Method and program for evaluating the concomitant clinical dementia rating and its future outcome using predicted age difference

Abstract

A method for quantitatively evaluating cognitive impairment and its future changes from medical images of an individual's brain, the method comprising scanning the brain of the individual with a scanning device to obtain at least one medical brain image; processing the medical brain image to obtain at least one feature of the image; using a pre-established prediction model to determine the condition of cognitive impairment and predicting its future changes based on the at least one obtained feature.

Description

[0001] Cross - reference to related patent applications

[0002] This application claims the benefit and priority of U.S. Provisional Patent Application No. 63 / 216,028, filed on June 29, 2020, which is hereby incorporated by reference in its entirety. Technical field

[0003] The present disclosure relates to the technical field of evaluating predicted age difference (PAD) based on corrected diffusion MRI data and using PAD to predict current and future incident clinical dementia ratings and related procedures, as well as distortion / artifact correction of diffusion magnetic resonance imaging (MRI) data and its procedures for an integrated framework. Background art

[0004] The background description provided herein is to generally present the background of the present invention. The subject matter discussed in the background art section should not be regarded as prior art merely because it is mentioned in the background art section. Similarly, problems mentioned in the background art section or associated with the subject matter of the background art section should not be regarded as having been recognized in the prior art. The subject matter in the background art section merely represents different approaches, which may themselves also be inventions.

[0005] In clinical practice, doctors use various cognitive tests to evaluate cognitive decline and / or impairment due to dementia. Cognitive tests include dementia assessment tests and neuropsychological tests. Dementia assessment tests are developed based on clinically relevant cognitive symptoms, while neuropsychological tests are developed based on performance in different psychological domains. Widely used dementia assessment tests include the Clinical Dementia Rating (CDR), Mini-Mental State Examination (MMSE), Montreal Cognitive Assessment (MoCA), AD8, Alzheimer's Disease Assessment Scale - Cognitive Subscale (ADAS-cog), Alzheimer's Disease Cooperative Study ADL Scale for Mild Cognitive Impairment (ADCS-ADL-MCI), and Neuropsychiatric Inventory (NPI). A typical neuropsychological assessment will evaluate a person's general intelligence, attention and concentration, learning and memory, reasoning and problem-solving, language, visuospatial skills (perception), motor and sensory skills, mood, and behavior. An example of such an assessment is the Repeatable Battery for the Assessment of Neuropsychological Status (RBANS).

[0006] The CDR is a score estimated during a semi-structured interview used to stage the severity of dementia in Alzheimer's disease (AD) (Hughes et al., 1982) or non-AD dementia (Kazui et al., 2016). Compared to neuropsychological tests, the CDR interview focuses on the patient's daily function and is thus more in line with the standard diagnostic criteria for dementia (Lim et al., 2007).

[0007] The CDR interview is conducted by a certified doctor or nurse and given to the patient and appropriate informants. The interview consists of six different domains, namely, memory, orientation, judgment and problem-solving, community affairs, home and hobbies, and personal care. Functional impairment in each domain is graded by a five-point scale system: none = 0, questionable = 0.5, mild = 1, moderate = 2, severe = 3. The scores on the six domains are then converted into an overall CDR score. The CDR score has been proven to be reliable for diagnosing dementia and grading the severity of dementia (Morris, 1993). To date, the CDR has been included in the guidelines for dementia care and diagnosis by professional groups worldwide (Petersen et al., 2018).

[0008] However, the current CDR has two obvious limitations. The first limitation of the CDR is its inability to predict the future outcome of the CDR in the upcoming two to three years. Although the CDR is reliable for grading the severity of dementia, it cannot predict the progression of dementia. The future outcome of the CDR is clinically important because medical professionals not only assess the severity of a patient's dementia but also want to know whether the dementia will worsen in the next few years and at what rate. This limitation is clinically relevant, especially in patients with very mild or questionable dementia (i.e., CDR = 0.5). The outcomes of these patients are heterogeneous, with some progressing to dementia (CDR = 1 or higher), some reverting to normal cognition (CDR = 0), and some remaining at the same CDR (Daly et al., 2000). To address this limitation, some studies have shown that the sum of scores in the six domains (i.e., sum of box, SOB) (O′Bryant et al., 2008) or the subscale score in the orientation domain (Kim et al., 2017) or the combination of SOB and an additional memory test (Lee et al., 2006) can be used to predict dementia progression. However, all of the above measures rely heavily on the interview and are thus affected by clinician judgment, informant reliability, and cultural differences (Lim et al., 2007), which is the second limitation of the CDR.

[0009] Regarding this second limitation, the CDR has strict conditions for a successful interview. It requires the availability of well-trained medical professionals and reliable informants to ensure the accuracy of the CDR. In underdeveloped or undersupported regions, there may be a lack of trained professionals or reliable informants, thus undermining the availability of an accurate CDR.

[0010] To address these limitations, Maillard et al. addressed both limitations by using the free water (FW) content, a metric derived from post-processing of brain diffusion MRI. They reported that higher baseline FW was significantly associated with higher baseline CDR and a higher probability of becoming a higher CDR score in a later interview (Maillard et al., 2019). Additionally, to address the second limitation, Beheshti et al. estimated PAD based on individual morphometric features of cerebral gray matter extracted from T1-weighted (T1w) MRI and found that PAD was associated with scores of several clinical assessments of dementia, including CDR (Beheshti et al., 2018).

[0011] However, all existing methods / findings that address the CDR limitation have their own limitations and drawbacks. For example, Beheshti et al. demonstrated an association between PAD and CDR, but it failed to report how PAD predicts future outcomes of CDR. O′Bryant et al., Kim et al., and Lee et al. used information obtained from interviews to predict CDR outcomes; however, all scores they measured were interview-dependent and thus were affected by clinician judgment, informant reliability, and cultural differences. Maillard's FW metric has been shown to be associated with CDR at baseline, and it can also predict CDR changes in later visits. However, FW may be confounded with brain atrophy in pixels adjacent to ventricles and sulci due to the partial volume effect of cerebrospinal fluid (CSF).

[0012] Therefore, there is an urgent need for a method based on individual patient MRI scan data that can predict dementia progression for a given CDR and is free from these limitations.

[0013] On the other hand, in this field, diffusion MRI is widely used as a tool to understand the morphometric features of the individual brain. However, diffusion-weighted images (DWIs) included in diffusion MRI (dMRI) data have various artifacts that impair medical professionals' ability to estimate PAD. Among them, artifacts of susceptibility-induced distortion are the most notorious because they greatly degrade the performance of dMRI analysis. Such artifacts are typically retrospectively corrected by measuring or estimating a "field map" that represents how points in the undistorted space are displaced (or how distorted in the context of distortion correction) to the distorted space. The field map is a 3D scalar field because DWI is generally considered to be distorted along the phase-encoding (PE) direction due to low acquisition bandwidth.

[0014] Correction methods can be roughly divided into three categories. The first category is field-map based, which explicitly measures the field map by acquiring two gradient-echo images with precise scan parameters except for the echo time (TE), and then the field map is the difference in phase between the two images divided by the difference in TE values (Jezzard et al., 1998). The second category is reverse-gradient based (Andersson et al., 2003; Irfanoglu et al., 2015; Hedouin et al., 2017), in which two b 0 images (or two complete dMRI datasets) are acquired. The b 0 image here refers to DWI with b-value = 0 s / mm 2 (i.e., without applying diffusion-sensitizing gradients). The basic idea behind this category is that the two b 0 images (or two dMRI datasets) will have reverse distortions, so the field map can be estimated by registering them to their "midpoint" image, which is theoretically the undistorted image. The last category is anatomy-image based (Huang et al., 2008), in which the real b 0 image is registered to a pseudo b 0 image, which is constructed from an anatomy image, such as a T1-weighted (T1w) image or a T2-weighted (T2w) image. Since the anatomy image has no distortion, the estimated displacement map from the registration can be considered as a surrogate for the field map. Technically, both the reverse-gradient based method and the anatomy-image based method greatly involve registration calculations, so some efforts have been made to combine them by treating the pseudo b 0 image as the midpoint image in the reverse-gradient based method; this method has been shown to improve the performance of distortion correction (Irfanoglu et al., 2015).

[0015] However, the above methods have the drawback of significantly limiting the accuracy of dMRI data.

[0016] In particular, the anatomy-image based method uses an anatomy image (T1w or T2w) to construct a pseudo b 0 image, which is used as the reference image to drive the correction. The correction methods in this category register the real b 0 image (distorted) nonlinearly to the pseudo b 0Image (distortion-free) alignment is used to estimate the field map. When the images are well aligned, the image difference should be low. Thus, the goodness of alignment is typically measured by a similarity metric such as the sum of squared differences (SSD) (Huang et al., 2008), normalized mutual information (NMI) (Bhushan et al., 2015), correlation ratio (CR) (Bhushan et al., 2015), or cross-correlation (CC) (Irfanoglu et al., 2015).

[0017] Nonlinear registration is an ill-posed problem and thus is prone to failure in nature, especially when the images have large differences. Two main sources contribute to the image differences. One is the spatial difference, which is caused by the distortion due to susceptibility, and the other is the contrast difference, which is the intensity deviation between the real b 0 image and the pseudo-b 0 image. Conventional correction methods assume that the image differences are entirely caused by the distortion due to susceptibility and neglect the contribution of the intensity deviation. In this way, the overfitting problem may become significant for regions with large intensity deviations, leading to incorrect point-by-point correspondences and incorrect field maps.

[0018] One solution is to use a cost function that is less sensitive to the contrast difference, such as NMI, CR, or CC. However, the estimation using these cost functions is prone to getting stuck in local minima (or maxima, depending on the cost function), resulting in the spatial differences due to the distortion caused by susceptibility not being well corrected (Bhushan C et al., 2015). Another solution is to adjust the hyperparameters of the correction method to make the field map smoother and thus reduce the sensitivity to the image differences. Similar to the previous method, the lower sensitivity will lead to the real distortion not being well corrected.

[0019] In addition, some previous studies used general 3D registration correction methods such as large deformation diffeomorphic metric mapping (LDDMM) (Beg et al., 2005; Miller et al., 2006) and advanced normalization tool (ANT) (Avants et al., 2008) for registration. Using 3D registration correction methods to solve the 1D distortion correction problem has at least two drawbacks. First, the calculated free-form deformation must be further constrained to be along the PE direction (Irfanoglu et al., 2015), resulting in unnecessary calculations. A more serious second drawback is that the convergence of registration may not always correspond to the convergence of distortion correction because the registration correction method has more degrees of freedom than the distortion problem itself.

[0020] Therefore, a method for correcting and / or reducing artifacts in the dMRI data of an individual's brain is necessary for effective calculation and prediction of dementia progression for a given CDR. SUMMARY OF THE INVENTION

[0021] Regarding the technical limitations in the prior art, the present disclosure provides a brain age score and its procedure for predicting the dementia severity of a person and predicting his / her dementia outcome in the following years. Specifically, the brain age score is the PAD, and it evaluates the patient's current CDR and predicts the CDR change of the patient in the next few years.

[0022] Brain age is a type of brain predicted age derived from machine learning methods based on brain images (Cole and Franke, 2017). It requires using a large number of brain images from a group of healthy people and the chronological age of each individual to train the brain age model. Typically, features of brain structure or function are extracted from brain MRI data, and the features are trained by machine learning algorithms. The algorithm produces a prediction model that predicts the chronological age of an individual based on his / her brain image, and thus is called the brain predicted age. Once the model is trained and validated, it can be used to estimate the brain age of an individual, and then calculate the PAD defined as the difference between the brain age and the chronological age of the individual to indicate the aging state of the brain.

[0023] PAD has been shown to be a potential biomarker of cognitive aging. It is associated with cognitive impairment (Liem et al., 2017), fluid intelligence (Cole, 2020; Cole et al., 2018), and various cognitive abilities (Boyle et al., 2020; Richard et al., 2018). Beheshti et al. estimated PAD using morphometric features of individuals extracted from T1w MRI (Beheshti et al., 2018). They found that PAD was associated with scores of several clinical assessments of dementia, including the CDR. The literature by Beheshti demonstrated the ability of PAD to predict the CDR at the time of MRI scan. However, they did not have longitudinal data to study the prediction of PAD in the change of CDR over subsequent years.

[0024] The Open Access Series of Imaging Studies (OASIS) is a database that collects MRI and PET data, neuropsychological tests, and clinical data (https: / / www.oasis-brains.org / ). It is publicly available for the scientific community with the aim of studying the role of healthy aging and AD. The OASIS-3 dataset was released in 2019 and includes longitudinal data of 1098 participants aged from 42 to 95 years (LaMontagne et al., 2019). The participants include (1) individuals who are overall healthy and cognitively normal (i.e., Clinical Dementia Rating (CDR)=0), with or without a family history of Alzheimer's disease (AD); and (2) otherwise healthy individuals with suspected, mild, moderate, or severe dementia, i.e., CDR = 0.5, 1, 2, or 3 respectively. All participants gave informed consent in accordance with the procedures approved by the Institutional Review Board of the University of Washington School of Medicine.

[0025] Healthcare professionals and patients are very concerned about being able to predict a person's CDR in the next few years because appropriate interventions and follow-up visits can be prescribed individually. Therefore, this method is based on a system of 258 cognitively normal adults in the OASIS data to establish a pre-established prediction model (which is a brain age model), and apply this model to participants with different CDRs including 0, 0.5, and 1 at baseline (i.e., the time of MRI scan), and study the association between PAD and the change of CDR during several years of follow-up.

[0026] Specifically, the present invention established three conclusions. First, PAD is different among participants with different CDR scores of 0, 0.5, and 1 at baseline. Second, the PAD of participants with a stable CDR of 0.5 is different from that of participants with a CDR of 0.5 at baseline but changing to 0 or 1 during follow-up. Third, the PAD of participants with a stable CDR of 0 is different from that of participants with a baseline CDR of 0 but changing to 0.5 during the follow-up years.

[0027] Moreover, the present invention uses microstructural features of white matter (WM) extracted from diffusion tensor imaging (DTI) to calculate the predicted age difference (PAD) (referred to as WM PAD) and establish a method for correlating WM PAD and CDR, rather than using morphological measurement features of gray matter from T1w MRI (Beheshti et al., 2018) or fractional anisotropy (FA) content from diffusion MRI (Maillard et al., 2019). Specifically, the WM PAD of an individual is correlated with the person's concomitant CDR. Additionally, given a person's CDR, the WM PAD is also correlated with changes in CDR over the subsequent two to three years. This method for correlating WMPAD and CDR allows medical professionals to use WMPAD as a surrogate marker to predict CDR and its future outcomes in the coming years.

[0028] In one embodiment of the present invention, a method and associated computer-based program have been designed to read designated brain image data of an individual as input and estimate the predicted age difference (PAD) of a person, and based on a predefined cutoff value of PAD, determine whether the person is likely or unlikely to be cognitively normal when he / she is evaluated by the Clinical Dementia Rating (CDR). The method and program are used for a retrospective study of the cohort of the OASIS-3 database, and the performance of PAD in identifying patients who are likely or unlikely to be cognitively normal (CDR = 0) is evaluated by comparing the results with the CDR scores of the patients.

[0029] In one embodiment, the present invention relates to a method for quantitatively evaluating the likelihood of a disease condition and / or a change in condition from a medical image of an individual's brain, the method comprising: (1) processing an original medical image to obtain a first medical image and a second medical image; (2) processing the first medical image to obtain at least one tissue segment of the first medical image; (3) obtaining a deformation map of the individual's brain based on the at least one tissue segment obtained in step (2); (4) correcting artifacts in the second medical image; (5) using the artifact-corrected second medical image obtained in step (4) and the deformation map obtained from the first medical image in step (3) to obtain at least one diffusion-related value of the tissue; (6) calculating a predicted age difference (PAD) by inputting the extracted diffusion-related value obtained in step (5) into an established brain age prediction model; (7) predicting the likelihood of a disease condition or a change in condition from the PAD using a disease condition or condition change prediction model.

[0030] In one embodiment of the present invention, the disease condition refers to concomitant CDR. In one embodiment, the change in disease condition refers to the change in CDR within 2 - 3 years. In one embodiment, the diffusion-related values include fractional anisotropy (FA) and mean diffusivity (MD) values.

[0031] In another aspect of the present invention, in a method for quantitatively assessing the likelihood of a disease condition or a change in condition from medical images of an individual's brain, the first image is a T1-weighted (T1w) image.

[0032] In another aspect of the present invention, in a method for quantitatively assessing the likelihood of a disease condition or a change in condition from medical images of an individual's brain, the second image is obtained from a diffusion tensor imaging (DTI) sequence.

[0033] In another aspect of the present invention, in a method for quantitatively assessing the likelihood of a disease condition or a change in condition from medical images of an individual's brain, at least one brain tissue includes white matter (WM), gray matter (GM), or both.

[0034] In another aspect of the present invention, in a method for quantitatively assessing the likelihood of a disease condition or a change in condition from medical images of an individual's brain, step (3) includes: registering the segments of at least one brain tissue obtained in step (2) to the MNI space.

[0035] In another aspect of the present invention, in a method for quantitatively assessing the likelihood of a disease condition or a change in condition from medical images of an individual's brain, step (4) includes: inverting the T1w contrast to synthesize a pseudo-b 0 image and registering the DTI to the pseudo-b 0 image to correct the artifacts of the DTI data.

[0036] In another aspect of the present invention, in a method for quantitatively assessing the likelihood of a disease condition or a change in condition from medical images of an individual's brain, the brain tissue in step (5) is WM.

[0037] In another aspect of the present invention, in a method for quantitatively assessing the likelihood of a disease condition or a change in condition from medical images of an individual's brain, the diffusion-related values in step (5) include fractional anisotropy (FA) and mean diffusivity (MD).

[0038] In another aspect of the present invention, in a method for quantitatively assessing the likelihood of a disease condition or a change in condition from medical images of an individual's brain according to any one of the preceding claims, the disease condition is a Clinical Dementia Rating (CDR).

[0039] In another aspect of the present invention, the present invention relates to a method for evaluating an individual's Clinical Dementia Rating (CDR) and future changes in CDR based on the individual's brain MRI data according to the Predicted Age Difference (PAD), comprising: (1) acquiring MRI data of the individual's brain using T1-weighted (T1w) imaging and Diffusion Tensor Imaging (DTI) sequences; (2) processing the acquired T1w data to obtain white matter (WM) and gray matter (GM) segments; (3) registering the WM and GM segments obtained in step (2) to the MNI space to obtain a deformation map of the individual's brain; (4) correcting the artifacts of the DTI data; (5) obtaining fractional anisotropy (FA) and mean diffusivity (MD) maps of the individual in the native space based on the artifact-corrected DTI data; (6) using the FA and MD maps obtained in step (5) and the deformation map obtained in step (3) to extract the FA and MD values of the WM region of the individual's brain; (7) calculating the WM PAD by inputting the extracted FA and MD of the WM into the established brain prediction age model; (8) predicting the individual's CDR and CDR changes according to the WM PAD using the CDR prediction model and the CDR change prediction model.

[0040] In another aspect of the present invention, in the method for estimating an individual's Clinical Dementia Rating (CDR) and future changes in CDR based on the individual's brain MRI data according to the Predicted Age Difference (PAD), step (2) comprises: (i) correcting the SI inhomogeneity of the T1w data; (ii) segmenting the Tissue Probability Map (TPM); (iii) obtaining the WM and GM segments.

[0041] In another aspect of the present invention, in the method for estimating an individual's Clinical Dementia Rating (CDR) and future changes in CDR based on the individual's brain MRI data according to the Predicted Age Difference (PAD), step (2) further comprises: inverting the T1w contrast to synthesize a pseudo-b 0 image.

[0042] In another aspect of the present invention, in the method for estimating an individual's Clinical Dementia Rating (CDR) and future changes in CDR based on the individual's brain MRI data according to the Predicted Age Difference (PAD), step (4) comprises: registering the DTI data to the pseudo-b 0 image to correct the artifacts of the DTI data.

[0043] In another aspect of the present invention, in a method for estimating an individual's Clinical Dementia Rating (CDR) and future changes in CDR based on an individual's brain MRI data according to the Predicted Age Difference (PAD), step (5) includes: (i) estimating the diffusion tensor at each image pixel of the artifact-corrected DTI data; (ii) calculating the FA and MD at each pixel using the diffusion tensor indices derived from the estimated diffusion tensors at each pixel; (iii) generating FA and MD maps in the native space using the FA and MD at each pixel.

[0044] In another aspect of the present invention, in a method for estimating an individual's Clinical Dementia Rating (CDR) and future changes in CDR based on an individual's brain MRI data according to the Predicted Age Difference (PAD), MD = (λ1 + λ2 + λ3) / 3; FA = [3(Δλ1 2 + Δλ2 2 + ΔX3 2 ) / 2(λ1 2 + λ2 2 + λ3 2 )] 1 / 2 . λ1, λ2, and λ3 respectively represent the 1st, 2nd, and 3rd eigenvalues of the diffusion tensor.

[0045] In another aspect of the present invention, in a method for estimating an individual's Clinical Dementia Rating (CDR) and future changes in CDR based on an individual's brain MRI data according to the Predicted Age Difference (PAD), step (6) includes: (i) registering the FA and MD maps to the MNI space using the deformation map in step (3); and (ii) masking the registered FA and MD maps with the WM segment.

[0046] In another aspect of the present invention, in a method for estimating an individual's Clinical Dementia Rating (CDR) and future changes in CDR based on an individual's brain MRI data according to the Predicted Age Difference (PAD), the brain prediction age model is established by using the Gaussian process regression method to regress the chronological age of more than one individual for the FA and MD values within the WM region.

[0047] In another aspect of the present invention, in a method for estimating an individual's Clinical Dementia Rating (CDR) and future changes in CDR based on an individual's brain MRI data according to the Predicted Age Difference (PAD), the individual's brain is composed of white matter regions.

[0048] In another aspect of the present invention, in a method for estimating an individual's Clinical Dementia Rating (CDR) and future changes in CDR based on an individual's brain MRI data according to the Predicted Age Difference (PAD), the CDR change is the individual's CDR in the next 2 - 3 years since the MRI scan date.

[0049] In another aspect of the present invention, the present invention relates to a non-transitory computer-readable medium storing a program that causes a computer to execute a process for predicting an individual's Clinical Dementia Rating (CDR) and future changes in CDR, the process comprising: (1) processing an original medical image to obtain a first medical image and a second medical image; (2) processing the first medical image to obtain a fragment of at least one brain tissue of the first medical image; (3) obtaining a deformation map of the individual's brain based on the fragment of at least one brain tissue obtained in step (2); (4) correcting artifacts in the second medical image; (5) using the artifact-corrected second medical image obtained in step (4) and the deformation map obtained from the first medical image in step (3) to obtain fractional anisotropy (FA) and mean diffusivity (MD) values of the brain tissue; (6) calculating a predicted age difference (PAD) by inputting the extracted FA and MD obtained in step (5) into a established dMRI brain age model; (7) predicting the likelihood of a disease condition or condition change from the PAD using a disease condition or condition change prediction model.

[0050] In another aspect of the present invention, in a non-transitory computer-readable medium storing a program that causes a computer to execute a process for predicting an individual's Clinical Dementia Rating (CDR) and future changes in CDR, the first image is a T1-weighted (T1w) image; the second image is a diffusion tensor imaging (DTI) sequence.

[0051] In another aspect of the present invention, in a non-transitory computer-readable medium storing a program that causes a computer to execute a process for predicting an individual's Clinical Dementia Rating (CDR) and future changes in CDR, at least one brain tissue includes white matter (WM), gray matter (GM), or both.

[0052] In another aspect of the present invention, DWI is corrected using a novel method and / or a computer-based program to more efficiently remove and / or reduce artifacts in dMRI data.

[0053] First, for spatial differences, a registration method dedicated to susceptibility-induced distortion is developed. The first-order derivative (analytical) and second-order derivative (approximate) are derived. Thus, the method can use a Gauss-Newton optimization scheme to estimate the field map, which is very efficient. Second, a term called "bias field" is introduced to model the contrast difference between the true b 0 image and the pseudo-b 0 image. In this way, the image difference will be explained by two factors: susceptibility-induced distortion caused by the field map and intensity bias caused by the bias field. During the registration process, the field map and the bias field are estimated simultaneously. Their values are automatically adjusted by the estimation method according to the magnitude of the susceptibility-induced distortion and the level of the intensity bias.

[0054] These and other aspects of the present invention will become apparent from the following description of the preferred embodiments in conjunction with the accompanying drawings, although changes and modifications may be made therein without departing from the spirit and scope of the novel concepts of the present disclosure. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] The drawings illustrate one or more embodiments of the present invention and are used in conjunction with the written description to explain the principles of the present invention. Wherever possible, the same reference numerals are used throughout the drawings to refer to the same or like elements of the embodiments.

[0056] Figure 1 A flowchart showing a process for estimating CDR according to WM PAD;

[0057] Figure 2 A flowchart showing a process for acquiring T1w and DTI data;

[0058] Figure 3 A flowchart showing a process for image processing of T1w;

[0059] Figure 4 A flowchart showing a process for registration of GM and WM to MNI space;

[0060] Figure 5 A flowchart showing a process for artifact correction of DTI data;

[0061] Figure 6 A flowchart showing a process for reconstruction of MD and FA;

[0062] Figure 7 A flowchart showing a process for extraction of FA and MD;

[0063] Figure 8 A flowchart showing a process for estimating WM brain age and PAD;

[0064] Figure 9 A flowchart showing a process for prediction of CDR and CDR change;

[0065] Figure 10 A comparison between CN1-to-[CN2], [nD1], and [nD2] in Example 2 is shown;

[0066] Figure 11 A comparison between [D1]-to-CN, [nD1], and [D1]-to-D2 in Example 2 is shown;

[0067] Figure 12 A comparison between CN1-to-[CN2], D1-to-[CN], and [CN]-to-D1 in Example 2 is shown;

[0068] Figure 13 shows the spectrum of the PAD;

[0069] Figure 14 shows the process of scanning a participant's brain with an MRI scanner;

[0070] Figure 15 shows the process of converting an anatomical image into a pseudo-b 0 image;

[0071] Figure 16 shows the process of extracting a true b from dMRI data 0 image;

[0072] Figure 17 shows the rigid registration of the pseudo-b 0 image to the true b 0 image;

[0073] Figure 18 shows image difference modeling;

[0074] Figure 19 shows the algorithm for estimating β and v;

[0075] Figure 20 shows the result of correcting the artifacts of the participant's dMRI data;

[0076] Figure 21 shows the PAD comparison between CN1-to-[CN2], [nD1], and [nD2] of Example 3;

[0077] Figure 22 shows the ROC curve analysis for evaluating the performance of distinguishing CDR = 0 from CDR > 0;

[0078] Figure 23 shows the confusion matrix of the results using a cut-off PAD of 6 and the distribution of the two populations (i.e., CDR = 0 and CDR > 0) with respect to PAD;

[0079] Figure 24 shows a conceptual data flow diagram illustrating the data flow between different components / devices in an exemplary device; and

[0080] Figure 25 shows a diagram illustrating an example of a hardware implementation of a device employing a processing system. DETAILED DESCRIPTION

[0081] The present invention will now be described more fully hereinafter with reference to the accompanying drawings, in which exemplary embodiments of the invention are shown. However, the present invention may be embodied in many different forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the invention to those skilled in the art. Like reference numerals refer to like elements throughout.

[0082] In the context of the present invention and in the specific context in which each term is used, the terms used in this specification generally have their ordinary meaning in the art. Certain terms used to describe the present invention are discussed below or elsewhere in the specification to provide additional guidance to the practitioner regarding the description of the present invention. For convenience, certain terms may be highlighted, for example, using italics and / or quotation marks. The use of highlighting has no effect on the scope and meaning of the term; in the same context, the scope and meaning of the term are the same whether or not it is highlighted. It should be understood that the same thing can be expressed in more than one way. Thus, alternative languages and synonyms may be used for any one or more of the terms discussed herein, and whether or not a term is elaborated or discussed in detail herein does not have any special significance. Synonyms for certain terms are provided. The recitation of one or more synonyms does not exclude the use of other synonyms. The use of examples anywhere in this specification (including examples of any term discussed herein) is merely illustrative and does not limit the scope and meaning of the present invention or any exemplary term. Likewise, the present invention is not limited to the various embodiments given in this specification.

[0083] It should be understood that, as used in the description herein and throughout the appended claims, the meanings of "a," "an," and "the" include plural references unless the context clearly dictates otherwise. Also, it should be understood that when an element is referred to as being "on" another element, the element can be directly on the other element or intervening elements may be present therebetween. In contrast, when an element is referred to as being "directly on" another element, no intervening elements are present. As used herein, the term "and / or" includes any and all combinations of one or more of the associated listed items.

[0084] It should be understood that although the terms first, second, third, etc. may be used herein to describe various elements, components, regions, layers, and / or sections, these elements, components, regions, layers, and / or sections should not be limited by these terms. These terms are only used to distinguish one element, component, region, layer, or section from another element, component, region, layer, or section. Thus, a first element, component, region, layer, or section discussed below may be referred to as a second element, component, region, layer, or section without departing from the teachings of the present invention.

[0085] In addition, relative terms such as "lower" or "bottom" and "upper" or "top" may be used herein to describe the relationship of one element to another, as illustrated in the figures. It should be understood that relative terms are intended to encompass different orientations of the device in addition to the orientation depicted in the figures. For example, if the device in one figure is flipped, an element described as on the "lower" side of another element will then be oriented on the "upper" side of the other element. Thus, depending on the specific orientation of the figure, the exemplary term "lower" can encompass both the "lower" and "upper" orientations. Similarly, if the device in one figure is flipped, an element described as "below" or "beneath" another element will then be oriented "above" the other element. Thus, the exemplary terms "below" or "beneath" encompass both the above and below orientations.

[0086] It should also be understood that the terms "comprises", "comprising", "has", "having", "contains", "containing", "involves", and the like are open-ended, i.e., they mean including but not limited to. When used in this disclosure, they specify the presence of the stated features, regions, integers, steps, operations, elements, and / or components, but do not preclude the presence or addition of one or more other features, regions, integers, steps, operations, elements, components, and / or groups thereof.

[0087] Unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs. It should also be understood that terms, such as those defined in commonly used dictionaries, should be interpreted as having a meaning that is consistent with their meaning in the context of the relevant art and this disclosure, and will not be interpreted in an idealized or overly formal sense unless expressly so defined herein.

[0088] As used in this disclosure, "about", "approximately", "approximate" or "substantially" generally should mean within 20% of a given value or range, preferably within 10%, more preferably within 5%. Numerical quantities given herein are approximate, meaning that the term "about", "approximately", "approximate" or "substantially" may be implicit if not expressly stated.

[0089] As used in this disclosure, the phrase "at least one of A, B, and C" should be interpreted to mean a logical (A or B or C) using non-exclusive logical OR. As used herein, the term "and / or" includes any and all combinations of one or more of the associated listed items.

[0090] Embodiments of the present invention are illustrated in detail below with reference to the accompanying drawings. The following description is merely illustrative in nature and is not intended to limit the present invention, its application, or its use. The broad teachings of the present invention may be implemented in a variety of forms. Thus, while the present invention includes specific examples, the true scope of the present invention should not be so limited since other modifications will become apparent after study of the drawings, the specification, and the appended claims. For clarity, like reference numerals will be used to identify like elements in the drawings. It should be understood that one or more steps within a method may be executed in a different order (or concurrently) without changing the principles of the present invention.

[0091] Embodiments of the present invention will be described in conjunction with the appended Figure 1 - 20 drawings. For purposes of the present invention, as specifically implemented and broadly described herein, the present invention in one aspect relates to methods and related apparatuses for calculating and predicting CDR based on individual WM PAD MRI data.

[0092] Figure 1 A general flowchart of the present invention is shown: a process of evaluating CDR and predicting CDR changes using WM PAD through a program. The present invention begins with acquiring MRI data using T1w imaging and diffusion-weighted imaging sequences (step 1). To ensure selection of the same WM components from different brains, spatial registration of the MRI images of each individual to a standard template in the MNI space is enforced (steps 2 and 3). In one embodiment, steps 2 and / or 3 are spatial normalization processes during which MRI images such as T1w images are segmented and fragments of certain brain tissues are registered to the MNI space. Thus, deformation maps of certain brain tissues are obtained during the spatial normalization process. To ensure accurate quantification of the diffusion-weighted indices (i.e., fractional anisotropy (FA) and mean diffusivity (MD)), artifacts in the diffusion-weighted MRI data should be corrected, followed by accurate estimation of the FA and MD indices (steps 4 and 5). After the FA and MD maps in native space and the deformation matrix between native space and MNI space have been obtained, we can transform the FA and MD maps to the MNI space and extract the FA and MD indices in the WM component (step 6). In one embodiment, steps 5 and / or 6 are feature quantification processes during which FA and MD maps based on artifact-corrected diffusion-weighted MRI data are obtained and the FA and MD indices of certain brain tissues are extracted. The resulting FA and MD values are used as inputs to an established dMRI-brain age model to estimate brain age and WMPAD (step 7). In the final step, WMPAD is used to determine CDR, and future changes in CDR are predicted using a model pre-established in the program (step 8). Detailed descriptions of each step are described below.

[0093] Figure 2Shows the process of obtaining T1w and DTI data through a program.

[0094] Step 1.1 is the MRI scan of the human brain. The person is scanned by a 1.5 Tesla or 3 Tesla MRI system with a phased array multi-channel head coil. MRI scans are contraindicated for participants within the first three months of pregnancy, with a pacemaker or defibrillator implanted, or who have had vascular clipping or stenting within 6 months.

[0095] Step 1.2 is the acquisition of T1w. T1w data is acquired using a 3D magnetization-prepared rapid gradient echo sequence. The imaging parameters are detailed as follows: repetition time (TR) = 2400 ms, echo time (TE) = 3.16 ms, inversion time (TI) = 1000 ms, field of view (FOV) = 256×256×176 mm3, and matrix size = 256×256×176.

[0096] Step 1.3 is the acquisition of DTI data. DTI data is acquired using diffusion-weighted 2D single-shot spin echo planar imaging, which uses 24 different magnitudes of diffusion sensitivity (b-values) corresponding to 24 different directions (b-vectors) evenly distributed in q-space.

[0097] Step 1.4 is to obtain the T1w and DTI data. Thus, one T1w data and one DTI data are obtained for each participant for processing in steps 2 and 4.

[0098] Figure 3 Shows the steps of image processing of T1w performed by the program during the spatial normalization process.

[0099] Step 2.1 is the T1w image of the person. The T1w image is processed using the segmentation toolbox in SPM12 software (Wellcome Trust Centre for Neuroimaging, University College London, London, UK).

[0100] Step 2.2 is the correction of SI inhomogeneity on T1w. The toolbox corrects the signal intensity inhomogeneity.

[0101] Step 2.3 is to segment the tissue probability map (TPM) using the segmentation toolbox in SPM12. After the SI inhomogeneity has been corrected, the toolbox segments the T1w image into various tissue segments, including gray matter (GM), white matter (WM), cerebrospinal fluid (CSF), bone, scalp, etc.

[0102] Step 2.4 is to obtain the GM and WM segments. Thus, six TPMs are obtained, and the GM and WM segments are selected from these TPMs for processing in step 3.

[0103] Step 2.5 is to invert the T1w contrast. In parallel with Steps 2.3 and 2.4, the contrast of the T1w image that has been corrected for signal intensity inhomogeneity is inverted to synthesize a pseudo-b 0 image.

[0104] Step 2.6 is to obtain the pseudo-b 0 image. In this way, we obtain the pseudo-b 0 image for each individual for processing in Step 4.

[0105] Figure 4 Shows the process of registering GM and WM to the MNI space during the spatial normalization process.

[0106] Step 3.1 is the GM and WM segments of the individual. The GM and WM segments are registered to the ICBM152 template defined in the MNI space.

[0107] Step 3.2 is to register the GM and WM segments to the ICBM152 template. This registration calls a process that is a variant of the large deformation diffeomorphic metric mapping (LDDMM) (Beg et al., 2005; Miller et al., 2006). Specifically, the velocity is iteratively estimated by emitting an initial velocity located in the ICBM152 space along the time dimension towards the native space of the individual. In the registration, the process is divided into 10 evenly spaced intervals, and an isotropic Gaussian filter with a full width at half maximum (FWHM) of 10 mm is used to ensure the smoothness of the initial velocity.

[0108] Step 3.3 is to obtain the deformation map of the individual. At the final stage of convergence, a deformation map is obtained that transforms the individual's TPM to match the TPM in the ICBM152 template. The deformation map is used for processing in Step 6.

[0109] Figure 5 Shows the process of artifact correction for DTI data.

[0110] Step 4.1 is the DTI image of the individual. The original DTI data has various artifacts, including susceptibility-induced distortion, eddy current-induced distortion, head motion, and intensity inhomogeneity, which vary from person to person and should be corrected to ensure accurate registration to the MNI space.

[0111] Step 4.2 is to register the DTI to the pseudo-b 0 image. The DTI data is processed using a novel registration-based method that corrects these artifacts in an integrated framework. The registration-based method registers the original DTI data of the individual to the pseudo-b 0 image of the same individual, and this pseudo-b 0 image is used as a reference image without distortion and intensity inhomogeneity.

[0112] Step 4.3 is to reduce artifacts using a series of artifact models. By invoking the artifact models in the integration framework, this process can reduce the artifacts in the DTI data.

[0113] Step 4.4 is to obtain the artifact-corrected DTI image. In this way, the artifact-corrected DTI image is obtained. When the estimation reaches convergence, the corrected DTI data is also aligned with the T1w image.

[0114] Figure 6 The process of reconstructing MD and FA during the feature quantization process is shown.

[0115] Step 5.1 is the artifact-corrected DTI data. This artifact-corrected DTI data is processed to estimate the diffusion tensor and the diffusion tensor indices derived therefrom.

[0116] Step 5.2 is to estimate the diffusion tensor at each image pixel. The diffusion tensor estimation is performed using the method proposed by (Koay et al., 2007). Briefly, first, weighted linear least squares is performed, and then the result is used as the initial estimate for the constrained nonlinear least squares to obtain the diffusion tensor, which is ensured to be positive definite (i.e., all the eigenvalues of the diffusion tensor are positive).

[0117] Step 5.3 is to calculate FA and MD at each pixel. The diffusion tensor indices (i.e., fractional anisotropy (FA) and mean diffusivity (MD)) are derived from the estimated diffusion tensor at each pixel. The MD and FA values are determined using the standard formulas: MD = (λ1 + λ2 + λ3) / 3, and FA = [3(Δλ1 2 + Δλ2 2 + Δλ3 2 ) / 2(λ1 2 + λ2 2 + λ3 2 )] 1 / 2 , where λ1, λ2, and λ3 represent the first, second, and third eigenvalues of the diffusion tensor, respectively, and Δλ1, Δλ2, and Δλ3 represent λ1 - MD, λ2 - MD, and λ3 - MD, respectively.

[0118] Step 5.4 is to obtain the FA and MD maps in the native space. In this way, the FA and MD maps of each individual in the native space are obtained.

[0119] Figure 7 The process of extracting FA and MD during the feature quantization process is shown.

[0120] Step 6.1 is the FA and MD maps of individuals in the native space. After the FA and MD maps have been obtained in step 5, the FA and MD values must be extracted from the normalized WM segments.

[0121] Step 6.2 is to register the FA and MD maps to the MNI space. To this end, the FA and MD maps in the native space are normalized to the MNI space by the deformation map obtained in Step 3.

[0122] Step 6.3 is to mask the registered FA and MD maps with the WM segment. After being normalized to the MNI space, the FA and MD values in the WM are masked with the standardized WM segment in the ICBM152 template.

[0123] Step 6.4 is to obtain the FA and MD values of the WM. In the final step, the FA and MD values of the WM of each individual are obtained. The results are used for the processing in Step 7.

[0124] Figure 8 The process for estimating the WM brain age and PAD is shown.

[0125] Step 7.1 is the FA and MD of the person in the WM. The FA and MD of the person are used as the input to the established dMRI-brain age model, which is pre-trained and validated using DTI data acquired from the same MRI scanner.

[0126] Step 7.2 is to input the FA and MD values into the established dMRI-brain age model to estimate the brain age. To establish the dMRI-brain age model, the Gaussian process regression (GPR) method is used to regress the chronological age of the participants against the FA and MD values within the WM region. The output of the model is the predicted age estimated based on the FA and MD values of the person and the GRP model, which is called the dMRI-brain age. The performance of the model has been tested and quantified according to the mean absolute error (MAE) and the Pearson correlation coefficient (r) between the dMRI-brain age and the chronological age.

[0127] Step 7.3 is to calculate the WM PAD according to the estimated brain age and chronological age of the person. After obtaining the dMRI-brain age, the predicted age difference (WM PAD) of the white matter is calculated by the following formula: dMRI-brain age - chronological age.

[0128] Step 7.4 is to obtain the WM PAD. In this way, the WM PAD is obtained for estimating the CDR and CDR change.

[0129] Figure 9 The process for determining the CDR and predicting the future change of CDR is shown.

[0130] Step 8.1 is the WM PAD of the person. The WM PAD of the person is used as the input to the established prediction model to determine his / her CDR and / or predict the future change of CDR.

[0131] Step 8.2 is to determine CDR from the established model that associates WM PAD with CDR. To establish a predictive model of CDR from WM PAD, a linear regression model is applied to a data set from a population in which WM PAD and incident CDR are pre-known. In one embodiment, a data set from a population in which WM PAD and incident CDR are pre-known is used to establish a brain age model.

[0132] Step 8.3 is to obtain the determined CDR. Using the established regression model of CDR and / or brain age model, the CDR of a person is determined from the WM PAD of the same person.

[0133] Step 8.4 is to predict future changes in CDR from the established model that associates WM PAD with future changes in CDR. To establish a predictive model of future changes in CDR from WMPAD, a linear regression model is applied to a data set from a population in which WMPAD and CDR records 2 to 3 years after WM PAD measurement are known. In one embodiment, such a model is a brain age model. The CDR change is defined as the difference between the CDR obtained simultaneously with the WM PAD measurement and the CDR obtained 2 to 3 years later.

[0134] Step 8.5 is to obtain the predicted CDR change. Using the established regression model of CDR change, the CDR change of the same person is predicted from the WM PAD of the person.

[0135] In one embodiment, the pre-established predictive model includes the dMRI-brain age model used in WM PAD calculated according to the extracted FA and MD values in step 7 and the models that relate WM PAD to CDR and CDR change in step 8.

[0136] In one embodiment, medical brain images of multiple patients are processed to extract at least one feature of each patient. Given a set of training data, the pre-established predictive model is constructed by associating the features with the cognitive impairment status using machine learning methods. For independent data, the model can be used to predict the status of cognitive impairment by inputting brain image features.

[0137] Figure 24 is a conceptual data flow diagram illustrating the data flow between different components / devices in the exemplary device 2400. In one embodiment, one or more image acquisition devices (e.g., MRI scanners) collect brain image data of a patient and send the data to the receiving component 2402 for further processing. Then, the receiving component 2402 sends the image data to the image processing component 2403. It should be noted that in other embodiments, the image acquisition device can directly send the image data to the image processing component 2403 without using the receiving component 2402.

[0138] The image processing component 2403 processes steps 2.1 - 2.6 to obtain the pseudo-b 0 image, as well as GM and WM segment information. Then, for steps 3.1 - 3.3, the image processing component sends the GM and WM segment information to the registration component 2406 to obtain the deformed image. In parallel, for steps 4.1 - 4.4, the pseudo-b 0 image information and DTI data are sent to the artifact correction component 2405, during which step, the artifact-corrected DTI image is obtained.

[0139] In one embodiment, the prediction component 2407 then receives the deformed image from the registration component 2406 and the artifact-corrected DTI data from the artifact correction component 2405. The prediction component 2407 can then proceed to steps 5 - 8. In another embodiment, steps 6.1 - 6.4 can be performed at the registration component 2406 to obtain the FA and MD values of the WM.

[0140] Figure 25 FIG. 2500 is a diagram illustrating an example of a hardware implementation of the processing system 2400. The processing system 2500 can be implemented with a bus architecture generally represented by a bus 2503. Depending on the specific application and overall design constraints of the processing system 2500, the bus 2503 can include any number of interconnected buses and bridges. The bus 2503 links together various circuits, including one or more processors and / or hardware components represented by one or more processors 2501, an image processing component 2403, an artifact correction component 2405, a registration component 2406, a prediction component 2407, and a computer-readable medium / memory 2502. The bus 2503 can also link various other circuits, such as a timing source, peripheral devices, voltage regulators, and power management circuits, etc.

[0141] The processing system 2500 can be coupled to an image acquisition device 2401, which can be an MRI scanner or other clinical image acquisition device.

[0142] The processing system 2500 includes one or more processors 2501 coupled to a computer-readable medium / memory 2502. The one or more processors 2501 are responsible for general processing, including executing software stored on the computer-readable medium / memory 2502. When executed by the one or more processors 2501, the software causes the processing system 2500 to perform the various functions described above for any particular device. The computer-readable medium / memory 2502 can also be used to store data manipulated by the one or more processors 2501 when executing the software. The processing system 2500 further includes at least one of an image processing component 2403, an artifact correction component 2405, a registration component 2406, and a prediction component 2407. These components can be software components that run in the one or more processors 2501, reside / store in the computer-readable medium / memory 2502, one or more hardware components coupled to the one or more processors 2501, or some combination thereof.

[0143] It should be understood that the specific order or hierarchy of the blocks in the disclosed process / flowchart is an illustration of an exemplary method. Based on design preferences, it should be understood that the specific order or hierarchy of the blocks in the process / flowchart can be rearranged. Further, some blocks can be combined or omitted. The appended method claims present the elements of the various blocks in an example order and are not meant to be limited to the specific order or hierarchy presented.

[0144] Example 1

[0145] This artifact correction step 4.1 - 4.3 and Figure 5 In Figures 14 to 20 is presented in a more detailed manner.

[0146] Specifically, the brains of participants are scanned by an MRI scanner. dMRI data and anatomical images (T1w or T2w) ( Figure 14 ) are acquired. The anatomical images are converted into pseudo-b 0 images, which have an image contrast similar to that of the real b 0 images ( Figure 15 ). Real b 0 images are extracted from the dMRI data ( Figure 16 ). The pseudo-b 0 images are rigidly registered to the real b 0 images. The registered images are denoted and the real b 0 b 0 images are Here, denotes the image domain. The coordinates of f 0 f 0 are r ∈ Ω r ∈ Ω. f 1 f 1The PE direction is ( Figure 17 ).

[0147] Assume that the bias field is multiplicative and is modeled by e β , where, Thus, will be similar to f 1 . The smoothness of the β field is achieved via the differential operator L β , where the norm of β is defined as ||β|| 2 = <L β β|β> 2 . According to Figure 18 , is used to represent the field map (in Hz). The relationship between the distortion map and the field map is:

[0148]

[0149] where k = τV, τ is the effective echo spacing time, and V is the field of view of the DW image in the PE direction. Let v = kψ be the initial velocity. The smoothness of v is achieved via the differential operator L v , where the norm of v is defined as ||v||| 2 = <L v v|v) 2· .

[0150] The distortion-corrected image is where, represents function composition( Figure 18 ).

[0151] Figure 19 Shows the process of estimating β and v. The cost function to be minimized is E 1 = E d + E β + E v ,

[0152] where,

[0153]

[0154] We use the sum of squared differences (SSD) to quantify the data match between the true b 0 image and the pseudo b 0 image (E d ), where regularization penalizes non-smooth bias fields (E β ) and non-smooth initial velocities (E v ). The parameters σ d , σ β and σ v are the weights of the data match term and the regularization term, respectively.

[0155] For convenience of annotation, we define some auxiliary functions. The function b = vec(A) converts the 3D field into a vector by cascading it column by column where n xyz is the total number of voxels. The reverse operation is achieved by A = ivec(b). Additionally, the function B = diag(b) converts the b vector into a diagonal matrix B.

[0156] The Gauss - Newton method is used to update the value of β. Let

[0157]

[0158] and b β = vec(β), g β = vec(g β ), and H β = diag(vec(H β )). The descent deviation is calculated via the following formula

[0159] δb β = (σ d H β + L β ) -1 (σ d g β + σ β L β b β ), (4)

[0160] Or in the form of a 3D field δβ = ivec(δb β ), and then the β field is updated via the following formula

[0161] β (iter+1) = β (iter) - γ β ·δβ (iter ), (5)

[0162] where 0 < γ β ≤ 1 is the scaling factor that controls the size of the descent. The differential operator L β is implemented via the following formula

[0163]

[0164] where the first term in the integral is the membrane energy and the second term is the bending energy, and λ β，1 and λ β，2 control their regularization strengths respectively. In fact, this differential operator is expressed in matrix form .

[0165] Similar to the deviation field, the Gauss-Newton method is used to update the initial velocity field v. Let

[0166]

[0167] Here, is the gradient along the p direction. Also let b v = vec(v), g v = vec(g v ), H v = diag(vec(H v ))), the descent deviation is calculated via the following formula

[0168] δb v = (σ d h v + L v )) -1 (σ d g v + σ v L v b v ), (8)

[0169] Or in the form of a 3D field δv = ivec(δb v ), then the v field is updated via the following formula

[0170] v (iter+1) = v (iter) - γ v ·δv (iter) ), (9)

[0171] where 0 < γ v ≤ 1 controls the descent magnitude. The differential operator L v is defined in the same form as L β , where the intensity of the film energy and the intensity of the bending energy are controlled by λ v，1 and λ v，2 respectively. This difference operator is encoded in the matrix .

[0172] Compared with existing methods, the current registration method is dedicated to the distortion caused by magnetic susceptibility, which is a 1D deformation along the PE direction. Although the first derivative (g v ) has been shown above, the approximate Hessian (H v ) of the second derivative is our invention. Therefore, this method uses the Gauss-Newton optimization scheme to estimate the field map, which is more efficient compared with existing methods.

[0173] Moreover, the term called "deviation field" is introduced to account for the difference between the true b 0 image and the pseudo b 0Model the contrast differences between images. Existing anatomical image-based methods do not explicitly consider the contrast differences during the estimation process.

[0174] Figure 20 An example of using the proposed method is shown. As can be observed, the reference pseudo-b 0 image ( Figure 20 (e)), the original b 0 image ( Figure 20 (a)) has large distortions around the frontal region. If the artifact model does not consider the intensity bias, the method falls into a local minimum, resulting in the distortion not being fully corrected, such as in the regions indicated by the green arrows in the b 0 image ( Figure 20 (b)) and the fractional anisotropy (FA) map ( Figure 20 (c)). On the other hand, using the bias field ( Figure 20 (i)) to model the intensity bias can lead to proper registration, so the b 0 image ( Figure 20 (f)) and the FA map ( Figure 20 (g)) are very well aligned with the pseudo-b 0 image. The field map estimated using the model with intensity bias ( Figure 20 (h)) is much smoother than the field map estimated using the model without intensity bias ( Figure 20 (d)). The interpretation of the results is the large local image differences between the real b 0 image and the pseudo-b 0 image. Recall that two main sources contribute to the image differences: spatial differences (caused by distortions due to susceptibility) and contrast differences (caused by intensity bias). If the algorithm ignores the contribution of the intensity bias, the algorithm will try to reduce the cost function (i.e., SSD) by spatially aligning the registered images. It is well known that registration, especially non-linear registration, is an ill-posed problem. Thus, registering regions with large intensity bias easily leads to incorrect point-by-point correspondences and incorrect field maps.

[0175] The results reveal (1) that the estimated field map can adequately correct the distortions caused by susceptibility, and (2) that modeling the intensity bias can facilitate the correction of the distortions caused by susceptibility.

[0176] Example 2

[0177] Method

[0178] 1. Subjects

[0179] Participants in OASIS-3 were recruited through different projects related to the Knight ADRC. The participants were (1) generally healthy and cognitively normal (CDR = 0) individuals with or without a family history of AD and (2) generally healthy individuals with CDR = 0.5, 1, or 2. Exclusion criteria included medical conditions that precluded longitudinal participation, such as end-stage renal disease requiring dialysis or contraindications for MRI studies, such as pacemaker implantation. All participants gave informed consent in accordance with procedures approved by the Institutional Review Board of the University of Washington School of Medicine.

[0180] 2. Data Selection

[0181] Participants in OASIS-3 underwent one to multiple sessions of brain MRI scans on three 3T MRI scanners (Siemens, Erlangen, Germany), which included two scanner models, namely Biograph and TIM Trio. To reduce scanner-related variability, the same acquisition protocol was used for T1w imaging and diffusion tensor imaging (DTI). MRI data acquired from the same scanner model (TIM Trio) were selected for analysis. Participants with images presenting severe image artifacts (such as motion blur on T1w or failure to pass artifact correction on DTI) were excluded due to the need for acceptable image quality of T1w and DTI data for analysis. Thus, a total of 529 participants were recruited for subsequent analysis out of 575 MRI scan sessions.

[0182] 3. MRI Data Acquisition

[0183] As presented above in Figure 2 and the related description, 3D magnetization-prepared rapid gradient echo sequences were used to acquire T1w data. The imaging parameters are detailed as follows: repetition time (TR) = 2400 ms, echo time (TE) = 3.16 ms, inversion time (TI) = 1000 ms, field of view (FOV) = 256×256×176 mm3, and matrix size = 256×256×176. DTI data were acquired using diffusion-weighted 2D single-shot spin echo planar imaging, which used 24 different magnitudes of diffusion sensitivity (b-values), corresponding to 24 different directions (b-vectors), repeated twice. The b-values and b-vectors are listed in Supplementary File 1. The imaging parameters are described as follows: TR = 14500 ms, TE = 112 ms, FOV = 224×224 mm 2 , matrix size = 112×112, slice thickness = 2 mm. Two DTI acquisitions were concatenated to form a single DTI dataset, so for each participant, there was one T1w data and one DTI data in each MR session.

[0184] 4. Image Processing

[0185] All MRI data are processed according to the following procedure. It requires tissue segmentation, artifact correction, diffusion tensor estimation, image registration, and diffusion index extraction.

[0186] 4.1 As presented in Figure 3 and the related description, for brain segmentation on T1w images, the T1w images are processed using the segmentation toolbox in SPM12 software (Wellcome Trust Centre for Neuroimaging, University College London, London, UK). The toolbox corrects the signal intensity inhomogeneity and segments the T1w images into various tissue components, including gray matter (GM), white matter (WM), cerebrospinal fluid (CSF), bone, scalp, etc. The output of the segmentation toolbox is six TPMs, each of which represents a brain tissue component, and the T1w image with intensity inhomogeneity is corrected. The corrected T1w image is used for subsequent artifact correction of DTI data, and the TPMs of GM and WM are used for spatial normalization of the image to the MNI space.

[0187] 4.2 Artifact correction of DTI data: As presented in Figure 5 and the related description, the original DTI data has various artifacts, including susceptibility-induced distortion, eddy-current-induced distortion, head motion, and intensity inhomogeneity. The DTI data is processed using a novel registration-based procedure that corrects these artifacts in an integrated framework. Briefly, this procedure registers the original DTI data to the pseudo-b 0 image, which is synthesized by inverting the contrast of the T1w image whose intensity inhomogeneity has been corrected in step 4.1. In this way, the pseudo-b 0 image serves as a reference image without distortion and intensity inhomogeneity. By incorporating the artifact model into the integrated framework, this procedure can reduce the artifacts in the DTI data. At the same time, when the estimation reaches convergence, the corrected DTI data is easily spatially aligned with the T1w image.

[0188] 4.3 Diffusion tensor estimation: According to Figure 6 , the artifact-corrected DTI data is processed using the method proposed by (Koay et al., 2007) to estimate the diffusion tensor. Briefly, first, weighted linear least squares is performed, and then the result is used as the initial estimate for constrained nonlinear least squares to obtain the diffusion tensor, which is ensured to be positive definite (i.e., all the eigenvalues of the diffusion tensor are positive). The associated fractional anisotropy (FA) and mean diffusivity (MD) are derived from the estimated diffusion tensor at each pixel. The MD and FA values are determined using standard formulas: MD = (λ1 + λ2 + λ3) / 3, and FA = [3(△λ1 2 +△λ2 2 +△λ3 2 ) / 2(λ1 2 +λ2 2 +λ32 )] 1 / 2 , where λ1, λ2, and λ3 represent the first, second, and third eigenvalues of the diffusion tensor, respectively, and Δλ1, Δλ2, and Δλ3 represent λ1-MD, λ2-MD, and λ3-MD, respectively.

[0189] 4.4 Spatial normalization to MNI space: Using a variant of the large deformation diffeomorphic metric mapping (LDDMM) algorithm (Hsu et al., 2012; Hsu et al., 2015; Beg et al., 2005; Miller et al., 2006), the TPMs of GM and WM segmented from the T1w image are registered to the ICBM152 template (defined in MNI space). Specifically, the velocity is iteratively estimated by emitting an initial velocity located in the ICBM152 space along the time dimension towards the native space of the individual. At convergence, the associated deformation map can transform the individual's TPM to match the ICBM152 TPM template. In the registration, the process is divided into 10 evenly spaced intervals, and an isotropic Gaussian filter with a full width at half maximum (FWHM) of 10 mm is used to ensure the smoothness of the initial velocity.

[0190] 4.5 Diffusion index extraction: The FA and MD maps in the native space are normalized to MNI space through the deformation map derived from the LDDMM registration. Masks created from the WM TPM of the ICBM152 template are used to extract the FA and MD values in the WM pixels. Since FA and MD are DTI indices representing white matter microstructural properties, these values are used as features for dMRI-brain age modeling.

[0191] All image processing procedures are performed using in-house programs in MATLAB (MathWorks, Natick, Massachusetts, USA), except for brain segmentation on T1w images using SPM12.

[0192] 5. Grouping according to CDR tracking records

[0193] According to the tracking records of the CDR values of the participants, the participants recruited in this study are divided into 9 groups. The 9 groups are named as follows: [CN]-modeling, [CN1]-to-CN2, CN1-to-[CN2], [CN]-to-D1, D1-to-[CN], [D1]-to-CN, [nD1], [D1]-to-D2, and [nD2]. All groups are mutually exclusive, except for [CN1]-to-CN2 and CN1-to-[CN2], which recruit the same participants for the first and second MRI scans, respectively. The characteristics of the 9 groups are described below. 475 individuals are included in the analysis after grouping: 258 participants are in the CN-modeling group, and 217 participants are in the remaining groups.

[0194] 5.1[CN]-Modeling

[0195] This group included participants with normal cognition (CDR=0, referred to as CN) and stability. CDR was 0 in all clinical records and had at least one such record within 1 year before or after the MRI scan. The data in this group was used as training data to establish the dMRI-brain age model. [CN] represents the cognitive status of CN around the time of the MR scan.

[0196] 5.2[CN1]-to-CN2

[0197] This group included 46 cognitively normal and stable participants. The screening criteria were the same as for the [CN]-modeling group. The data in this group were used as test data for dMRI-brain age modeling. [CN1] indicates that the participant was CN around the time of the first MR scan.

[0198] 5.3CN1-to-[CN2]

[0199] The participants in this group (N=46) were the same as those in [CN1]-to-CN2. As mentioned above, the MRI scan date in this group was later than the MRI scan date in [CN1]-to-CN2, with an interval of 2.91±0.67 years. [CN2] indicates that the participant was CN around the time of the second MR scan.

[0200] 5.4[CN]-to-D1

[0201] Participants in this group (N=34) were cognitively normal within an interval of ±180 days from the MRI scan date, but converted to very mild symptomatic AD (CDR=0.5, referred to as D1) after this interval.

[0202] 5.5D1-to-[CN]

[0203] Participants in this group (N=25) were cognitively normal within an interval of ±180 days from the MRI scan date, but were at D1 before that interval.

[0204] 5.6[D1]-to-CN

[0205] Participants in this group (N=26) were in D1 within an interval of ±180 days from the MRI scan date, but transitioned to CN after this interval. [D1] indicates cognitive status at D1 around the time of the MRI scan.

[0206] 5.7[nD1]

[0207] Participants in this group (N = 34) were in D1 within the interval of ±180 days from the MRI scan date. After this interval, 19 participants had at least one clinical record showing that they remained in the D1 stage, while the remaining participants (N = 15) had no available records of CDR. Therefore, this group was named nominal D1 (abbreviated as nD1). [nD1] indicates that the participants were in nominal D1 near the MRI scan time.

[0208] 5.8[D1]-to-D2

[0209] Participants in this group (N = 28) were in D1 within the interval of ±180 days from the MRI scan date, but changed to mild symptomatic AD (CDR = 1, abbreviated as D2) after this interval.

[0210] 5.9[nD2]

[0211] Participants in this group (N = 24) were in D2 within the interval of ±180 days from the MRI scan date. After this interval, 13 participants changed to moderate symptomatic AD (CDR = 2, abbreviated as D3), 6 participants had at least one clinical record showing that they remained unchanged, and 5 participants had no CDR records. Therefore, this group was named nominal D2 (abbreviated as nD2). [nD2] indicates that the participants were in nominal D2 near the MRI scan time.

[0212] Table 1 lists the demographic data of the people in the analysis, and Table 2 summarizes the statistical results of pairwise comparisons in the seven groups.

[0213] Table 1: Demographic Data

[0214]

[0215] MMSE: Mini-Mental State Examination; CDR: Clinical Dementia Rating; CN: CDR = 0 (cognitively normal); D1: CDR = 0.5; D2: CDR = 1; PAD: Predicted Age Difference.

[0216] The time elapsed here represents the scan interval between [CN1]-to-CN2 and CN1-to-[CN2].

[0217] APOE e4 refers to the percentage of the group population with one or two e4 alleles.

[0218] Table 2: Adjusted p-values for pairwise comparisons (two-sample t-test; adjusted for multiple comparisons using the Benjamini-Hochberg method)

[0219]

[0220]

[0221] CDR: Clinical Dementia Rating; CN: CDR = 0 (cognitively normal); D1: CDR = 0.5; D2: CDR = 1; PAD: Predicted Age Difference; FDR: False Discovery Rate.

[0222] 6. dMRI - Brain Age

[0223] Data in the CN - modeling group were used to train the dMRI - brain age model. The Gaussian process regression method was used to regress the chronological age of participants against FA and MD values within the WM region. After the training process, the model was applied to the individual MRI data in other groups to estimate the dMRI - brain age. The Predicted Age Difference (PAD) was calculated by subtracting the chronological age from the dMRI - brain age. The performance of the model was tested in the CN1 group and quantified by the mean absolute error (MAE) and the Pearson correlation coefficient (r) between the dMRI - brain age and the chronological age.

[0224] 7. Statistical Inference

[0225] The PAD values were compared between any two of the seven groups (i.e., CN1 - to - [CN2], [CN] - to - D1, D1 - to - [CN], [D1] - to - CN, [nD1], [D1] - to - D2, and [nD2]); a total of 21 pairwise comparisons. For each pair of groups, a two - sample t - test was performed. The Benjamini - Hochberg procedure with a false discovery rate (FDR) of 0.05 was used to determine whether the test was considered statistically significant.

[0226] Results

[0227] i. Comparison between different CDR levels

[0228] Figure 10 Groups with different CDR levels are shown, i.e., CN1 - to - [CN2], [nD1], and [nD2]. The means and standard deviations of the PAD values for CN1 - to - [CN2], [nD1], and [nD2] were - 0.75 ± 5.53, 5.48 ± 6.68, and 12.33 ± 10.89 years, respectively. All comparisons for the three pairs of groups showed statistically significant differences (Table 2, #4, #6, #20). The results indicate that the higher the CDR level, the higher the PAD value.

[0229] ii. Comparison within participants with a baseline CDR of 0.5

[0230] Figure 11Shows participants with a CDR of 0.5 at the time of MRI scan, who reverted to CN ([D1]-to-CN), remained relatively stable ([nD1]), or converted to D2 ([D1]-to-D2) within 2 to 3 years ( Figure 11 ). The mean and standard deviation of the PAD values for [D1]-to-CN, [nD1], and [D1]-to-D2 were 2.99 ± 7.16, 5.48 ± 6.68, and 9.21 ± 6.14 years, respectively. The PAD value in [D1]-to-D2 was significantly higher than those in [D1]-to-CN (adjusted p = 0.0031) and [nD1] (adjusted p = 0.0406) (Table 2, #17, #19). The PAD value in [D1]-to-CN was lower than that in [nD1], but the comparison between the two groups did not reach statistical significance (adjusted p = 0.2227) (Table 2, #16).

[0231] iii. Comparison within participants with a baseline CDR of 0

[0232] Figure 12 Shows participants who were in CN at the time of MRI scan but showed a relatively stable state (CN1-to-[CN2]) or an unstable state ([CN]-to-D1 and D1-to-[CN]) within 1 to 2 years. The mean and standard deviation of the PAD values for CN1-to-[CN2], [CN]-to-D1, and D1-to-[CN] were -0.75 ± 5.53, 3.10 ± 7.72, and 2.90 ± 7.19 years, respectively. The PAD values for the two groups [CN]-to-D1 (adjusted p = 0.0213) and D1-to-[CN] (adjusted p = 0.0318) were significantly higher than those in the CN1-to-[CN2] group (Table 2, #1, #2). There was no significant difference between the [CN]-to-D1 group and the D1-to-[CN] group (adjusted p = 1.0000) (Table 2, #7).

[0233] iv. PAD spectra

[0234] Figure 13 Shows seven groups sorted according to the mean PAD value, and we can observe the following five subgroups. The first subgroup includes participants with relatively stable normal cognition (CN1-to-[CN2]) and an average PAD of approximately 0 years. The second subgroup includes participants showing transitions between CN and D1 (i.e., [CN]-to-D1, D1-to-[CN], and [D1]-to-CN), with an average PAD of approximately 3 years. The third, fourth, and fifth subgroups are participants in [nD1], [D1]-to-D2, and [nD2], with average PADs of 5.48, 9.21, and 12.33 years, respectively.

[0235] In summary, for the elderly with CDR of 0, 0.5, and 1, the PAD derived from DTI corresponds to the CDR score, and the PAD differs between those with relatively stable CDR and those whose CDR becomes a higher score over several years. The results generated by the present invention indicate that PAD is necessary for grading dementia severity and predicting changes in severity over several years. The ability of PAD can address the current limitations of CDR, which is subject to inaccurate information obtained from unreliable informants and cannot predict CDR changes in the next few years. Therefore, the solution of the present invention can be an alternative marker for incident CDR and, in addition, a predictive marker for CDR changes from baseline over 2 to 3 years.

[0236] Association of PAD with mild to moderate dementia severity

[0237] In the present invention, it was demonstrated that PAD derived from white matter microstructural properties as indicated by FA and MD of DTI is associated with CDR of 0, 0.5, and 1( Figure 10 ).

[0238] PAD in patients with a baseline CDR of 0.5 is associated with CDR changes within 1 to 2 years

[0239] Among patients with a CDR of 0.5, they presented different outcomes in approximately 1 to 2 years( Figure 11 ). The present invention demonstrated that the PAD in these patients was significantly different. The average PAD of patients whose CDR changed from 0.5 to 1 in about 2 years was 9.21 years, significantly higher than that of patients with a constant CDR of 0.5 (average PAD = 5.48 years, adjusted p = 0.0406) or patients with a baseline CDR of 0.5 but whose CDR changed to 0 in about 1.5 years (average PAD = 2.99 years, adjusted p = 0.0031). The results are clinically significant. Currently, exercise and cognitive training are recommended interventions for MCI patients with a CDR mainly of 0.5 (Petersen et al., 2018). However, there is no information available to predict cognitive decline in individual patients, and thus, interventions cannot be further customized for individual patients according to the risk of decline. The findings of the present invention imply that PAD can be used to classify MCI patients into a high-risk group (CDR changes from 0.5 to 1) and a low-risk group (constant CDR of 0.5 or CDR changes from 0.5 to 0). Appropriate interventions can then be designed for different risk groups.

[0240] PAD in patients with a baseline CDR of 0 is associated with CDR changes within 2 to 3 years

[0241] The present invention also shows that participants with a constant CDR of 0 and those with a baseline CDR of 0 but who became 0.5 in about 2.5 years have significantly different PADs (average PAD: -0.75 years vs. 3.10 years, adjusted p = 0.0213; Figure 12 ). The results indicate that there are changes in PAD even among cognitively normal individuals. Multiple brain age studies in cognitively normal individuals have demonstrated an association between PAD and lifestyle factors such as physical exercise, alcohol and tobacco consumption, and social relationships, as well as cardiovascular risks such as blood pressure, diabetes, and body mass index (Cole, 2020; Hatton et al., 2018; Kolbeinsson et al., 2020; Ning et al., 2020; Ronan et al., 2016; Stefiner et al., 2016). These studies suggest that to some extent, changes in PAD are caused by health-related risk factors of individual individuals, and the more risk factors, the higher the PAD. Moreover, many epidemiological studies have shown that poor health-related risk factors have a higher risk of being troubled by dementia in later life (Akbaraly et al., 2019; Cations et al., 2016; Fayosse et al., 2020; Kivipelto et al., 2006; Mukadam et al., 2019; WHO, 2019). In summary, previous studies have reported the association between PAD and lifestyle risk factors and the association between lifestyle factors and dementia. This study further shows that for cognitively healthy individuals over 60 years old (approximate mean - SD = 69.10 - 7.86 years in Table 1, CN1 - to - [CN2]), those with a PAD higher than 10 years (approximate mean + SD = 3.10 + 7.72 years in Table 2, [CN] - to - D1) may be more likely to become cognitively impaired (CDR = 0.5) within 3 years than those with a PAD lower than -6 years (approximate mean - SD = -0.75 - 5.53 years in Table 2, CN1 - to - [CN2]).

[0242] Correspondence between PAD and stable and metastable CDR states

[0243] The present invention finds that some groups in the study population showed relatively stable CDR scores within 2 to 3 years, namely CN1 - to - [CN2], [nD1], and [nD2], while some groups showed relatively rapid CDR transitions, namely [CN] - to - D1, D1 - to - [CN], [D1] - to - CN, and [D1] - to - D2 (Table 2). Notably, when each group is marked with PAD, spanning relatively stable and metastable groups reveals a series of continuous PADs ( Figure 13)。The PAD was the smallest in the cognitively normal group (CN, -0.75 years), followed by the three metastable groups ([CN]-to-D1, D1-to-[CN], and [D1]-to-CN, presenting 3.10 years, 2.90 years, and 2.99 years, respectively), the relatively stable group with CDR = 0.5 ([nD1], 5.48 years), another metastable group ([D1]-to-D2, 9.21 years), and was the largest in the relatively stable group with CDR = 1 or more ([nD2], 12.33 years). The PAD spectra corresponding to the stable and metastable states of CDR provided new perspectives for the CDR of the elderly. Given the CDR scores of 0, 0.5, or 1 for an individual, we can use PAD to predict the probability of remaining cognitively normal (CDR = 0) or transitioning to different CDR stages (CDR = 0.5 or 1) within several years.

[0244] Example 3

[0245] Another example of the present invention provides proof of concept to support the proposed intended use of the device. A computer-based program reads the specified brain image data of an individual as input and estimates the PAD of the person, and based on the predefined cut-off value of the PAD, the program determines whether the person is likely / unlikely to be cognitively normal when he / she passes the CDR assessment.

[0246] The present invention performs a retrospective study on the cohort of the OASIS-3 database, which evaluates the performance of PAD in identifying patients who are likely / unlikely to be cognitively normal (CDR = 0) by comparing the results with the CDR scores of the patients.

[0247] Positive predictive value (PPV), negative predictive value (NPV), sensitivity, specificity, odds ratio (OR), and the accuracy of discrimination between CDR>0 and CDR = 0 were obtained at different cut-off values of PAD.

[0248] Data Selection

[0249] Similar to Example 3, the participants in OASIS-3 of Example 3 underwent one-phase to multi-phase brain MRI scans on three 3T MRI scanners (Siemens in Erlangen, Germany), and the three scanners included two scanner models, namely Biograph and TIM Trio. To reduce the variability in terms of scanners, the same acquisition protocol was used for T1-weighted (T1w) imaging and diffusion tensor imaging (DTI). The MRI data acquired from the same scanner model (TIM Trio) were selected for analysis. Since the analysis required acceptable image quality of T1w and DTI data, participants with images presenting severe image artifacts (such as motion blur on T1w or failure to pass artifact correction on DTI) were excluded. A total of 529 participants were recruited for subsequent analysis meeting the selection criteria out of 575 MRI scans.

[0250] Grouping According to CDR

[0251] According to the tracking records of the CDR values of the participants, the participants enrolled in this study were divided into 5 groups. The 5 groups are named as follows: [CN] - Modeling, [CN1] - to - [CN2], CN1 - to - [CN2], [nD1], and [nD2]. All groups are mutually exclusive, except for [CN1] - to - [CN2] and CN1 - to - [CN2], which recruited the same participants for the first and second MRI scans, respectively. The characteristics of the 5 groups are described below.

[0252] i) [CN] - Modeling: This group included 258 cognitively normal (CDR = 0, abbreviated as CN) and stable participants. The CDR was 0 in all clinical records and had at least one such record within 1 year before or after the MRI scan. The data in this group were used as training data to establish the dMRI - brain age model. [CN] represents the cognitive status of CN near the MR scan time.

[0253] ii) [CN1] - to - CN2: This group included 46 cognitively normal and stable participants. The screening criteria were the same as those for the [CN] - Modeling group. The data in this group were used as test data for dMRI - brain age modeling. [CN1] represents that the participants were CN near the time of the first MR scan.

[0254] iii) CN1 - to - [CN2]: The participants in this group (N = 46) were the same as those in [CN1] - to - [CN2]. As described above, the MRI scan date in this group was later than that in [CN1] - to - [CN2], and the scan interval was 2.91 ± 0.67 years. [CN2] represents that the participants were CN near the time of the second MR scan.

[0255] iv) [nD1]: The participants in this group (N = 34) were in D1 within the interval of ±180 days from the MRI scan date. After this interval, 19 participants had at least one clinical record showing that they remained in the D1 stage, while the remaining participants (N = 15) had no available records of CDR. Therefore, this group was named nominal D1 (abbreviated as nD1). [nD1] represents that the participants were in nominal D1 near the MRI scan time.

[0256] v) [nD2]:Participants in this group (N = 24) were in D2 within the interval of ±180 days from the MRI scan date. After this interval, 13 participants transitioned to moderately symptomatic AD (CDR = 2, abbreviated as D3), 6 participants had at least one clinical record showing that they remained unchanged, and 5 participants had no CDR record. Therefore, this group was named nominal D2 (abbreviated as nD2). [nD2] indicates that the participants were in nominal D2 near the MRI scan time.

[0257] Analysis and Statistics

[0258] Brain age modeling: The data in the [CN]-modeling group were used to train the dMRI-brain age model. The Gaussian process regression method was used to regress the chronological age of the participants against the DTI indices of the brain. After the training process, the model was applied to the individual MRI data in other groups to estimate the dMRI-brain age. The predicted age difference (PAD) was calculated by subtracting the chronological age from the dMRI-brain age. The performance of the model was tested in the [CN1]-to-CN2 groups, and this performance was quantified by the mean absolute error (MAE) and the Pearson correlation coefficient (r) between the dMRI-brain age and the chronological age.

[0259] Group comparison: The PAD values of any two of the three groups (i.e., CN1-to-[CN2], [nD1], and [nD2]) were compared, for a total of 3 pairwise comparisons. For each pair of groups, a two-sample t-test was performed. The Benjamini-Hochberg procedure with a false discovery rate (FDR) of 0.05 was used to determine whether the test was considered statistically significant.

[0260] Performance for intended use: To evaluate the performance of PAD in excluding cognitively normal conditions, we combined the [nD1] and [nD2] groups to form a new group (named [nD1]+[nD2]) to represent the group with CDR > 0, which would be excluded by our intended use. Receiver operating characteristic (ROC) curve analysis was performed to evaluate the accuracy of PAD in distinguishing between CDR = 0 and CDR > 0. The positive predictive value (PPV), negative predictive value (NPV), sensitivity, specificity, and odds ratio (OR) for the distinction between CDR > 0 and CDR = 0 were evaluated at different cut-off values of PAD.

[0261] Results

[0262] Demographics: Table 3 summarizes the demographics of 5 groups, including [CN]-modeling (training data), [CN1]-to-CN2 (test data), and 3 comparison groups, namely CN1-to-[CN2], [nD1], and [nD2]. Qualitatively, we can see the following trends: from CN1-to-[CN2], [nD1] to [nD2], CDR and PAD increase, and MMSE decreases.

[0263] Table 3: Demographics

[0264]

[0265]

[0266] Performance of brain age prediction: The model performance tested in the [CN1]-to-CN2 group (test data) gives MAE = 4.35 ± 3.39 years and r = 0.77 (p-value = 3.64×10-10).

[0267] Group comparison: Figure 21 Groups with different CDR levels are shown, namely CN1-to-[CN2], [nD1], and [nD2]. The mean and standard deviation of the PAD values for CN1-to-[CN2], [nD1], and [nD2] are -0.75 ± 5.53, 5.48 ± 6.68, and 12.33 ± 10.89 years, respectively. All comparisons of the three group pairs show statistically significant differences. The results indicate that the higher the CDR level, the higher the PAD value.

[0268] ROC curve analysis: Figure 22 The results of the ROC curve analysis are shown. The area under the curve (AUC) is 0.82. Table 4 lists the discrimination performance between CDR = 0 and CDR > 0 at different cut-off values of PAD. The optimal cut-off value of PAD is found to be 3 years. Using PAD = 3 as the cut-off value, the discrimination performance between CDR = 0 and CDR > 0 shows sensitivity = 0.79, specificity = 0.74, PPV = 0.79, NPV = 0.74, OR = 10.86, and accuracy = 0.77.

[0269] Table 4: Discrimination performance between CDR = 0 and CDR > 0 at different cut-off values of PAD.

[0270]

[0271] Abbreviations: PAD = predicted age difference, TP = true positive, FP = false positive, FN = false negative, TN = true negative, OR = odds ratio, ACC = accuracy, PPV = positive predictive value, NPV = negative predictive value, Sens. = sensitivity, Spec. = specificity, AUC = area under the curve.

[0272] Conclusion

[0273] The present invention shows that the optimal cut-off value for differentiating PAD with CDR = 0 and CDR>0 is 3 years, resulting in a sensitivity = 0.79, specificity = 0.74, PPV = 0.79, NPV = 0.74, OR = 10.86 and accuracy = 0.77. However, the intended use is to exclude CDR = 0 in patients suspected of having dementia. In this intended use, the performance of PPV, specificity and OR rather than sensitivity and accuracy become the primary endpoints of concern. In this case, PAD = 6 years seems to be the optimal cut-off value, giving a PPV = 0.89, specificity = 0.91 and OR = 13.86. Figure 23 The confusion matrix showing the results using a cut-off PAD of 6 and the distribution of the two groups (i.e., CDR = 0 and CDR>0) with respect to PAD is shown. The results imply that for participants with PAD >= 6 years, 89% of them are not cognitively normal (i.e., CDR>0), and for cognitively normal (i.e., CDR = 0) participants, 91% of them have a PAD < 6 years. The OR of 13.86 defined as (TP×TN) / (FP×FN) implies that the ratio of true results (TP and TN) to false results (FP and FN) is greater than 10-fold. On the other hand, using PAD = 6 years as the cut-off value, the sensitivity = 0.57 and NPV = 0.63. The results imply that for participants who are not cognitively normal (i.e., CDR>0), 57% of them have a PAD >= 6 years, while for participants with a PAD < 6 years, 63% are cognitively normal (i.e., CDR = 0). In summary, using PAD = 6 years as the cut-off value has satisfactory performance in identifying patients who are unlikely to be cognitively normal, but it performs poorly in identifying cognitively normal patients. Therefore, we conclude that the satisfactory performance of PPV and OR supports the intended use of the device we proposed.

[0274] Importantly, it should be noted that the present invention is a pioneer in the field of CDR assessment and prediction in at least three aspects. First, the brain age model established by the present invention is based on white matter microstructure metrics, named dMRI-brain age and the resulting WMPAD. Second, what WMPAD does is to predict the concomitant measure of human CDR. Third, additionally, WM PAD also predicts the CDR change of the same person from 2 to 3 years from the time when the dMRI-brain age is measured.

[0275] In contrast to the present invention, existing methods for estimating CDR either establish a brain age model based on morphometric features of gray matter obtained from T1M MRI (termed GM brain age) and the resulting GM PAD, and find an association between GM PAD and a comorbid measure of CDR (Beheshti et al., 2018); or calculate the FW content from diffusion MRI and find an association between FW and comorbid measures of CDR and future states of CDR (Maillard et al., 2019).

[0276] Other research and inventions in the art have examined in detail the association between MRI metrics and the entire array of cognitive performance measured by validation tools. Among these results, specific MRI metrics associated with specific cognitive measures have been reported. However, the present invention establishes a unique correspondence between WM PAD and comorbid CDR, as well as a correspondence between WM PAD and changes in CDR 2 to 3 years after baseline. This specific correspondence between dMRI-brain age and CDR has not been reported previously and is thus not obvious to contemporary inventors.

[0277] WM PAD has advantages over existing solutions because (1) it is more sensitive to CDR than GM brain age, and (2) it is less sensitive to cerebrospinal fluid (CSF) partial volume effects than the FW metric.

[0278] First, WM PAD is more sensitive than GM brain age in differentiating CDR scores. In the paper by Beheshti et al., they only showed the correlation between GM brain age and CDR scores (from 0, 0.5 to 1). They did not show the comparison results between brain age and CDR, probably due to invalid results caused by significant overlap between CDR scores (in Beheshti's paper Figure 7 b).

[0279] On the other hand, compared to the FW metric, dMRI-brain age may be less sensitive to CSF partial volume effects caused by brain atrophy. Hypothetically, the FW content is mainly attributed to CSF partial volume effects in white matter adjacent to the ventricles or sulci. The FW content will likely increase as the brain atrophies with age. In other words, the FW content is confounded by brain atrophy. In contrast, WM PAD is derived from FA and MD in whole-brain white matter. Although FA and MD values may change in white matter pixels adjacent to the ventricles and sulci, this effect is negligible because the number of pixels with partial volume effects is quite small compared to the whole-brain white matter pixels.

[0280] The above description of the exemplary embodiments of the present invention is presented only for purposes of illustration and description and is not intended to be exhaustive or to limit the invention to the precise form disclosed. Many modifications and variations are possible in light of the above teachings.

[0281] Although several alternative embodiments of the present invention have been shown, it should be understood that certain changes known to those skilled in the art can be made without departing from the basic scope of the present invention as discussed and set forth above and below (including the claims and the drawings). In addition, the above embodiments and the claims set forth below are only intended to illustrate the principles of the present invention and are not intended to limit the scope of the present invention to the disclosed elements.

[0282] References cited in this application (which may include patents, patent applications, and various publications) are cited and discussed in the specification of the present invention. The citation and / or discussion of these references are provided solely to clarify the description of the present invention and not to admit that any of these references are "prior art" of the invention described herein. All references cited and discussed in the specification of the present invention are hereby incorporated by reference in their entirety and to the same extent as if each reference had been incorporated by reference individually.

Claims

1. A method for evaluating the Clinical Dementia Rating (CDR) and future changes of CDR of an individual based on the predicted age difference (PAD) according to the individual's brain MRI data, comprising: (1) Scanning the brain of the individual using a scanning device to obtain MRI data of the individual's brain using T1-weighted (T1w) imaging and diffusion tensor imaging (DTI) sequences; (2) Processing the obtained T1w data to obtain white matter (WM) and gray matter (GM) image segments; (3) Registering the WM and GM image segments obtained in step (2) to the MNI space to obtain a deformation map of the individual's brain; (4) Correcting the artifacts in the DTI data; (5) Obtaining the fractional anisotropy (FA) and mean diffusivity (MD) maps of the individual in the native space based on the artifact-corrected DTI data; (6) Using the FA and MD maps obtained in step (5) and the deformation map obtained in step (3) to extract the FA and MD values of the WM region of the individual's brain; (7) Calculating the WM PAD by inputting the extracted FA and MD of the WM into a established dMRI-brain age model; (8) Predicting the CDR and CDR changes of the individual according to the WM PAD using a CDR prediction model and a CDR change prediction model; (9) Outputting the status of cognitive impairment and its future changes at an output terminal; wherein step (2) includes: (i) correcting the SI inhomogeneity of the T1w data; (ii) segmenting the tissue probability map (TPM); (iii) obtaining the WM and GM segments.

2. The method for evaluating the Clinical Dementia Rating (CDR) and future changes of CDR of an individual based on the predicted age difference (PAD) according to the individual's brain MRI data according to claim 1, wherein, step (5) includes: (i) Estimating the diffusion tensor at each image pixel of the artifact-corrected DTI data; (ii) Calculating the FA and MD at each pixel using the diffusion tensor indices derived from the estimated diffusion tensor at each pixel; (iii) Generating the FA and MD maps in the native space using the FA and MD at each pixel.

3. The method for evaluating the Clinical Dementia Rating (CDR) and future changes of CDR of an individual based on the predicted age difference (PAD) according to the individual's brain MRI data according to claim 1, wherein, step (6) includes: (i) Registering the FA and MD maps to the MNI space using the deformation map in step (3); and (ii) Masking the registered FA and MD maps with the WM segments.

4. The method for evaluating the Clinical Dementia Rating (CDR) and future changes of CDR of an individual based on the predicted age difference (PAD) according to the individual's brain MRI data according to claim 1, wherein, the dMRI-brain age model is established by regressing the chronological age of more than one individual for the FA and MD values in the WM region using the Gaussian process regression method.

5. A method for evaluating the Clinical Dementia Rating (CDR) and future changes in CDR of an individual based on predicted age differences (PAD) according to the individual's brain MRI data as claimed in claim 1, wherein, the CDR change is the change in CDR of the individual within the next 2 - 3 years since the MRI scan date.

Images