WO2016141449A1 - Computer-assisted focused assessment with sonography in trauma - Google Patents

Abstract

Systems, methods, and devices relating to the detection and characterization of a patient's kidney using 3-D ultrasound. An operator is provided with instructions for proper placement of a probe on a patient. The system determines if the probe is properly placed, otherwise, the operator is continuously prompted to properly place the probe. Ultrasound images are obtained using the probe and the patient's kidney is detected in the images using an organ atlas database. Once the kidney is detected, if the probe is misaligned, corrective instructions for the proper alignment of the probe are sent to the operator.

Description

COMPUTER-ASS ISTED FOCUSED ASSESSMENT WITH SONOGRAPHY IN TRAUMA FIELD OF THE INVENTION

[0001] The present invention relates to the field of

ultrasound imaging. More specifically, this invention relates to computer-assisted probe placement which facilitates Focused Assessment with Sonography in Trauma (FAST) examination by non-trained operators.

BACKGROUND

[0002] The term trauma refers to an internal free-fluid

caused by internal bleeding. Emergency surgical intervention is often reguired to save trauma patients' lives. The ability for non-skilled technicians to rapidly detect trauma is desirable and potentially life-saving.

[0003] Trauma may be detected by way of computed tomography

(CT), magnetic resonance imaging (MRI) or ultrasound imaging. Trauma detection by ultrasound is

particularly favorable when compared to CT and MRI as it is non-invasive, does not impose restriction of use on patients, is portable, and may provide real-time imaging .

[0004] Recent technological advancements in ultrasound

beamforming have resulted in rapid three-dimensional (3-D) ultrasound imaging. 3-D ultrasound provides the location of the ultrasound signal in space along with a volumetric representation of internal organs. Accordingly, 3-D ultrasound images are also sometimes referred to as ultrasound volumes.

[0005] Knowledgeable practitioners in the field of sonography often refer to trauma detection by ultrasound as Focused Assessment with Sonography in Trauma, or FAST. FAST reguires a trained FAST examiner, such as a radiologist, to control and move an ultrasound probe around a patient's body. A FAST examiner, referred to in this document as an operator, views images while moving an ultrasound probe. In doing so, the operator searches the images for darkened regions representing low-echoic, blood-filled areas. When a dark region is located, the operator uses his or her technical expertise and knowledge to determine whether the detected dark area represents trauma, a fluid-carrying organ, or a shadow.

^0006] Placement of an ultrasound probe which provides a view of the anatomical region known as Morison's pouch is known by those skilled in the art as being highly sensitive for trauma detection. This view is of particular interest as it shows both the right side of the liver and the right kidney, effectively

visualizing the hepatorenal recess of subhepatic space, commonly referred to as Morison's pouch. Morison's pouch view also precludes possible

misdiagnosis of trauma due to shadows cast by the presence of kidney stones.

^0007] A correct Morison's pouch view shows the kidney's

unigue and easily distinguishable structure in a particular location and orientation. As such, it is of vital importance that an operator properly segment, or identify, the kidney when obtaining a Morison's pouch view to provide an accurate trauma diagnosis. However, the inherent low guality of ultrasound images and the natural gaps along the kidney boundary may cause a technician to misalign the ultrasound probe leading to poor kidney detection and segmentation

(identification) .

[0008] A Morison's pouch view is optimally acguired by

placing the ultrasound probe at the intersection of Horizontal Subxiphoid (HS) line and the right mid- auxiliary line, with the probe marker directed toward the head. The HS line is an imaginary horizontal line connecting the xiphoid process to the mid-auxiliary line .

[0009] When a trauma patient is laid in a supine position,

Morison's pouch fills with blood. In FAST examination, this appears as a dark (low-echoic) region between the kidney and the liver. An expert with knowledge of human anatomy and ultrasound imaging would be able to position the ultrasound properly on the trauma patient's body and determine the presence of trauma. An examiner lacking such skill and knowledge, however, would be unlikely to properly place the probe and segment the kidney, leading to potential misdiagnosis and loss of life. Therefore, the current state of the art reguires that a FAST examiner have expertise in human anatomy and ultrasound imaging.

[0010] Methods and technigues have been developed to aid

operators in segmenting internal organs such as kidneys for trauma detection. However, such methods reguire operator intervention to determine the presence of trauma and to obtain the valuable 3-D ultrasound images that indicate the precise location of the trauma on the patient in real-time. These methods, therefore, reguire operator expertise and may be time consuming. [0011] Among all internal organs, the kidney has a unique appearance structure in 3D ultrasound volumes, such that its distinct shape distinguishes it from all other internal organs. Since the right kidney is visible in the Morison's pouch view, kidney

segmentation can be used as a cornerstone of a ultrasound automated diagnosis system. To provide an automated kidney segmentation, applying a robust and accurate kidney detection is crucial to eliminating the need for human intervention to manually initialize the kidney segmentation process

[0012] For example, in 2005, Fernandez and Lopez reported a method aiding operators in ultrasound kidney

segmentation (Medical Image Analysis, vol. 9, no. 1, pages 1-23) . Fernandez and Lopez's semi-automated approach requires the operator to manually outline kidney boundaries on individual input two-dimensional (2-D) ultrasound images using Markov random field and active contours. These manipulated images are then automatically reconstructed into a 3-D kidney shape. Such a method is time consuming and requires technical expertise on the part of the FAST examiner.

[0013] In 2012, Prevost et al . proposed another method to aid operators in ultrasound kidney segmentation in a report entitled "Kidney detection and real-time segmentation in 3d contrast-enhanced images." at the 9^th IEEE International Symposium on Biomedical Imaging (ISBI) . In this method, input 3-D ultrasound images are enhanced to show the kidney as brighter than other regions. Prevost et al . then apply a 3-D deformable model to these images. In this method, the operator is required to manually select landmarks inside and/or outside the kidneys . Their method then directly imposes the selected landmarks on the energy

functional to prevent the segmentation process from becoming trapped in a local minima. The success rate of this approach depends on the guality of the input contrast enhanced ultrasound images. The performance of this approach has not been reported on 3-D ultrasound images. Therefore, similar to the method proposed by Fernandez and Lopez, the method proposed by Prevost et al . does not provide an automated and rapid method to aid a FAST examiner in kidney segmentation and trauma detection.

[0014] There is therefore a need to mitigate, if not

overcome, the shortcomings of the prior art and to, preferably, provide an automated approach for segmenting kidneys for FAST examiners lacking expertise and knowledge in ultrasound imaging and human anatomy.

SUMMARY OF INVENTION

[0015] The present invention provides systems, methods, and devices relating to the detection and characterization of a patient's kidney using 3-D ultrasound. An operator is provided with instructions for proper placement of a probe on a patient. The system determines if the probe is properly placed, otherwise, the operator is continuously prompted to properly place the probe. Ultrasound images are obtained using the probe and the patient's kidney is detected in the images using an organ atlas database. Once the kidney is detected, if the probe is misaligned, corrective instructions for the proper alignment of the probe are sent to the operator. [0016] The present invention provides a computer assisted probe placement that facilitates the focused

assessment with sonography in trauma (FAST)

examination by non-trained operators using 3D ultrasound systems.

[0017] The present invention provides an automated ultrasound probe placement for FAST examination.

[0018] The present invention facilitates FAST examination by non-trained operators in emergency situations.

[0019] In a first aspect, the present invention provides a method for detecting a presence of trauma in a patient, the method comprising: a) providing an operator with an indication of a location on said patient for an initial placement of a probe; b) obtaining images of an internal area of said patient using said probe; c) determining if at least one specific organ is present in images obtained in step b) ; d) in the event said at least one specific organ is not present in said images, repeating steps a) - c) until said at least one specific organ is present in said images; e) determining an alignment of said at least one specific organ in said images; f) determining if a desired view of said internal area is found in said images; g) in the event said desired view is not found in said images, relating said alignment of said at least one specific organ to an alignment of said probe; h) determining corrective instructions for probe placement based on results of step g) , said corrective instructions being operative to adjust a view of said internal area to thereby provide said operator with a view with a correct alignment of said at least specific organ; i) sending said corrective instructions to said operator ; j ) obtaining images of said internal area of said patient using said probe; k) determining if at least one specific organ is present in images obtained in step j);

1) repeating steps e) - k) until said desired view is found in said images. In a second aspect, the present invention provides a method for compiling a database for use in detecting a presence of a specific organ in an ultrasound image, the method comprising: a) pre-processing at least one training image; b) manually segmenting an organ in said at least one training image; c) generating an average shape model for said organ based on said at least one training image and said organ segmented in step b) ; d) extracting texture features from said organ segmented in step b) ; e) training kernel based support vector machines to classify portions of images as being organ or non-organ based on said at least one training image; f) storing said kernel based support vector machines and said organ average shape model in said database. In a third aspect, the present invention provides a system for detecting an organ in at least one

ultrasound image of an internal area of a patient, the system comprising:

- processor for processing said at least one ultrasound image;

- data storage containing a database, said database including :

- a reference image of said organ;

- at least one training image containing at least one segmented organ;

- at least one kernel based support vector machine, said support vector machine being for classifying portions of said at least one image as being organ or non-organ based on said at least one training image; and

- an average shape model for said organ based on said at least one training image. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] The embodiments of the present invention will now be described by reference to the following Figures, in which identical reference numerals in different figures indicate identical elements, and in which:

Figure 1 shows the overall block diagram of the automated ultrasound probe navigation system;

Figure 2 shows a graphical representation of the initial probe placement;

Figure 3 schematically illustrates the steps for the first stage of generating a kidney atlas database, including selecting training ultrasound volumes, selecting a reference volume, registering the training ultrasound volumes on the reference volume, and manually segmenting the registered volumes;

Figure 4 is a block diagram for the second stage of generating a kidney atlas database including

generating the average kidney shape model;

Figure 5 schematically illustrates the steps in the third stage of generating a kidney atlas database of the present invention, including training spatially distributed Kernel-based Support Vector Machines (KSVM) ;

Figure 6 schematically details steps in kidney segmentation of input ultrasound volumes;

Figure 7 is a flow chart detailing the steps in an automated kidney segmentation process by generation of a binarized mask specifying voxels with a higher probability of being kidney tissue; Figure 8 shows the graphical user interface of the automated ultrasound probe navigation; and

Figure 9 shows the relation of the detected kidney orientation and ultrasound probe rotation.

^0023] The Figures are not to scale and some features may be exaggerated or minimized to show details of particular elements while related elements may have been eliminated to prevent obscuring novel aspects. Therefore, specific structural and functional details disclosed herein are not to be interpreted as limiting but merely as a basis for the claims and as a representative basis for teaching one skilled in the art to variously employ the present invention.

DETAILED DESCRIPTION

[0024] In one aspect of the invention, the present invention involves an organ atlas database, an automated kidney segmentation process, an ultrasound probe navigation system, and an automated trauma diagnosis system to help guide a FAST examiner, or an untrained operator, to detect trauma. Input 3D ultrasound images are automatically compared with the 3D images from an organ atlas database using real-time image processing misalignment calculations to obtain a correct view. Automated organ segmentation using the organ atlas database is then performed and probe navigational commands are generated and sent to an operator to find the correct Morison's pouch view. After verifying that a correct view has been obtained, trauma is diagnosed using the segmented kidney. [0025] For the explanation and description below, the organ of interest will be referred to as a kidney and a desired view will be referred to as a Morison's pouch view. However, as should be clear to a person of skill in the art, the present invention may be applied to any other organ and to any desirable view of any anatomical area of a mammal.

[0026] Once the operator is provided with indications as to where to place the ultrasound probe, it should be clear that the ultrasound image gathered should include a view of the kidney. A partial or full image of the kidney should be present and this can be detected by the system. Detection of at least a portion of the kidney is possible as the kidney's unigue shape and positioning can be compared to predetermined reference images of the kidney stored in the database. If any portion of the acguired image is not found to contain at least a section of a kidney that conforms to the expected shape and location of the kidney, then the operator must have placed the probe at an incorrect location. The operator is thus prompted to start the process anew. Once the kidney has been detected in the acguired images, the placement and orientation of the kidney is compared to its expected location and orientation as determined by the training images in the database. By continuously comparing the shape of the acguired image of the kidney with the expected shape of a kidney,

translations and rotations to conform the shape of the acguired image of the kidney to the expected shape can be calculated. These calculated translations and rotations are then sent to the operator as corrective instructions for the placement of the probe. [0027] Figure 1 is a flowchart detailing the steps involved in a method according to one aspect of the invention. In this method, the operator is provided with automated guidance for a placement of the ultrasound probe (step 10) . The relevant ultrasound volumes are then gathered using the probe (step 20) . A kidney detection is then performed (Step 30) . Decision 40 then determines if the kidney has been detected or if the kidney exists in the ultrasound volumes. If a kidney has not been detected in the ultrasound images, the operator is again guided to a proper placement of the ultrasound probe. On the other hand, if the kidney has been detected, then a calculation of the kidney misalignment is performed (step 50). Decision 60 then determines if a proper Morison's pouch view has been achieved. In the event the proper view has not been achieved, the probe's misalignment is determined based on a misalignment of the kidney from the ultrasound images (step 70) . A corrective probe alignment command is then sent to the operator (step 80) and the method loops back to that of acguiring ultrasound images (step 20).

[0028] Returning to step 60, if the proper Morison's pouch view has been achieved, then the kidney in the ultrasound images is segmented (step 90) . This data is then used for an automated trauma diagnosis (step 100) .

[0029] In one aspect, the invention provides an operator with instructions or indications for a proper placement of the ultrasound probe. Ultrasound images are then gathered using the probe. Whether the probe is properly placed or not is automatically determined by analyzing the gathered images using the kidney as a reference point. If the kidney or a portion of the kidney is not detected from the ultrasound images, the operator is again prompted with instructions for a proper initial placement of the probe. If at least a portion of the kidney is detected, then calculations are performed to determine any misalignment of the kidney in the images. Any misalignment of the kidney is related back to a misalignment of the probe placement. Corrective instructions for the probe placement are then automatically provided to the operator. The process repeats until the images indicate a correct view of the area around the kidney. Using this correct view of the area, the kidney is segmented and the images are then used for trauma diagnosis .

[0030] Figure 2 is a sample graphical user interface (GUI) which may be used with the invention. The GUI presents the operator with a front image of a human body and an indication of where the ultrasound probe is to be placed on a patient's body for a proper initial placement. As can be seen in Figure 2, for a proper Morison's pouch view, the operator is guided to place the ultrasound probe at the intersection of the HS and mid-auxiliary lines to ensure that the right kidney is at least partially present and is in the correct orientation.

[0031] It should be noted that while the GUI illustrated in

Figure 2 or a variant thereof may be used, other implementations are also possible. As an example, the indication as to where to place the probe on the patient may actually be projected on the actual patient. Similarly, automated, pre-recorded auditory instructions, predetermined written instructions (with or without accompanying diagrams), and any other visual or auditory means to guide the operator may also be used. As with the GUI example, these initial instructions are provided to the operator so that the operator is enabled to locate a proper initial probe placement .

^0032] The automated navigation system then processes the acguired 3-D ultrasound images, also referred to as ultrasound volumes, to detect the kidney.

^0033] The position and orientation of the detected kidney are used to calculate potential ultrasound probe misalignment. In the event that probe misalignments are found, corrective instructions for the operator are calculated. The navigation system's instructions guide the operator to move or rotate the ultrasound probe toward the correct placement . Similar to the initial placement of the probe, the corrective instructions to the operator can take multiple forms . The operator can be guided by a projection of the calculated misalignment onto the region where the ultrasound probe is placed. Alternatively, the operator can be guided by an auditory navigation command sent to the operator. These corrective navigation commands can also be displayed on a GUI or projected on the patient. The corrective instructions can take the form of text or symbols presented to the operator using any suitable means.

^0034] Ultrasound probe misalignments are determined by the navigation system using an organ atlas database. The atlas database is generated in advance using manually segmented organs in a training set of 3-D ultrasound images . The organ atlas database stores information required for segmenting an organ of interest in input ultrasound volumes. Such information may include: the reference volume, segmented organs in the training set of ultrasound images, the organ average shape model, and trained spatially distributed kernel based support vector machines (KSVMs) .

[0035] The organ atlas database-based approach consists of

(i) generating an organ atlas database, and (ii) segmenting the organ of interest in input 3-D ultrasound images. Essentially, the organ atlas database is used to determine if a potential organ of interest in the input image (i.e. the image obtained with the probe) is a kidney and if the kidney is in a proper alignment in the image. It should be noted that, in one implementation of the invention, the organ atlas database is part of the navigation system. The database is generated separately and is used whenever the system is used.

[0036] To generate an organ atlas database, multiple steps are involved and these steps are outlined in the schematic charts of Figures 3, 4 and 5. In the description of the steps for generating such an organ atlas database, the organ of interest will be the right kidney and the desired view will be, as above, a view of Morison's pouch. Of course, other organs of interest and other desired views may be used. Generating the requisite organ atlas database involves three main steps: a) pre-processing training volumes and segmenting these training volumes, b) generating an organ average shape model, and c) extracting features from the training volumes and training KSVMs. These three main steps are illustrated schematically in Figures 3, 4 and 5. [0037] In Figure 3, the sub-steps for the main step of preprocessing and manually segmenting training volumes are : i. Selecting a training set of 3-D ultrasound images ;

ii . Reducing speckle noise in the training set of 3-D ultrasound images;

iii . Removing intensity inhomogeneity field using a bias correction approach;

iv . Selecting a reference volume;

v. Manually selecting corresponding points on source and reference 3-D Ultrasound Images (landmark selection) ;

vi . Registering the training set volumes on the reference volume via a rigid-body transformation ;

vii . Manually segmenting the registered

volumes ;

[0038] In Figure 4, the main step is that of generating an organ average shape model. The sub-step is that of storing the segmentations as binarized masks in the atlas database.

[0039] In Figure 5, the main step is that of extracting

texture features. This is accomplished by: i) Extracting features as 3-D volumes from a

registered volume using Gabor filters; ii) Extracting sub-volumes from 3-D feature volumes; iii) Training Kernel-based Support Vector Machines (KSVMs) for sub-volumes, and storing trained KSVMs in the atlas database; iv) Using binarized masks of the manually segmented kidneys to generate an organ average shape model, and storing the average shape model in the atlas database.

[0040] Referring to the steps illustrated in Figure 3, the process begins with the selection of a training set of 3-D ultrasound images (step 200). Each training ultrasound volume, V[_{r r} where i e {1,2, . . . , N_training} , should have a representative organ, such as a kidney, represented. A graphical user interface (GUI) designed to allow a user to select and load ultrasound volumes from the computer storage is used in one

implementation of the invention.

[0041] After the training volumes have been selected, speckle noise of each training ultrasound volume is reduced using an anisotropic diffusion filter (step 210).

Partial differential eguations (PDE) are employed to spatially control smoothing power based on the distance of voxels to object edges. Voxels located away from object edges are more highly smoothed. Voxels close to object edges are only smoothed in the direction parallel to the object edge. The anisotropic diffusion filter operates based on the following PDE formula,

Eguation (1) ^^j^ = div[c(q(x,y,z,-t))^■ VV_ds(x,y,z,-t)], where V and div are gradient and divergence operators, respectively. C(.) is the diffusion coefficient, and q (x, y, z; t) is the instantaneous coefficient of

variation, which is itself a function of the magnitude of the volume gradient, and intensifies edges. V_ds is the despeckled volume and is initiated as V_ds (x, y, z; t = 0) = V(x,y,z). The expression c(q(x,y,z; t)) is defined using the following eguations :

Eguation (2) c(q(x,y,z; £)) =—_r„ ¹ _λ , , ,

⁺ [¾§ω(ι+¾§ω)] and,

Eguation (3) q(x,y,z;t)

where q₀( ~ QO^exP[^— t] is the speckle scale function, and q₀ is the speckle coefficient of variation. q x,y,z; t) operates as an edge detector. Details regarding discretization can be found in reference [9]

identified below.

[0042] In step 220, the intensity inhomogeneity of the input ultrasound volume is removed. This may also be referred to as "bias correction". The intensity inhomogeneity is modeled as a multiplicative bias field,

Eguation (4) V_ds(x,y,z) = V_bc(x,y,z)f(x,y,z)_r where f(x,y,z) is a bias field, and V_bc(x,y,z) is the bias corrected volume .

[0043] To linearize eguation (4), a logarithmic

transformation is applied,

Eguation (5) V_ds(x,y,z) = V_bc(x,y,z) +f(x,y,z) , where V_ds = log(V_ds , V_bc = log (V_bc), and / = log (/) . Now, let Pv_dsr Pv_bcr and Pf be the probability density functions of V_ds, V_bc, and /, respectively. The probability density functions are related as, Equation

;0044] Pv_ds is calculated at each voxel based on the histogram of a sub-volume centered at the voxel. Pv_bc is unknown, and Pf is simplified to be a Gaussian distribution with zero mean and an unknown variance. By taking the Fourier transform of the probability density

functions, Pv_bc r which is the Fourier transform of y_h , is approximated as follows,

Equation (7) P_{Vbc= 2+z2}Py_ds where Pf and P_v are Fourier transforms of Pf and Pv_ds r respectively. Then, an estimation of / is achieved as,

Equation (8) f_es(x,y,z) = V_ds(x,y,z) -

where ,

Equation (9)

S_∞ V_bc(x,y,z)Pf(v_ds(x,y,z)- V_bc (x,y,z))P_Vbc( _bc(x,y,z)^dV_bc(x,y,z)

f- o Pf(v_ds{x,y,z)- V_bc{x,y,z))p_Vbc( _bc{x,y,z))dV_bc{x,y,z)

[ 0 045 ] After estimating f(x,y,z) at all voxels, V_bc(x,y,z) is

recovered using equation ( 5 ) , V_bc(x, y, z) = exp (v_bc(x, y, z) is obtained.

[ 004 6 ] The training ultrasound volumes, V^_h = V_b ^l _c, where

i e {1, ... , N_Training}, are then enhanced, and a reference volume is selected (step 230 ) . It should be noted that, to allow a user to perform the selection, a suitable user interface may be used. Preferably, the best quality ultrasound volume is selected as the reference volume.

[0047] Once the reference volume has been selected, all other training ultrasound volumes are then registered onto the reference volume using a landmark-based rigid-body registration (step 240). This preserves the

variability in a given organ shape among individuals while removing training ultrasound volume

misalignments. For optimal results, a GUI designed to allow the operator to manually select landmarks can be used. Such a GUI may provide the ability to select pair-voxels on the reference volume, _ef i ^an each training ultrasound volume, X^_rc .

[0048] After selecting landmarks for each training ultrasound volume, an affine transformation is fitted on the selected landmarks, with an aim to register the training ultrasound volume on the reference volume.

The registered volume is V_rreg · For example, an affine transformation with 7 parameters may be applied, such as 3 translations, 3 rotations and 1 scale, as = [m_x, .y, m_z, θ_χ, 9y, θ_ζ, s] . The affine transformation matrix is defined as follows,

Equation (10) X^k _src = AX^kref where,

Equation (11) A =

where R_x, R_y, and R_z are rotation matrices of the x-, y- and z- axes, respectively. Matrix A is defined to minimize the following error,

Eguation (12)

[0049] In the next step (step 250), the organs, such as

kidneys, in the registered training volumes are manually segmented. The results are stored as binarized masks, Bⁿ where n e {1, ... ,N_Training} . In these binarized masks, the image consists of Is inside the organ and 0s outside of the organ as shown in one section of Figure 4. These results are stored in the organ atlas database (step 260).

[0050] After pre-processing and manually segmenting the

training volumes, a kidney average shape model is generated to initialize kidney segmentation (Figure 4) . This may result in a faster and more accurate segmentation. The organ's average shape model is generated using the manual segmentations that are already aligned, as described above and illustrated in Figure 3 . As such, the voxel-wise average is

calculated on the segmentations using, y^NTraining _n . .

Eguation (13) KASM(x,y,z) = ^i=J il

N Training where KASM is the kidney average shape model. KASM is then saved in the kidney atlas database (Figure 4) .

[0051] Once the average shape model has been developed,

texture features from the registered ultrasound volumes can be extracted. The texture features are joined with the manual segmentations and these features are then used to train KSVM classifiers to automatically segment the kidney portion of an input ultrasound volume.

^0052] Gabor filters are highly effective for extracting

texture features in image analysis applications. 3-D Gabor wavelets, sinusoidal waves modulated by 3-D Gaussian functions, may be used to extract texture features from ultrasound volumes as follows,

Equation (14) G_ffii (x, y, z) =

and,

Equation (15) u = fsin<Pcos9, v = fsin<Psin9

where S and / are a normalization scale and the amplitude of the complex sinusoids, respectively. σ_χ, Oy , and σ_ζ are the Gaussian envelop widths in the x-,y- , and z- axes, respectively. R is the 3-D rotation matrix with 3 parameters: θ_χ, 9_y, and θ_ζ .

^0053] A simplification of the 3-D Gabor wavelet equation may be applied, σ_χ = o_y= σ_ζ = σ. Each extracted feature from a registered training ultrasound volume, V_r ^l _eg, has a specific set of parameters [<?,θ_χ,θ_γ,θ_ζ], and is named

Ρσθ_χθγθ_ζ· ^Any number can be selected to define o,9_x,9_y, or θ_ζ . For example, if θ_χ, 9_y, θ_ζ e {0,45,90,135} and

oe {0.4,0.6}, it would result in 2 * 4³ = 128 features, F{ = F^_Wz\le[l 128]}.

[0054] Texture information of a given organ, such as a

kidney, highly changes throughout an ultrasound volume. Therefore, using a single classifier may not be effective to classify kidney and non-kidney tissues in ultrasound volumes. To overcome this problem, texture information may be locally classified by dividing the entire volume into a set of sub-volumes . In such a case, each volumetric feature, F{ , may be divided into

K sub-volumes, K'^k where ke{l,...,K}. Furthermore, the sub-volumes may be divided unevenly, or evenly such as with 50 percent overlapping. For each pair of sub- volumes k and features I, a sub-set of N_Tra£n£n5-related sub-volumes, {F_; ^1,fe,

_may collected.

[0055] Figure 5 schematically shows the remaining stages of generating the kidney atlas database. As can be seen, the various stages are applied in parallel to multiple datasets once the relevant texture features have been extracted using the Gabor wavelets . The sub-volumes are extracted and then the related sub-volumes in the sub-sets are converted into vectors in a process called vectorization . In vectorization, vectors in each sub-set are vertically concatenated to create a single vector for each sub-set, F^k . Then, vectors related to each sub-volume are horizontally

concatenated and form a sub-volume feature matrix, M^k = [F^k, ... , F^k] . Following the previous example, the matrix would be M^k = [F^,^, ... , F^₂₈] . [0056] A KSVM classifier may then be used to classify voxels into kidney and non-kidney tissues. Rather than minimize the classification error, KSVM provides a maximum margin distance between two classes to obtain a maximum generalization ability. In addition, using a kernel-based operation, KSVM is able to solve nonlinear classification problems by mapping samples into a higher dimensional space.

[0057] In one embodiment of the present invention, the

Gaussian kernel function may be selected for the KSVMs . To train spatially distributed KSVMS, the manual segmentations may be used to specify classes of each row in each sub-volume feature matrix, M^k . Thus, sub-volumes, B^l,k, are extracted from each B^l where i e [1, ... , N_Training\ and k e {1, ... , K] . Each sub-volume, B ^l,k , is then vectorized and sub-volumes related to the same k are concatenated, creating B^k for all k e {l, ... , K} . Following these calculations, M^k and B^k may both be used to train KSVM^k . The trained KSVM classifiers,

{KSVM_t, ... , KSVM_K] , may then be stored in the atlas database .

[0058] The above described atlas database is applied with the automated trauma diagnosis system to segment the organ of interest and diagnose trauma (Figure 7) . The segmenting of organs in the input 3-D ultrasound images involves a number of steps, as follows: i. Reducing speckle noise from the training set of 3-D ultrasound volumes and removing intensity inhomogeneity field using a bias correction approach;

ii . Registering the kidney in the input 3-D ultrasound image on the kidney in the reference volume (stored in the atlas

database), using a rigid transformation; iii . Extracting features as 3-D volumes from the registered volume using Gabor filters; iv . Extracting sub-volumes from 3-D feature volumes ;

v. Using trained KSVM in the atlas database to classify voxels into kidney (ones) and non-kidney (zeros) voxels;

vi . Initializing a deformable model with the kidney average shape model (stored in the atlas database) ;

vii . Concatenating classified sub-volumes to form a binarized mask;

viii . Initializing a level-set function by the generated kidney average shape model;

ix . Deforming the level-set function via a rigid transformation to fit the level-set function inside the binarized mask;

x. Applying level-set propagation to

accurately segment kidneys .

xi . Fitting the deformable model on the

classification result of step (6),

providing the final segmented kidney.

[0059] Steps iii to vii above are illustrated schematically in Figure 6.

[0060] To segment the organ, first, the ultrasound volumes, Vⁱⁿ , are enhanced by reducing speckle noise. This may be effectuated by an anisotropic diffusion filter. Bias correction is then performed to remove intensity inhomogeneity . The enhanced volume is noted by . [0061] Second, the misalignment of a kidney in an input ultrasound volume with respect to the kidney in the reference volume is removed. This may be performed by applying an automated rigid-body registration. Since the trained classifiers are spatially distributed, a misalignment in the kidney position results in failing to correctly detect voxels pertinent to the kidney tissue. For this purpose, a landmark-based

registration cannot be used as kidney segmentation should preferably be automated. Thus, an optimal direction strategy must be applied to modify the rigid-body registration parameters .

[0062] The rigid-body registration uses an affine

transformation with several parameters. For example, the affine transformation may include 7 parameters such as 3 translations, 3 rotations, and 1 scaling.

The parameters vector, 9^lter, at each iteration would then consist of θ_η e {t_x, t_y, t_z, θ_χ, 9_y, θ_ζ, s} where ne{l,...,7}. A parameters vector is then initialized with 9° =

[0,0,0,0,0,0,1]^T and is iteratively modified to maximize the fitness of the reference volume, ^', with the input ultrasound volume, . At each iteration, each parameter, θ_η, is modified as follows,

Equation (16)

where ||-||i is the norm-1, and δ_η is the changing step of the n^th parameter of the affine transformation, a adjusts the updating weight of parameters. The iterative process may continue until there are no further improvements. The registered input volume is referred to as V_r ^l _g , and the calculated registration parameters are saved as 9^_eg .

3-D Gabor wavelets may then be used to extract features from V_R ^L^_GR for example {F ³, ... , F^_Q]■ Sub-volumes may then be extracted from the extracted features, for example, [f₍ ^re5'^fe] where k e {l, ... , K] and I e {1, ...,128} .

The extracted sub-volumes may then be vectorized and concatenated for each sub-volume to generate the features matrix, M^re9,k .

Next, voxels in each sub-volume are classified into kidney and non-kidney tissues using the related trained spatially distributed classifier, KSVM_k . The feature matrix M^rea,k is the input to KSVM_k . The output is a classification result as a vector of 'zero's and 'one's. The vector is reshaped to form a 3-D binarized sub-volume. All binarized sub-volumes may then be combined to generate the 3-D binarized mask, B^rea , as shown in Figure 6 .

Once the binarized mask has been generated, a

deformable model with the average kidney shape model, stored in the kidney atlas database, must be

initialized. A combination of global and local deformations may be applied to finely segment kidneys. An implicit representation, Φ e R , may be used to define the deformable model. For example, <P(x, y, z; t) > 0 specifies voxels inside the segmentation region, <P(x, y, z; t) = 0 specifies voxels on the segmentation boundary, and 0(x, y, z; t) < 0 specifies voxels outside the segmentation region. In this example, <P(x, y, z; t = 0) = 2 x (KASM(x, y, z)— 0.5) must be set for initializing. An affine deformation according to eguation (16) must then be applied to fit <P(x,y,z;t) = 0 on the binarized mask, B^rea .

[0067] The calculated affine deformation to globally fit

0(x,y,z;t) = 0 on the binarized mask, B^rea, is represented by 7 parameters as §?_eg = [t_x, t_y, t_z, θ_χ, 9_y, θ_ζ, s] . The sum of

&reg ^and ®reg is the calculated misalignment of the kidney image with respect to the reference kidney shape, 9_mi_san_gnment 9_reg + Q_reg .

[0068] The affine deformation is a global deformable model used to align the deformable model on the organ of interest such as a kidney (highlighted in the binarized mask). Region-based level-set propagation must then be applied as a local deformable model to finely segment the kidney. Considering the 3-D image domain to be Ω, average intensity levels of kidney to be c_lr and non-kidney regions to be c₀, all calculated based on V_r ^l _g and B^re9, the following regularized function is defined as,

Eguation (17)

t

+ v j Κ( (χ,γ,χ))Λπ ύ

j

n a i where V is the gradient operation, δ and H are Dirac delta and Heaviside functions, respectively, and μ, v, λ and λ₂ are regulation parameters.

[0069] The Euler-Lagrange eguation may be used to minimize J

with respect to Φ, and the following propagation eguation may be obtained,

Eguation (18)

[0070] For organ segmentation, the above eguation must be iteratively repeated until a convergence is achieved. The segmented kidney in the registered volume is defined as V™g = (Φ > 0) . Finally, an inverse affine transformation may be applied to generate the kidney segmentation result in the original input ultrasound volume as V_s% = {AT^V^ .

[0071] To adjust for misalignments of the probe,

misalignments of the kidney in the resulting images have to be related to misalignments of the probe. As noted above, the navigation system will continuously provide the operator with commands to adjust the positioning and/or orientation of the probe to arrive at the desired view of the specific internal area of the patient. Figure 8 schematically illustrates the navigation window that provides the operator with instructions in one embodiment of the invention. As well, Figure 8 also shows that, in this embodiment, the probe is provided with a marker to indicate a specific probe orientation (e.g. a "top" for an initial probe placement) . [0072] To correct for misalignments, it should be noted that the calculated registration parameter Θ is related to the probe misalignment on the patient body. The automated probe navigation system of the present invention may begin by asking the operator to compensate for improper orientation of the ultrasound probe from the initial placement to remove θ_χ by rotating the ultrasound probe. The coordinate system used for the following explanation is illustrated in Figure 9. After removing θ_χ, corrective probe

translation commands can be sent to the operator.

[0073] From the image of the kidney, if the x-axis is

misaligned, it is due to the depth of the kidney inside the patient's body rather than to probe misalignment. Therefore, only t_y and t_z should be considered. t_y represents the misalignment of the ultrasound probe in the direction of mid-auxiliary line toward cephalad and t_z represents the misalignment perpendicular to the mid-auxiliary line. Translation commands are iteratively sent to the operator until the desired view of Morison's pouch is attained.

[0074] The kidney image misalignment, e_misalignment, is obtained in two steps: (1) aligning the enhanced input 3D image on the reference 3D image and (2) aligning the binarized 3D image on the expected alignment of the kidney shape model in the generated atlas database. Step (1) provides a rough alignment of the kidney shape image on the expected kidney alignment, while step (2) performs a fine alignment of the input kidney image on to the reference kidney shape. It should be noted that step (1) is crucial for step (2) to correctly work. If step (1) is not properly done, the KSVM classifiers would not be able to properly generate the binarized volume.

[0075] For clarity, it should be clear that the probe

navigation system always tracks on the calculated kidney image misalignment and stops probe navigation when the misalignment falls within a predetermined acceptable range. The kidney image misalignment,

^misalignment ⁼ [tx> ty> tz> x> y> z> ^s] r i-^s recalculated after each time the operator moves and/or rotates the ultrasound probe. If the translation and orientation parameters, including t_x, t_y, t_z, θ_χ, 9_y and θ_ζ , are smaller than their related threshold values, the navigation system concludes that the probe is aligned within an acceptable range from the reference kidney alignment and the navigation process stops. The threshold values for the translation and orientation parameters are set a ^s [^Tt_x = ¹⁰P^X> ^Tt_y = ¹⁰P^X> ^Tt_z = lOpx, Τ_θχ = 5^°, Tg_y = 5^°, Τ_θζ = 5^°] .

[0076] In another aspect of the invention, there is provided a system for automated organ detection using

ultrasound images. The system includes an ultrasound imaging sub-system which has a probe for use in gathering the ultrasound images. As well, the system includes at least one processor for processing the gathered images and for analyzing these images. Finally, the system includes an organ atlas database that has reference images of the organ, trained classifiers which determine if portions of the image are organ tissue or not, as well as an average organ shape model. These reference images, classifiers, and average organ shape model are used by the system to detect the presence (partial or full) of the organ in the gathered images . The database can be stored in any suitable data storage device or system such as a hard drive, solid state drive, or an online storage system .

[0077] Various aspects of the invention may be implemented as a built-in package in a 3D ultrasound machine or as an add-on package installed on a personal computer connected to a 3D ultrasound imaging device. As an add-on package, any 3D ultrasound machine which supports the ultrasound research interface (URI) or any other protocol exporting ultrasound raw data may be used with the invention. Examples of suitable ultrasound imaging devices for use with the various aspects of the invention include the Siemens SONOLINE Antares ultrasound system and the Hitachi HiVision 5500.

[0078] To better understand the invention, the following

references may be consulted. These references are hereby incorporated by reference in their entirety.

[1] S. Stergiopoulos , Advanced Signal Processing Handbook: Theory and Implementation for Radar, Sonar, and Medical Imaging Real-Time Systems, CRC Press, Inc, 2009.

[2] S. Stergiopoulos and P. Shek, "portable 4d ultrasound diagnostic imaging system, " in IEEE UFFCS, 2011.

[3] S. Stergiopoulos and A. Dhanantwari, "High Resolution 3D Ultrasound Imaging System Deploying a Multi-Dimensional Array of Sensors and Method for Multi-Dimensional Beamforming Sensor Signals". US Patent 6,719,696, 13 April 2004. [4] S. Stergiopoulos and A. Dhanantwari, "High

Resolution 3D Ultrasound Imaging System Deploying a Multi-Dimensional Array of Sensors and Method for Multi-Dimensional Beamforming Sensor Signals". US Patent 6,482,160, 19 November 2002.

[5] G. M. Treece, "Volume measurement and surface visualisation in seguential freehand 3D ultrasound, " Doctoral dissertation, Ph. D. Thesis, Cambridge

University, 2000.

[6] Y. Yu and S. Acton, "Speckle reducing anisotropic diffusion, " IEEE Transactions on Image Processing, vol. 11, no. 11, pp. 1260-1270, 2002.

[7] A. Belaid, D. Boukerroui, Y. Maingourd and L. J. F., "Implicit active contours for ultrasound images segmentation driven by phass information and local maximum likelihood.," in IEEE International Symposium on Biomedical Imaging: From Nano to Macro, 2011.

[8] M. Martin-Fernandez and C. Alberola-Lopez , "An approach for contour detection of human kidneys from ultrasound images using Markov random fields and active contours," Medical image analysis , vol. 9, no. 1, pp. 1-23, 2005.

[9] R. Prevost, B. Mory, J. Correas, L. D. Cohen and R. Ardon, "Kidney detection and real-time segmentation in 3d contrast-enhanced ultrasound images.," in 9th IEEE International Symposium on Biomedical Imaging (ISBI), 2012.

[10] S. Stergiopoulos, P. Shek, K. Plataniotis and M. M, "Computer Aided Diagnosis for Detecting Abdominal Bleeding with 3D Ultrasound Imaging" . US Patent

Application 14/159744, 21 January 2014. [11] H. Shokoohi, K. S. Boniface and A. Siegel,

"Horizontal subxiphoid landmark optimizes probe placement during the Focused Assessment with

Sonography for Trauma ultrasound exam, " European Journal of Emergency Medicine, pp. 1-5, 201 1.

[12] J. G. Sled, A. P. Zijdenbos and A. C. Evans, "A Nonparametric Method for Bias Correction of Intensity Non-unifomiity in MRI Data, " IEEE Transaction on Medical Imaging, vol. 17, no. 1, pp. 87-97, 1998.

[13] B. S. Manjunath and W. Y. Ma, "Texture Features for Browsing and Retrieval of Image Data, " IEEE

TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, vol. 18, no. 8, pp. 837-842, 1996.

[14] L. Shen and L. Bai, "3D Gabor wavelets for evaluating SPM normalization algorithm, " Medical Image Analysis, vol. 12, pp. 375-383, 2008.

[15] Y. Zhan and D. Shen, "Deformable Segmentation of 3-D Ultrasound Prostate Images Using Statistical Texture Matching Method, " IEEE TRANSACTIONS ON MEDICAL IMAGING, vol. 25, no. 3, pp. 256-272, 2006.

[16] C. J. C. Burges, "A Tutorial on Support Vector Machines for Pattern Recognition, " Data Mining and Knowledge Discovery, vol. 2, pp. 121-167, 1998.

[17] T. F. Chen and L. A. Vese, "Active Contours Without Edges," IEEE TRANSACTIONS ON IMAGE PROCESSING, vol. 10, no. 2, pp. 266-277, 2001. The method steps of the invention may be embodied in sets of executable machine code stored in a variety of formats such as object code or source code. Such code is described generically herein as programming code, or a computer program for simplification. Clearly, the executable machine code may be integrated with the code of other programs, implemented as subroutines, by external program calls or by other technigues as known in the art.

[0080] The embodiments of the invention may be executed by a computer processor or similar device programmed in the manner of method steps, or may be executed by an electronic system which is provided with means for executing these steps . Similarly, an electronic memory means such computer diskettes, CD-ROMs, Random Access Memory (RAM) , Read Only Memory (ROM) or similar computer software storage media known in the art, may be programmed to execute such method steps. As well, electronic signals representing these method steps may also be transmitted via a communication network.

[0081] Embodiments of the invention may be implemented in any conventional computer programming language. For example, preferred embodiments may be implemented in a procedural programming language (e.g."C") or an object oriented language (e.g. "C++")· Alternative embodiments of the invention may be implemented as pre-programmed hardware elements, other related components, or as a combination of hardware and software components.

Embodiments can be implemented as a computer program product for use with a computer system. Such

implementations may include a series of computer instructions fixed either on a tangible medium, such as a computer readable medium (e.g., a diskette, CD- ROM, ROM, or fixed disk) or transmittable to a computer system, via a modem or other interface device, such as a communications adapter connected to a network over a medium. The medium may be either a tangible medium (e.g., optical or electrical

communications lines) or a medium implemented with wireless techniques (e.g., microwave, infrared or other transmission techniques). The series of computer instructions embodies all or part of the functionality previously described herein. Those skilled in the art should appreciate that such computer instructions can be written in a number of programming languages for use with many computer architectures or operating systems. Furthermore, such instructions may be stored in any memory device, such as semiconductor, magnetic, optical or other memory devices, and may be

transmitted using any communications technology, such as optical, infrared, microwave, or other transmission technologies. It is expected that such a computer program product may be distributed as a removable medium with accompanying printed or electronic documentation (e.g., shrink wrapped software), preloaded with a computer system (e.g., on system ROM or fixed disk), or distributed from a server over the network (e.g., the Internet or World Wide Web) . Of course, some embodiments of the invention may be implemented as a combination of both software (e.g., a computer program product) and hardware. Still other embodiments of the invention may be implemented as entirely hardware, or entirely software (e.g., a computer program product) . Embodiments of the invention may also be implemented using sequential or parallelized programming methods. Parallelized implementations of the methods and processes of the invention may be used with multi-core processors, such as Intel Xeon processors or Intel Extreme processors, or they can be implemented using GPU processors based on CUDA language. In one variant, the generated software package may be used either as a solution built into a 3D ultrasound imaging device or as an installable or executable addon software on a personal computer (including desktop workstation or laptop) connected to a 3D ultrasound imaging device supporting the ultrasound research interface or any other protocol to transfer ultrasound raw data. A person understanding this invention may now conceive of alternative structures and embodiments or

variations of the above all of which are intended to fall within the scope of the invention as defined in the claims that follow.

Claims

What is claimed is:

1. A method for detecting a presence of trauma in a patient, the method comprising: a) providing an operator with an indication of a location on said patient for an initial placement of a probe ; b) obtaining images of an internal area of said patient using said probe; c) determining if at least one specific organ is present in images obtained in step b) ; d) in the event said at least one specific organ is not present in said images, repeating steps a) - c) until said at least one specific organ is present in said images ; e) determining an alignment of said at least one specific organ in said images; f) determining if a desired view of said internal area is found in said images; g) in the event said desired view is not found in said images, relating said alignment of said at least one specific organ to an alignment of said probe; h) determining corrective instructions for probe

placement based on results of step g) , said corrective instructions being operative to adjust a view of said internal area to thereby provide said operator with a view with a correct alignment of said at least specific organ; i) sending said corrective instructions to said operator; j ) obtaining images of said internal area of said patient using said probe; k) determining if at least one specific organ is present in images obtained in step j ) ;

1) repeating steps e) - k) until said desired view is found in said images.

2. The method according to claim 1, wherein said at least one organ comprises a kidney.

3. The method according to claim 1, wherein said probe is an ultrasound probe.

4. The method according to claim 3, wherein said images are ultrasound images .

5. The method according to claim 4, wherein said images are 3-D ultrasound images.

6. The method according to claim 1, wherein step e) includes the steps of :

- reducing noise in said image; and

- correcting for intensity bias in said image.

7. The method according to claim 1, wherein step e) is executed by referring to an organ atlas database, said database containing at least one average organ shape model.

8. The method according to claim 7, wherein said at least one average organ shape model is used in at least one of steps c) and e ) .

9. The method according to claim 1, wherein step c) includes using a global deformation to detect said organ in said images .

10. The method according to claim 9, wherein said global deformation is based on an affine registration of an organ in said image to a reference image in said database.

11. The method according to claim 1, wherein step c) includes using a local deformation to detect said organ in said images.

12. The method according to claim 11, wherein said local deformation is based on a region-based level-set method for deforming an organ in said image to register with a reference image in said database .

13. The method according to claim 1, wherein said method uses at least one of :

- 3-D Gabor wavelets as 3-D filters for extracting features of said organ in said images;

- localized classification of portions of said images to detect said organ; and

- spatially distributed classifiers to classify portions of said images to detect said organ.

14. A method for compiling a database for use in detecting a presence of a specific organ in an ultrasound image, the method comprising: a) pre-processing at least one training image; b) manually segmenting an organ in said at least one training image; c) generating an average shape model for said organ based on said at least one training image and said organ segmented in step b) ; d) extracting texture features from said organ segmented in step b) ; e) training kernel based support vector machines to classify portions of images as being organ or non-organ based on said at least one training image; f) storing said kernel based support vector machines and said organ average shape model in said database .

15. The method according to claim 14, wherein said database further contains at least one of :

- a reference image of said organ; and

- at least one segmented organ in said at least one training image.

16. The method according to claim 14, wherein step a) includes reducing noise in said at least one training image.

17. The method according to claim 14, wherein step a) includes applying a bias correction to remove intensity inhomogeneity in said at least one training image.

18. The method according to claim 14, further including a step of registering said at least one training image to a reference image of said organ.

19. The method according to claim 17, wherein said step of registering uses a landmark-based rigid-body registration.

20. A system for detecting an organ in at least one

ultrasound image of an internal area of a patient, the system comprising :

- processor for processing said at least one ultrasound image ;

- data storage containing a database, said database including : - a reference image of said organ;

- at least one training image containing at least one segmented organ;

- at least one kernel based support vector machine, said support vector machine being for classifying portions of said at least one image as being organ or non-organ based on said at least one training image; and

- an average shape model for said organ based on said at least one training image.