US Patent for Kidney glomeruli measurement systems and methods Patent (Patent # 10,045,728 issued August 14, 2018) (2024)

CROSS-REFERENCE TO RELATED APPLICATIONS

This application is a continuation of PCT Application No. PCT/US2014/059545 filed on Oct. 7, 2014 and entitled “KIDNEY GLOMERULI MEASUREMENT SYSTEMS AND METHODS”. PCT Application No. PCT/US2014/059545 claims priority to, and the benefit of, U.S. Provisional Application Ser. No. 61/887,668 filed on Oct. 7, 2013 and entitled “KIDNEY GLOMERULI MEASUREMENT SYSTEMS AND METHODS”. Each of the above applications is hereby incorporated by reference in their entirety.

FEDERALLY SPONSORED RESEARCH OR DEVELOPMENT

This invention was made with government support under grant number R21-DK091722 awarded by the National Institute of Health. The government has certain rights in the invention.

TECHNICAL FIELD

The present disclosure relates to image processing, and specifically to systems and methods to measure glomeruli in the kidneys.

BACKGROUND

The glomeruli of the kidney perform the key role of blood filtration. Each functioning glomerulus consists of a tuft of capillaries and several types of cells, and is surrounded by a Bowman's capsule forming the corpuscle. The number of glomeruli in a kidney is correlated with susceptibility to chronic kidney and cardiovascular disease, driving interest in technology to measure glomerular morphology. Cationic ferritin nanoparticles have been used to target, image, and count individual glomeruli in the whole kidney with magnetic resonance imaging (MRI). Accumulated nanoparticles create spots in MR images at each glomerulus. However, previous techniques have been unable to perform fast, reliable measurements of glomeruli in the image. Accordingly, improved techniques are desirable.

DETAILED DESCRIPTION

The following description is of various exemplary embodiments only, and is not intended to limit the scope, applicability or configuration of the present disclosure in any way. Rather, the following description is intended to provide a convenient illustration for implementing various embodiments including the best mode. As will become apparent, various changes may be made in the function and arrangement of the elements described in these embodiments without departing from the scope of the present disclosure.

Prior techniques were unable to perform fast, reliable measurements of kidney glomeruli in a 3 dimensional (3D) image. Traditionally, glomeruli, and thereby nephrons, have been counted using stereological or morphometric techniques applied to histological sections. The gold standard method is known as the dissector/fractionator stereological technique. While this technique provides good estimates of total glomerular number, it is limited to studies of kidneys obtained at autopsy, as representative 3D samples from the kidney are required. Recently, superparamagnetic, cationic ferritin (CF) nanoparticles have been developed as contrast agents for magnetic resonance imaging (MRI). CF binds to anionic proteoglycans in the glomerular basem*nt membrane after intravenous injection, and the accumulation of CF may be detected with T2*-weighted MRI in 3D. This technique can be used to detect, measure, and count every glomerulus in the whole kidney ex vivo and in vivo. While exciting, the widespread use of this technique for preclinical and clinical studies is compromised by the lack of image processing tools to reliably and accurately segment glomeruli in the magnetic resonance (MR) images. In accordance with principles of the present disclosure, a computational pipeline to process 2D and/or 3D MRI images to accurately, and robustly count glomeruli in whole kidney in vivo is presented.

In accordance with principles of the present disclosure, in various exemplary embodiments an image processing system for measuring glomeruli in a kidney comprises a suitable computing device configured for image processing. In an exemplary embodiment, an image processing system comprises a high-performance workstation (e.g., configured with a multi-core processor such as a six-core XEON brand processor supplied by Intel Corporation) having a suitable amount of RAM (e.g., 4 GB or more) and a suitable graphics card (e.g., a workstation-class graphics card supplied by NVIDIA Corporation, Advanced Micro Devices, or similar).

In an exemplary embodiment, with reference now to FIG. 1, a method 100 of measuring glomeruli in a kidney comprises identification of a candidate glomerular region from a 3 dimensional (3D) image of a kidney, segmentation of individual glomeruli in the 3D image of a kidney via a hessian approach (step 110), extraction of features of the individual glomeruli (step 120), and development of a model to identify the individual glomeruli in the 3D image (step 130). In accordance with an exemplary embodiment, this method may be used to detect, measure and count every glomerulus in a whole kidney ex vivo and in vivo. The features utilized in step 120 may include, but are not limited to, average intensity, divergence, distance to kidney boundary, region volume, shape index, and/or Laplacian of Gaussian (LoG).

In accordance with principles of the present disclosure, a 3 dimensional (3D) image may be generated using superparamagnetic cationic ferritin (CF) nanoparticles or other suitable method. In accordance with an exemplary embodiment, cationic ferritin nanoparticles may be used as contrast agents for MRI. CF nanaoparticles bind to anionic proteoglycans in the glomerular basem*nt membrane when injected intravenously and the accumulation of CF may be detected with T2*-weighted MRI in a 3D image.

In accordance with principles of the present disclosure, a candidate glomerular region may be identified via a suitable method, for example via a Hessian matrix-based approach. In accordance with an exemplary embodiment, a convexity property from a Hessian Matrix may be used to identify candidate glomerular regions from the whole kidney image. The convexity property may be further used to separate touching glomeruli. In an exemplary embodiment, use of the convexity property may reduce data size, allowing diverse features from each candidate glomerular object to be explored for accurate segmentation.

In various exemplary embodiments, segmentation of glomeruli may be the separation of touching glomeruli. Touching glomeruli may be glomeruli that intersect with each other in an image, forming a clump. A Hessian matrix may be applied to split the touching glomeruli based on the convexity property. The segmentation of glomeruli may allow for each individual glomeruli to be counted, providing an accurate count of glomeruli in a kidney.

In various exemplary embodiments, extraction of features of the individual glomeruli may be used. Exemplary features include, but are not limited to, one or more of average intensity A_T, regional blobness R_T, regional flatness S_T, average intensity (AI), divergence (Div), distance to kidney boundary (Dis), region volume(s) (Vol), shape index (SI), and Laplacian of Gaussian (LoG). In various embodiments, one or more features may be derived from domain knowledge or imaging geometrics. The application of these features may contribute to the segmentation of the glomeruli.

In accordance with principles of the present disclosure, the development of a model to identify the individual glomeruli in the 3D image may be used. In accordance with an exemplary embodiment, a model may be a Variational Bayesian Gaussian Mixture Model (VBGMM). While it will be appreciated that other models may be utilized, VBGMM may advantageously be utilized due to computational efficiency and accuracy.

In certain exemplary embodiments, a method of detecting kidney disease comprises obtaining a 3 dimensional (3D) image of a kidney, identification of a candidate glomerular region from the 3D image, segmentation of individual glomeruli in the 3D image of a kidney, and extraction of features of the individual glomeruli and development of a model to identify the individual glomeruli in the 3D image, wherein the number of glomeruli correlates with susceptibility to chronic kidney and cardiovascular disease.

It will be appreciated that Chronical Kidney Disease (CKD) develops when there are too few nephrons to maintain the homeostatic role in waste and fluid management. There are both congenital and acquired forms of nephropenia. The kidney's unique ability to adapt to fewer nephrons is reflected in the compensatory enlargement of the remaining glomeruli to maintain a constant filtration surface area and bulk glomerular filtration rate (GFR). Because of this compensatory mechanism, both endogenous and exogenous modalities that measure GFR are inadequate diagnostic tools for early detection of kidney diseases which involve changes in nephron number and glomerular volume. Studies have shown that even the mild stages of CKD are not benign and result in both higher risks of cardiovascular disease.

Exemplary results of method 100 are illustrated in FIGS. 2 through 7.

Turning now to FIGS. 8 through 12, various blob detectors have been developed over the years. These include interest point detectors such as Radial-Symmetry, SIFT, and SURF, and interest region detectors such as Harris-Affine detector, Hessian-affine detector and Hessian-Laplace detector. However, performance of these blob detectors for medical images (e.g., pathological and fluorescence images) are often unsatisfactory. Another type of detector, Laplacian of Gaussian (LoG) under the scale space theory has gained its popularity in the medical applications. In the scale space theory, one 2D image (or a slice of 3D image) is treated as a stack of images controlled by a scale parameter t. A multi-scale Gaussian scale space representation of the image is derived as the convolution of the raw image over the Gaussian kernel with respect to scale t to preserve the important spatial properties of the imaging structures. Given that as the scale parameter increases, the number of local minima in a dark blob does not increase, and the number of local maxima in a bright blob does not decrease, a diffusion process may apply to mark the identifiable blobs. For the individual blobs with similar sizes, one “optimal” scale often exists. Detectors generated via LoG kernels have been successfully applied for some blob detections. However, the symmetric nature of the LoG detector limits its applications to rotational asymmetric blobs.

This limitation has been attempted to be addressed by utilizing a generalized Laplacian of Gaussian (gLoG) in an attempt to detect rotational asymmetric structures by using different Gaussian kernels. The gLoG is thus able to detect general elliptical structures such as rotationally symmetric and asymmetric blobs. However, this approach is computationally intensive, and as such, these prior approaches are typically utilized only on 2D histopathological images and/or 2D microscopic images.

In contrast, principles of the present disclosure enable assessment, including blob detection, segmentation, and/or the like, on 3D images as well as 2D images, for example in order to segment many small structures. One challenge in 3D MR blob detection is that MR imaging is known to have acquisition noise, partial volume effect (the mixture of several tissue signals in a voxel) and bias field (spatial intensity inhom*ogeneity). Additionally, to enable high-throughput in vivo studies, a highly efficient detector is desirable. In addition, unlike the situation in histological sections, the glomeruli's small size corresponds to a high spatial frequency typically close to that of image noise. Thus, the segmentation is sensitive to local noises. For example, the average volume of rat glomeruli is approximately 6 to 7*10⁵ um³, which is less than 10 voxels with a resolution of 62×62×62 μm³ in a 3D MR image. Accordingly, exemplary methods disclosed herein provide a robust detector tolerable to image noises.

To address these and other outstanding challenges, a Hessian-based Difference of Gaussian (HDoG) detector is disclosed. The exemplary detector improves the detectability of glomeruli in 3D. HDoG utilizes a Difference of Gaussians (DoG) approach which has great computational advantages over LoG by using the approximation of LoG in the detection process. With reference now to FIG. 8, in an exemplary HDoG method 200, a DoG transformation of a raw image that enhances the blob structures and smoothes the local noises is first obtained (step 210). Hessian analysis is then applied on the transformed imaging matrix to pre-segment and delineate the blob (glomeruli) candidates (step 220). For each glomerular candidate identified by the Hessian analysis, exemplary regional features are derived (step 230), for example average intensity feature A_T, regional blobness R_T, and regional flatness S_T. These features are fed to an exemplary tuning-free unsupervised clustering algorithm, for example a Variational Bayesian Gaussian mixture model (VBGMM) for post pruning (step 240).

Exemplary method 200 may be validated, for example using precision, recall and F-score metrics. In an exemplary validation test of 15 pathological images and 200 2D fluorescence microscopy cell images, HDoG outperforms both LoG and gLoG detectors with less computing time. Additionally, results on human and rat kidney images confirm that HDoG is able to segment glomeruli automatically and accurately.

In various exemplary embodiments, utilization of method 200 provides a computationally efficient approach for identifying potential glomeruli. In an exemplary embodiment, method 200 was implemented on a Windows PC configured with an Intel Xeon processor operating at 2.0 Ghz and with 32 G of RAM. An exemplary set of 15 600×800 pathologic images were processed via method 200. The average operational time of method 200 was about 9.4 seconds per image. By way of comparison, an exemplary prior method, gLoG, exhibited an average operation time of about 30 seconds per image to process the same set of 15 images. Similarly, method 200 was utilized to process 200 256×256 fluorescence-light microscopy images of cells, and achieved an average operational time of about 1.0 seconds per image. The exemplary prior method, gLoG, exhibited an average operation time of about 10.0 seconds per image. Stated another way, method 200 offers many hundreds of percent improvements in computational efficiencies over prior widely-accepted methods.

In method 200, a glomerulus may be considered as a blob, i.e., a region that is darker than surrounding, and the convexity of intensity function within a blob considered to be consistent. Yet in reality, the convex property of intensity function within a blob region may have discontinuities due to the image noise. Accordingly, it is desirable to apply a smoothing process to filter the noise and make the blob region asymptotic convex (or concave). A DoG filter may desirably be utilized because (i) it can smooth the image noise by enhancing the objects at the selected scale, (ii) it is the approximation of a Laplacian of Gaussian (LoG) filter, which can highlight the blob structure, and (iii) it is computationally efficient while offering similar accuracy. These properties are desirable to detect/segment blobs on 3D and/or 2D images.

In method 200, let an 3D image be f: R³→R, the scale space representation L(x, y, z; t) at point (x, y, z) with scale parameter t is the convolution of image f (x, y, z) with Gaussian kernel (x, y, z; t):
L(x,y,z;t)=G(x,y,z;t)*f(x,y,z)  (Equation 1)

Where * is the convolution operator and

$G\left(x,y,z;t\right)=\frac{1}{{\left(2\pi t^2\right)}^{\frac{3}{2}}}\exp \left(-\frac{x^2+y^2+z^2}{2t^2}\right).$
The Laplacian of L(x, y, z; t) is:
∇²L(x,y,z;t)=L_xx+L_yy+L_zz  (Equation 2)

Since

$\frac{1}{2}{\nabla }^{2}L\left(x,y,z;t\right)={\partial }_{t}L\left(x,y,z;t\right),$
we have:

$\begin{array}{cc}{\nabla }^{2}L\left(x,y,z;t\right)={\lim }_{\delta t\to 0}\frac{L\left(x,y,z;t+\delta t\right)-L\left(x,y,z;t-\delta t\right)}{\delta t}\approx L\left(x,y,z;t+\delta t\right)-L\left(x,y,z;t-\delta t\right)& \left(\mathrm{Equation}3\right)\end{array}$

That is,
∇²L(x,y,z;t)≈f(x,y,z)*(G(x,y,z;t+δt)−G(x,y,z;t−δt))  (Equation 4)

To locate an optimal scale of blob, γ-normalization may be added to method 200 detector as a normalized LoG detector t^γ∇²L(x, y, z; t), thus an approximation of normalized LoG may be:
DoG_nor(x,y,z;t)=t^γ(x,y,z)*(G(x,y,z;t+δt)−G(x,y,z;t−δt))  (Equation 5)

In method 200, during a normalized DoG transformation, a dark glomerular blob is converted to a bright glomerular blob and vice versa. A blob after the normalized DoG operation as may be referred to herein as a transformed blob. Exemplary discussion herein focuses on a dark blob (i.e., a transformed bright blob), and it will be appreciated that a similar process may be utilized for a bright blob (i.e., a transformed dark blob).

In various exemplary embodiments, in method 200 the eigenvalues of the Hessian matrix of a blob-like structure can be used to describe the structure's geometry. After a target image is smoothed via DoG, then for any voxel (x, y, z) in the normalized DoG image DoG_nor(x, y, z; t) at scale t, the Hessian Matrix for this voxel is:

$\begin{array}{cc}H({\mathrm{DoG}}_{\mathrm{nor}}(x,y;t))=[\begin{array}{ccc}\frac{{\partial }^2{\mathrm{DoG}}_{\mathrm{nor}}\left(x,y,z;t\right)}{\partial x^2}& \frac{{\partial }^2{\mathrm{DoG}}_{\mathrm{nor}}\left(x,y,z;t\right)}{\partial x\partial y}& \frac{{\partial }^2{\mathrm{DoG}}_{\mathrm{nor}}\left(x,y,z;t\right)}{\partial x\partial z}\\ \frac{{\partial }^2{\mathrm{DoG}}_{\mathrm{nor}}\left(x,y,z;t\right)}{\partial x\partial y}& \frac{{\partial }^2{\mathrm{DoG}}_{\mathrm{nor}}\left(x,y,z;t\right)}{\partial y^2}& \frac{{\partial }^2{\mathrm{DoG}}_{\mathrm{nor}}\left(x,y,z;t\right)}{\partial y\partial z}\\ \frac{{\partial }^2{\mathrm{DoG}}_{\mathrm{nor}}\left(x,y,z;t\right)}{\partial x\partial z}& \frac{{\partial }^2{\mathrm{DoG}}_{\mathrm{nor}}\left(x,y,z;t\right)}{\partial y\partial z}& \frac{{\partial }^2{\mathrm{DoG}}_{\mathrm{nor}}\left(x,y,z;t\right)}{\partial z^2}\end{array}]& \left(\mathrm{Equation}6\right)\end{array}$

Since the transformed-bright blob is concave elliptic in shape (i.e., brightness is faded isotropically), every voxel within the blob is concave elliptic. Accordingly, method 200 may apply the proposition that in a transformed 3D normalized DoG image, every voxel of a transformed-bright blob has a negative definite Hessian. This is because, given geometric classification as a voxel and specific orientation patterns, if voxel (x, y, z) is concave elliptic, all of the eigenvalues λ₁, λ₂, λ₃ of H(x, y; t) are negative, meaning λ₁<0, λ₂<0 and λ₃<0. Since every voxel in the transformed-bright blob is concave elliptic, its eigenvalues are all negative, and thus the Hessian matrix of the voxel is negative definite.

If a voxel resides in a transformed-bright blob, the Hessian matrix of the voxel is negative definite. But a voxel having negative definite Hessian may not be from a transformed-bright blob. Therefore, method 200 may utilize the following definition: a blob candidate T in normalized DoG space is a 6-connected component of set U={(x, y, z)|(x, y, z)ϵ DoG_nor(x, y, z; t), I(x, y, z; t)=1}, where I(x, y, z; t) is the binary indicator such that if the voxel (x, y, z) has a negative definite Hessian then I(x, y, z; t)=1, otherwise I(x, y, z; t)=0.

In various exemplary embodiments, in method 200 definiteness of the Hessian can be assessed by the leading principal minors instead of calculating its eigenvalues of the matrix. Specifically, if D_k is the k-th leading principal minor of matrix M, we conclude it is negative definite if and only if (−1)^kD_k>0.

As illustrated in FIGS. 9 and 10, application of step 220 not only detects glomeruli candidates (usually, a set of all true glomeruli with some false identifications), but also delineates the boundary of each glomeruli candidate.

In prior approaches, classical geometric features in blob detection are R_B (the likelihood of blobness) and S_B (flatness—the second order structureness). These features are based on solving eigenvalues (assuming |λ₁|≤|λ₂|≤|λ₃|) of Hessian at each voxel. That is,

$\begin{array}{cc}{R}_B=\frac{\mathrm{Volume}\mathrm{of}\mathrm{Blob}\text{/}\left(4\pi \text{/}3\right)}{{\left(\mathrm{Largest}\mathrm{Cross}\mathrm{section}\text{/}\pi \right)}^{\text{3/2}}}=\frac{{\lambda }_1}{\sqrt{{\lambda }_2{\lambda }_3}}& \left(\mathrm{Equation}7\right)\\ S_B=\sqrt{{\lambda }_1^2+{\lambda }_2^2+{\lambda }_3^2}& \left(\mathrm{Equation}8\right)\end{array}$

Because the eigenvalues λ₁, λ₂, λ₃ indicate the magnitudes of corresponding orthogonal curvatures represented by eigenvectors of Hessian, R_B describes the likelihood of the blob-like structure, and can attain maximum 1 when the region is an idealized blob. S_B describes the deviation of the three orthogonal curvatures from “flat”. The higher value S_B is, the stronger contrast can be achieved of the blob region against the background.

However, to calculate R_B, eigenvalues of Hessian need to be solved at each voxel. This procedure requires intensive computations. Accordingly, in order to improve computational efficiency, method 200 utilizes a new approach where modified features R_T and S_T are utilized. These two features are built upon regional Hessian evaluated at each glomerular candidate T instead of each voxel. The regional Hessian over the DoG transformation (smoothed images) is defined as

$\begin{array}{cc}{H}_T\left({\mathrm{DoG}}_{\mathrm{nor}}\left(x,y;t\right)\right)=\sum _{\left(x,y,z\right)\in T}[\begin{array}{ccc}\frac{{\partial }^2{\mathrm{DoG}}_{\mathrm{nor}}\left(x,y,z;t\right)}{\partial x^2}& \frac{{\partial }^2{\mathrm{DoG}}_{\mathrm{nor}}\left(x,y,z;t\right)}{\partial x\partial y}& \frac{{\partial }^2{\mathrm{DoG}}_{\mathrm{nor}}\left(x,y,z;t\right)}{\partial x\partial z}\\ \frac{{\partial }^2{\mathrm{DoG}}_{\mathrm{nor}}\left(x,y,z;t\right)}{\partial x\partial y}& \frac{{\partial }^2{\mathrm{DoG}}_{\mathrm{nor}}\left(x,y,z;t\right)}{\partial y^2}& \frac{{\partial }^2{\mathrm{DoG}}_{\mathrm{nor}}\left(x,y,z;t\right)}{\partial y\partial z}\\ \frac{{\partial }^2{\mathrm{DoG}}_{\mathrm{nor}}\left(x,y,z;t\right)}{\partial x\partial z}& \frac{{\partial }^2{\mathrm{DoG}}_{\mathrm{nor}}\left(x,y,z;t\right)}{\partial y\partial z}& \frac{{\partial }^2{\mathrm{DoG}}_{\mathrm{nor}}\left(x,y,z;t\right)}{\partial z^2}\end{array}]& \left(\mathrm{Equation}9\right)\end{array}$

Equation 9 is the summation of Hessian matrix over candidate region T. In an exemplary embodiment, this matrix describes the second-order derivative distribution within the region of the blob candidate. The derivatives are equally weighted averaged (sum over the region T) at the centroid of T over the region. The eigenvalues of this matrix represent the three principal curvatures of the centroid over the blob candidate, and can be utilized to measure the blobness over the region.

To efficiently measure the likelihood of blobness, a modified version of Equation (7) is utilized as:

$\begin{array}{cc}{R}_T=\frac{1}{\left(\frac{\left(\mathrm{xy}\mathrm{plane}\mathrm{Cross}\mathrm{section}\text{/}\pi \right)}{{\left(\mathrm{Volume}\mathrm{of}\mathrm{blob}\text{/}\left(4\pi \text{/3}\right)\right)}^{2\text{/}3}}+\frac{\left(\mathrm{xz}\mathrm{plane}\mathrm{Cross}\mathrm{section}\text{/}\pi \right)}{{\left(\mathrm{Volume}\mathrm{of}\mathrm{blob}\text{/}\left(4\pi \text{/3}\right)\right)}^{2\text{/}3}}+\frac{\left(\mathrm{yz}\mathrm{plane}\mathrm{Cross}\mathrm{section}\text{/}\pi \right)}{{\left(\mathrm{Volume}\mathrm{of}\mathrm{blob}\text{/}\left(4\pi \text{/3}\right)\right)}^{2\text{/}3}}\right)\text{/}3}=\frac{3\times {{\lambda }_1^{\prime }{\lambda }_2^{\prime }{\lambda }_3^{\prime }}^{\frac{2}{3}}}{{\lambda }_1^{\prime }{\lambda }_2^{\prime }+{\lambda }_2^{\prime }{\lambda }_3^{\prime }+{\lambda }_1^{\prime }{\lambda }_3^{\prime }}& \left(\mathrm{Equation}10\right)\end{array}$

In Equation 10, λ′₁, λ′₂, λ₃′ are eigenvalues of regional Hessian H_T. Since the Hessian matrix is negative definite at every voxel within a blob candidate, the sum of the Hessian matrix over a blob candidate is negative definite, meaning λ′₁, λ′₂, λ′₃<0. Thus, we have

$\begin{array}{cc}{R}_T=\frac{3\times {{\lambda }_1^{\prime }{\lambda }_2^{\prime }{\lambda }_3^{\prime }}^{\frac{2}{3}}}{{\lambda }_1^{\prime }{\lambda }_2^{\prime }+{\lambda }_2^{\prime }{\lambda }_3^{\prime }+{\lambda }_1^{\prime }{\lambda }_3^{\prime }}=\frac{3\times {\mathrm{tr}\left(H_T\right)}^{\frac{2}{3}}}{\mathrm{pm}\left(H_T\right)}& \left(\mathrm{Equation}11\right)\end{array}$

where pm(H_T)=λ₁′λ₂′+λ₂′λ₃′+λ₁′λ₃′. Additionally, pm(H_T) can be obtained by calculating three 2 by 2 principal minors of H_T, that is

$\begin{array}{cc}{\lambda }_1^{\prime }{\lambda }_2^{\prime }+{\lambda }_2^{\prime }{\lambda }_3^{\prime }+{\lambda }_1^{\prime }{\lambda }_3^{\prime }=\det \left(\begin{array}{cc}{H}_T^{1,1}& H_T^{1,2}\\ H_T^{2,1}& H_T^{2,2}\end{array}\right)+\det \left(\begin{array}{cc}{H}_T^{2,2}& H_T^{2,3}\\ H_T^{3,2}& H_T^{3,3}\end{array}\right)+\det \left(\begin{array}{cc}{H}_T^{1,1}& H_T^{1,3}\\ H_T^{3,1}& H_T^{3,3}\end{array}\right)& \left(\mathrm{Equation}12\right)\end{array}$

In method 200, to calculate the flatness over blob candidate T, we can utilize:

$\begin{array}{cc}{S}_T=\sqrt{{\left({\lambda }_1^{\prime }+{\lambda }_2^{\prime }+{\lambda }_3^{\prime }\right)}^2-2\left({\lambda }_1^{\prime }{\lambda }_2^{\prime }+{\lambda }_2^{\prime }{\lambda }_3^{\prime }+{\lambda }_1^{\prime }{\lambda }_3^{\prime }\right)}=\sqrt{{\mathrm{tr}\left(H_T\right)}^2-2\times \mathrm{pm}\left(H_T\right)}& \left(\mathrm{Equation}13\right)\end{array}$

In method 200, modified features R_T and S_T greatly reduce the computational burden as compared to prior approaches (illustrated in Eq. (7) and Eq. (8)) in that: (1) R_T and S_T are based on the regional Hessian evaluated at each blob region instead of at every voxel, and (2) R_T and S_T only require the calculations of trace and determinant, instead of requiring to find the roots (eigenvalues) of the characteristic equation of the Hessian matrix in Eq. (7) and Eq. (8). In method 200, R_T and S_T may be utilized together with the feature A_T, the average intensity value over region T. In various exemplary embodiments, these three features are input to a clustering algorithm, for example a Variational Bayesian Gaussian Mixture Model, to reduce and/or remove false identifications from the glomeruli candidate pool.

In various exemplary embodiments, method 200 utilizes a post pruning approach (step 240) in order to eliminate false glomeruli candidates. In method 200, a Variational Bayesian Gaussian Mixture Model (VBGMM) may be utilized because: (1) compared to the maximum likelihood Gaussian Mixture Model, the variational model is free from being trapped to a singularity solution; and (2) the variational model is able to automatically identify the number of clusters for optimum performance without the need for initialization and subjective parameter settings. In the VBGMM, given a 3D MR image, it can be considered that several multivariate Gaussian distribution components form the entire image. One of the components is a group of glomeruli, with others belonging to the background and image noises. In method 200, X={X¹, . . . , X^N} is the observation, and N(Xⁱ|μ, Λ) is the multivariate Gaussian distribution with mean μ and inverse covariance Λ that Xⁱ follows. The mixture distribution for M components is:
P(Xⁱ|π,μ,Λ)=Σ_j=1^Mπ_jN(Xⁱ|μ,Λ)  (Equation 14)

where π_j is the weight for component j.

In method 200, the elements in X may be assumed to be independent to each other, and a binary latent variable, Z={Z¹, . . . , Z^NM} where z_im=1 may be introduced. This indicates that Xⁱ belongs to class m and Σ_j=1^MZ^ij=1. The conditional probability of the image data set is:
P(X|Z,μ,Λ)=Π_i=1^NΠ_j=1^MN(Xⁱ|μ,Λ)^Zij  (Equation 15)

In method 200, the variational Bayesian Gaussian Mixture model can approximate the posterior P(θ|X) given any distribution P(X) and unknown parameters θ, by a simpler distribution Q(θ) that marginalizes the unknown parameter θ. In various exemplary embodiments, observation Xⁱ is a vector of three features A_T, R_T, S_T for a glomerular candidate region. The observations (glomerular candidate regions) form a multivariate Gaussian mixture and therefore they can be clustered into glomerular regions and non-glomerular regions, for example by using a Bayesian inference method.

Validation of method 200 may be conducted against known results on known images, for example a set of 15 pathological images with known results. The comparison result is shown in Table 1.

As shown in Table 1, across all the range of d, method 200 outperforms gLoG on Recall while underperforming gLoG on precision. This leads to a comparable result of F-score on method 200 compared to gLoG (for some values of d, HDoG outperforms gLoG while for the rest values of d, gLoG outperforms HDoG). Comparing to LoG detector, method 200 provides better results on Precision, Recall and F-score for all d. In addition, the results of standard deviation show that method 200 is more robust and has fewer variations as compared to gLoG and LoG.

Another comparison result of method 200 against known results on known images (a validation dataset of 200 256×256 fluorescence-light microscopy cell images) is shown in Table 2.

As shown in Table 2, though the Recall on method 200 is comparable to gLoG, on Precision and F-Score, method 200 outperforms gLoG for all d. Compared to LoG, method provides better performance on all three metrics: Precision, Recall and F-Score across the range of d. Additionally, method 200 has the least variations and thus is quite robust.

Method 200 may be validated by operation on human kidney images, as shown in Table 3:

Exemplary results of method 200 are illustrated in FIGS. 11A through 12B. Additional details regarding certain exemplary principles of the present disclosure, for example methods 100 and/or 200, may be found in M. Zhang et al., “Small Blob Identification in Medical Images Using Regional Features from Optimum Scale” published in IEEE Transactions on Biomedical Engineering, Vol. PP, Issue 99 on Sep. 25, 2014, the contents of which are incorporated herein by reference in their entirety.

It will be appreciated that, while principles of the present disclosure have been discussed in terms of kidneys and glomeruli, such principles may desirably be applied to other organs and/or disorders, for example those wherein blob detection is desirable.

While the principles of this disclosure have been shown in various embodiments, many modifications of structure, arrangements, proportions, the elements, materials and components, used in practice, which are particularly adapted for a specific environment and operating requirements may be used without departing from the principles and scope of this disclosure. These and other changes or modifications are intended to be included within the scope of the present disclosure and may be expressed in the following claims.

The present disclosure has been described with reference to various embodiments. However, one of ordinary skill in the art appreciates that various modifications and changes can be made without departing from the scope of the present disclosure. Accordingly, the specification is to be regarded in an illustrative rather than a restrictive sense, and all such modifications are intended to be included within the scope of the present disclosure. Likewise, benefits, other advantages, and solutions to problems have been described above with regard to various embodiments. However, benefits, advantages, solutions to problems, and any element(s) that may cause any benefit, advantage, or solution to occur or become more pronounced are not to be construed as a critical, required, or essential feature or element of any or all the claims.

As used herein, the terms “comprises”, “comprising”, or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but may include other elements not expressly listed or inherent to such process, method, article, or apparatus.