Sacolick L, Wiesinger F, Hancu I, Vogel M. B 1 mapping by Bloch-Siegerl shift

Magn Reson Med. 2010;63(5):1315-22.

[0235] 65. Cheng H-LM, Wright G A, Rapid high-resolution T(l) mapping by variable flip angles: accurate and precise measurements in the presence of radiofrequency field inhomogeneily. Magn Reson Med. 2006 Mar,55(3):566-74.

[0236] 66. Yuan J, Chow SKK, Yeung DKW, Ahuja AT, King AD. Quantitative evaluation of dual-flip-angle Tl mapping on DCE-MRI kinetic parameter estimation in head and neck. Quant Imaging Med Surg. 2012 Dec;2(4):245-53.

$02371 67. Insko EK, Bolinger L. Mapping of the Radiofrequency Field. J Magn Reson

Ser A. 1993 Jun;103(l):82-5. [0238] 68. Akoka S, Franconi F, Scguin F, Le Pape A, Radiofrequency map of an NMR coil by imaging. Magn Reson Imaging. 1993 Jan;l 1(3):437-41.

[0239] 69. Sanz-Requena R, Prats-MontaMn JM, Marti-Bonmati L, Albcrich-Bayarri A,

Garcia-Marti G, P6rez R, et al. Automatic individual arterial input functions calculated from PCA outperform manual and population-averaged approaches for the pharmacokinetic modeling of DCE- MR images. J Magn Reson Imaging. 2014 Nov 20;

[0240] 70. Yin J, Sim H, Yang J, Guo Q. Automated detection of the arterial input function using normalized cut clustering to determine cerebral perfusion by dynamic susceptibility contrast- magnetic resonance imaging. J Magn Reson Imaging. 2015 Apr 2l;41(4):1071-8.

[0241] 71. Van Osch MJP, Vonken EJP a, Bakker CJG, Viergever M a. Correcting partial volume artifacts of the arterial input function in quantitative cerebral perfusion MRI. Magn Reson Med. 2001;45(3):477-85.

[0242] 72. Simonis FF, Sbrizzi A Beld E, Lagendijk JJ, Berg CA van den. Improving the

Arterial Input Function in Dynamic Contrast Enhanced MRI by fitting the signal in the complex plane. Ismrm. 2015;23:193.

[0243] 73. Schwarz G. Estimating the Dimension of a Model. Ann Stat 1978;6(2):461-4.

[0244] 74. Zhu Y, Guo Y, Lingala SG, Bames S, Lebel RM, Law M, et al. Evaluation of

DCE-MR1 data sampling, reconstruction, and model fitting using a digital brain phantom. In: Proc ISMRM 2015. 2015. p. 3052.

[0245] 75. Bosca RJ, Jackson EF. Creating an anthropomorphic digital MR phantom— an extensible tool for comparing and evaluating quantitative imaging algorithms. Phys Med Biol. IOP Publishing; 2016;61(2):974-82.

[0246] 76. Ellingson BM, Bendszus M, Boxerman J, Barboriak D, Erickson BJ, Smits M, et al. Concensus recommendations for a standardized Brain Tumor Imaging Protocol in clinical trials. 2015;17(August):l 188-98. [0247] 77. Liu JP, Ma MC, Wu CY, Tai JY. Test of equivalence and non-inferiority for diagnostic accuracy based on the paired areas under ROC curves. Stai Med. 2006;25(7): 1219-38.

[0248] Sparse Pre-Contrast T1 Mapping for DCE-MRI Calibration

[0249] References

[0250] 1. Guo, Y., Lebel, R. M., Zhu, Y., Lingala, S. G., Shiroishi, M. S., Law, M. andNayak,

K. (2016), High-resolution whole-brain DCE-MRI using constrained reconstruction: Prospective clinical evaluation in brain tumor patients. Med. Phys., 43 : 20.13-2023. doi: 10.1118/1.4944736

[0251] 2. Lebel, R. M., Guo, Y., Lingala, S. G., Frayne, R.,Nayak, K. S. "Highly Accelerated

DCE imaging with integrated T1 mapping." Proc. 1SMRM 25th Scientific Sessions, Honolulu, April

2017, pl38.

[0252] 3. Maier O., Schoormans J., Schloegl M., et al. Rapid T1 quantification from high resolution 3D data with model-based reconstruction. Magn Reson Med. 2018;00:1-18. https://doi.Org/l 0.1002/mrm.27502

[0253] Sparse DCE-MRI using a Temporal Constraint Learned from Clinical Data

[0254] 1. Heye AK, Culling RD, Hernandez MD, et al. Assessment of blood-brain barrier disruption using dynamic contrast-enhanced MRL A systematic review. Neuroimage Clin. 2014;6:262-274.

[0255] 2. Sourbron SP, Buckley DL. Classic models for dynamic contrast-enhanced MRL

NMR Biomed. 2013^6(8): 1004-1027.

[0256] 3. Pianey CR, Welch EB, Xu L, et al. Temporal sampling requirements for reference region modeling of DCE-MRI data in human breast cancer. J Magn Reson Imaging. 2009;30(1):121-

134.

[0257] 4. Paldino MJ, Barboriak DP. Fundamentals of quantitative dynamic contrast-enhanced

MR imaging. Clin North Am. 2009; 17(2):277-289. [0258] 5. Tudorica LA, Oh KY, Roy N, et al. A feasible high spatiotemporal resolution breast

DCE-MRI protocol for clinical settings. Magn Reson Imaging. 2012;30(9):1257-1267.

[0259] 6. Lustig M, Donoho D, Pauly JM. Sparse MRI: the application of compressed sensing for rapid MR imaging. Magn Reson Med. 2007;58(6): 1182-1195.

[0260] 7. Lebel RM, Jones J, Ferre JC, et al. Highly accelerated dynamic contrast enhanced imaging. Magn Reson Med, 2014;71(2):635-644.

[0261[ 8. Guo Y, Lingala SG, Bliesener Y, et al. Joint arterial input function and tracer kinetic parameter estimation from undersampled dynamic contrast-enhanced MRI using a model consistency constraint Magn Reson Med. 2018;79(5):2804-2815.

[0262] 9. Zhang T, Pauly JM, Levesque IR. Accelerating parameter mapping with a locally low rank constraint. Magn Reson Med. 2015;73(2):655-661.

[0263] 10. Aharon M, Elad M, Brockstein A. K-SVD: An algorithm for designing overcomplete dictionaries for sparse representation. IEEE Trans. Signal Process. 2006;54(ll):431 l- 4322.

[0264] 11. Tropp JA, Anna, and Gilbert C. Signal recovery from random measurements via orthogonal matching pursuit IEEE Trans. Inf. Theory. 2007;53(12):4655-4666.

[0265]

Claims

WHAT IS CLAIMED IS:

1. A blade box magnetic resonance system comprising:

a magnetic resonance imaging subsystem that includes a magnetic assembly having a polarizing magnetic coil and gradient coil assembly and receiving coils that collect magnetic resonance imaging data from a subject being treated for a tumor from which dynamic contrast magnetic resonance images are constructed; and

a control system that is operable to:

perform calibration scans; and

determine values for MRI biomarkers from magnetic resonance images obtained from the subject, the MRI biomarkers including tracer kinetic parameters obtained from a tracer kinetic model.

2. The system of claim 1 wherein values for MRI biomarkeifs) are used to provide a prognosis for treatment.

3. The system of claim 1 wherein values for MRI biomaiker(s) are used to monitor a treatment course.

4. The system of claim 1 wherein the subject is being treated with a chemotherapeutic agent and/or radiation.

5. The system of claim 1 wherein the control system is operable to compare values or time trends of MRI biomarker values to standard biomarkers values or trends determined from prior patient trials.

6. The system of claim 1 wherein the control system jointly estimates contrast concentration versus time images and tracer parameter maps from under-sampled (k,t) space data.

7. The system of claim 1 wherein the tracer kinetic model is a Patlak model and Q includes and where K

is a transfer constant from blood plasma into extracellular extravascular space (EES) and v_p is fractional plasma volume.

8. The system of claim 1, wherein the tracer kinetic model is an extended Tofts model and q includes

and v_e where

is a transfer constant from blood plasma into extracellular extravascular space (EES), v_p is fractional plasma volume,

is a transfer constant from EES back to the blood plasma, and v_e is a fractional EES volume.

9. The system of claim 1 wherein the control system is operable to construct dynamic images using a consistency constraint is applied to construct dynamic images from the magnetic resonance imaging data, the consistency constraint including a sum of a data consistency component and a model consistency component.

10. The system of claim 1 wherein the control system performs one or more or all of: provides automatic arterial input function selection;

perform motion correction;

perform hierarchical TK modeling; and

measuring biomafkers in a region -of -i nterest.

11. The system of claim 1 wherein the calibration scans include B1+ mapping, and pre- contrast Ti mapping with spatial resolution and coverage that match an area from which the magnetic resonance imaging data is obtained.

12. The system of claim 1 wherein the control system is operable to determine an arterial input function or vascular input function from the magnetic resonance imaging data, wherein the arterial input function includes a time variation of a magnetic resonance contrast agent at one or more predetermined locations in an artery of the subject.

13. The system of claim 1 wherein the control system is operable to automatically select the tracer kinetic model by specifying a collection of possible models (nested or not nested) from which a model identification method is applied to select a model for each voxel or region.

14. The system of claim 1 operable to perform -weighted dynamic susceptibility

contrast MR1.

15. The system of claim 1 operable to perform Ti -weighted DCE-MR1.

16. A method for determining a patient prognosis or monitoring patient treatment with a magnetic resonance imaging subsystem that includes a magnetic assembly having a polarizing magnetic coil and gradient coil assembly and receiving coils that collect magnetic resonance imaging data from a subject being treated for a tumor from which dynamic contrast magnetic resonance images are constructed; and

a control system that is operable to:

perform calibration scans; and

determine values for MRI biomarkers from magnetic resonance images obtained from the subject, the MRI biomarkers including tracer kinetic parameters obtained from a tracer kinetic model, the method comprising:

a) administering a magnetic resonance contrast agent to a subject b) collecting magnetic resonance imaging data from the subject for a tissue or organ;

c) selecting a tracer kinetic model to be applied to the magnetic resonance imaging data, the tracer kinetic model being defined by a plurality of tracer kinetic parameters;

d) reconstructing dynamic images from the magnetic resonance imaging data e) applying the tracer kinetic model to the reconstructed dynamic images to estimate tracer kinetic parameter maps; and

f) determining a patient prognosis and-'or monitoring effectiveness of treatment; and

g) adjusting patient treatment if necessary.

17. The method of claim 16 wherein the tumor is from a tissue or organ selected from the group consisting of brain, breast, prostate, liver, kidney, lung, heart, thyroid, pancreas, spleen, intestine, uterus, ovary, limbs, spine, bones, and eyes.

18. The method of claim 16 wherein the subject has a brain tumor.

19. The method of claim 16 wherein steps b)-c) are repeated for multiple slices through the subject being imaged and three-dimensional image data are reconstructed.