Ultrasonic method

Abstract

A computer-implemented method for automatically identifying a reference blood vessel in an ultrasound image, the method comprising: segmenting the image to identify an organ; localizing a predetermined anatomical landmark specific to the organ; identifying a vasculature suitable for the organ; and selecting a reference blood vessel from the vasculature having a size within a predetermined size range and a position within a predetermined distance range from the predetermined anatomical landmark.

Description

Technical Field

[0001] The present invention relates to ultrasonic imaging and measurement, and in particular to the use of power Doppler ultrasound technology to measure blood flow. Background Art

[0002] Power Doppler ultrasound is a technique that uses the amplitude of the reflected signal as a function of the frequency shift of the transmitted signal to determine the movement of tissue, particularly blood flow, in a subject. Power Doppler (PD) ultrasound is quantified by integrating the amplitude of the ultrasound scatter within an area of interest (2D-PD) or volume (3D-PD). This is achieved by summing the power Doppler signals within the region of interest, which PD signals represent the local concentration of red blood cells (which scatter ultrasound). Typically, power Doppler ultrasound can only provide information about relative blood flow because the amount of attenuation of the echo returned between the measurement site and the sensor is unknown. The attenuation depends on the distance between the measurement site and the sensor and the nature of the intervening tissue, for example, scar tissue attenuates strongly. Knowing absolute blood flow would improve the diagnosis of various diseases. Known techniques for normalizing power Doppler signals involve manually identifying a region of interest to be used as a normalization reference to determine the normalization point. The normalization point used is the "knee value", where power Doppler intensity values above the normalization point are assigned a value of 1, and values below this normalization point are scaled to the range of 0 to 1 [1]. Summary of the Invention

[0003] There is a need for a technique capable of making an absolute measurement of blood flow.

[0004] According to the present invention, there is provided a computer-implemented method for automatically identifying a reference blood vessel in an ultrasound image, the method comprising:

[0005] Segmenting the image to identify an organ;

[0006] Locating a predetermined anatomical landmark specific to the organ;

[0007] Identifying a vasculature suitable for the organ; and

[0008] Selecting a reference blood vessel from the vasculature having a size within a predetermined size range and located at a predetermined distance range from the predetermined anatomical landmark.

[0009] Ideally, the diameter of the predetermined size range is greater than 3 mm.

[0010] Ideally, the segmentation includes using a trained fully convolutional neural network.

[0011] Ideally, the locating includes using a multi-class transfer learning model.

[0012] Ideally, the multi-class transfer learning model includes two independent paths: a first path and a second path. The first path has parameters initialized using a segmentation model, and the second path has parameters initialized using variance scaling.

[0013] Ideally, the identification includes using a method based on automated multi-seed region growing.

[0014] Ideally, the identification further includes using a 3D medial axis-based refinement method on the output of the method based on automated multi-seed region growing.

[0015] Ideally, the organ is the placenta, and preferably, the anatomical landmark is the uterine-placental interface. In this case, ideally, the predetermined distance range is from about 0.5 cm to about 1.5 cm.

[0016] According to the present invention, there is provided a computer-implemented method for mapping blood flow in a power Doppler image of an organ, the method comprising:

[0017] Identifying a reference vessel in the power Doppler image using the above method;

[0018] Determining a reference power Doppler value based on the reference vessel; and

[0019] Scaling the power Doppler values in the power Doppler image based on the reference power Doppler value to obtain a standardized blood flow image.

[0020] Ideally, determining the reference power Doppler value includes fitting a model vessel distribution to the reference vessel and determining standardized points that do not fall within the high-shear vessel margins.

[0021] Ideally, fitting the model vessel distribution includes detecting local maximum intensity points and using a region generation method to obtain an initial vessel region where all internal intensities are greater than a threshold.

[0022] Ideally, applying an iterative gradient descent technique to minimize a cost function.

[0023] According to the present invention, there is provided a computer-implemented method for determining the moving blood volume fraction in an organ, the method comprising:

[0024] Obtaining a power Doppler image of the organ;

[0025] Mapping blood flow in the power Doppler image using the above method to obtain a standardized blood flow image; and

[0026] Determining the moving blood volume fraction using the standardized blood flow image.

[0027] According to the present invention, there is provided a computer-implemented method for calculating the risk of adverse pregnancy outcomes (such as fetal growth restriction and / or preeclampsia), comprising:

[0028] Determining a fractional moving blood volume in the placenta of a subject according to the above method;

[0029] Measuring the placenta volume of the subject; and

[0030] Calculating a risk score based at least in part on the fractional moving blood volume and the placenta volume.

[0031] Thus, the present invention standardizes the absolute signal, so that the signal integration on the region of interest is always baselined, thereby realizing a fully automatic comparison of quantitative blood flow estimation between different patients with different tissue attenuations. For this purpose, large blood vessels with "100% vascular distribution" are identified at a tissue depth level similar to the region of interest, and the numerical value recorded by the PD signal in this blood vessel is used as the "standardization point" [2]. If the signal in this large blood vessel is divided by the PD signal of the entire region of interest, then other smaller blood vessels in this region of interest will always have the same proportional signal intensity related to the blood vessel with 100% vascular distribution. This standardization process produces a measurement method called fractional moving blood volume (FMBV), which is the only effective method for quantitatively measuring perfusion using 2D and 3D ultrasound.

[0032] The present invention provides a new method for FMBV estimation, which overcomes the limitations of the prior art and can improve the accuracy of this technology. For example, the present invention avoids the difficulties caused by defining the region of interest for analysis, where the standardized value is a function of the size of the region of interest and the nature of the tissue included in this region of interest. In addition, the technology of the present invention is not easily affected by the signal-to-noise ratio in the acquired images; in many cases, it is extremely insensitive to noise. In addition, the present invention is not easily affected by the position of medium-sized blood vessels in the region of interest on the "knee point" and the subsequently calculated standardized value.

[0033] The present invention is capable of automatically identifying the power Doppler standardized value from a single blood vessel to adjust the tissue attenuation of the ultrasound signal. Embodiments of the present invention provide a robust and effective method for automatically identifying the power Doppler standardized value from a single blood vessel extracted from a complex blood vessel network imaged by standard 3D power Doppler ultrasound. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] The present invention will be further described below with reference to exemplary embodiments and the drawings, in which:

[0035] Figure 1 Depicts the architecture of a fully convolutional neural network (fCNN) that can be used in embodiments of the present invention.

[0036] Figure 2 Depicts example images of placental segmentation using a 2D B-mode plane (left figure), placental segmentation using semi-automatic random walk segmentation (middle figure), and placental segmentation using Figure 1 segmentation by the fCNN (right figure).

[0037] Figure 3 Shows a graph comparing the placental volume calculated by the Figure 1 fCNN with the placental volume calculated by the semi-automatic random walk technique and the placental volume calculated by a commercially available VOCAL TM technique.

[0038] Figure 4 Depicts how to identify the utero-placental interface (UPI) once the amniotic-placental interface (API) and the fetal-placental interface (FPI) are known.

[0039] Figures 5(a) to 5(c) Depicts the fully convolutional neural network architecture of the model used in embodiments of the present invention. Specifically, Fig. 5(a) is a placenta segmentation (PS) model; Fig. 5(b) is a hybrid (HB) model; and Fig. 5(c) is a symbolic description.

[0040] Fig. 6(a) depicts an example of the skeletonization of a vascular tree and the vascular centerline showing the cross-sectional distribution, and Fig. 6(b) depicts a vascular tree with separated vascular segments.

[0041] Fig. 7(a) depicts flash artifacts, and Fig. 7(b) depicts the average Doppler signal intensity on the scan plane.

[0042] Fig. 8(a) depicts a B-mode scan covered by a power Doppler signal showing periodic noise, Fig. 8(b) is a side view, and Fig. 8(c) depicts a linear kernel.

[0043] Figure 9 Depicts the 3D fast Fourier transform of 3D-US data showing periodic noise.

[0044] Figure 10 Is a flowchart of the method according to the present invention.

[0045] In the various figures, like parts are designated by like reference numerals. DETAILED DESCRIPTION

[0046] Theory: Power Doppler signal distribution in blood vessels

[0047] Jannson et al. pointed out that in a phantom model containing scatterers, within a certain power range, the Doppler power is proportional to the velocity, and they concluded that this may apply to a range of machines and setups [3]. In any elliptical cross-section of a blood vessel, we expect the distribution of the flow velocity or intensity I to have the following shape:

[0048]

[0049] where a and b are the major-axis radius and minor-axis radius of the blood vessel cross-section, x and y are the Cartesian coordinates that define the major axis and minor axis of the blood vessel, and I max is the true (noise-free) maximum intensity within the blood vessel, γ is a parameter for "flattening" the velocity distribution, γ = 2 represents the classical Poiseuille flow distribution in a pipe with an ellipsoidal cross-section, and increasing γ > 2 enables the distribution to be flattened due to the pulsatile nature of the blood flow (especially in large arteries) [4].

[0050] The greatest risk factor for stillbirth is fetal growth restriction (FGR), which is usually secondary to poor placental implantation. Unfortunately, currently available methods for allocating FGR risk perform very poorly, and many women who are considered "low risk" do not realize that their babies are not growing properly until stillbirth occurs. If classified as "high risk", women will undergo a series of growth ultrasound scans to identify babies that fail to develop, with the aim of delivering before stillbirth. The current risk assessment methods perform so poorly that many women who are considered "high risk" give birth to well-developed, healthy babies after pregnancy, under a cloud of unnecessary stress and anxiety.

[0051] There is an urgent need for a robust and reliable early pregnancy screening method to assess the risk of FGR. This would mean that women who are actually at high risk of stillbirth secondary to FGR would undergo the available serial growth ultrasound scans.

[0052] Early pregnancy screening test

[0053] It has long been known that placental volume (PlVol) in the early pregnancy is related to the full-term birth weight, and it was proposed as early as 1981 that PlVol measured by B-ultrasound could be used to screen for FGR. Since then, many studies have shown that low PlVol between 11 and 13 weeks of pregnancy can predict adverse pregnancy outcomes, including small for gestational age (SGA, which is an alternative marker for FGR) and preeclampsia. Since PlVol has also been shown to be independent of other biomarkers, such as pregnancy associated plasma protein A (PAPP-A) and nuchal translucency (NT), a recent systematic review concluded that it could be successfully integrated into future multivariate screening methods for FGR, similar to the "combined test" currently used to screen for fetal aneuploidies (including Down syndrome) in the first trimester of pregnancy. Early pilot data also suggest that the combination of PlVol and estimates of placental perfusion may be able to distinguish different adverse pregnancy outcomes, with low PlVol but normal vascular distribution being associated with an increased risk of FGR, while low PlVol and low vascular distribution increase the risk of preeclampsia.

[0054] Although there is reliable background data to support their value, in order for PlVol and placental vascular distribution to become useful imaging biomarkers, a reliable, real-time, operator-independent technique is needed to estimate them.

[0055] So far, the only way to estimate PlVol is by manually outlining the placenta in three-dimensional ultrasound (3D-US) (drawing around the placenta in each 2D slice that makes up the 3D volume) or using semi-automatic tools such as VOCAL TM (GE, USA) to label the placenta. VOCAL TM requires the operator to "draw" on several, but not all, 2D slices around the placenta and then uses a rotational interpolation algorithm to estimate the final volume. Both of these strategies are too time-consuming and operator-dependent to be part of a population-based screening test and neither provides any other morphological metrics, such as shape or surface area. For VOCAL TM there are also significant controversies regarding the reproducibility and repeatability of the volume results generated based on the number of distributions used (degree of interpolation) and the variability of the organ of interest.

[0056] There is currently no product that can automatically measure validated vascular metrics, such as fractional moving blood volume (FMBV).

[0057] The present inventors have developed and validated a new fully convolutional neural network (OxNNet) [5] (see https: / / insight.jci.org / articles / view / 120178 ), which is capable of automatically identifying the placental volume from 3D-US. As further discussed below, the present invention provides further development to enable automatic identification of relevant anatomical landmarks, thereby facilitating fully automated vascular measurements.

[0058] To develop OxNNet, the placenta was segmented from 2393 early pregnancy 3D-US volumes using the gold standard semi-automatic random walk technique [6]. This was quality controlled by three operators to generate a "ground truth" dataset. A fully convolutional neural network (fCNN) called OxNNet was created using the TensorFlow (v1.3) framework and a 3D architecture inspired by 2D U-net [7] (see Figure 1 ). The effect of training set size on model performance was investigated by training with samples of 100, 150, 300, 600, 900, and 1200 cases for 25,000 iterations.

[0059] To evaluate PlVol segmentation, 1097 placentas were used as a training set with 100 validation images for 2-fold cross-validation. OxNNet was trained 8 times and ran for 26 hours. Then 1196 cases were tested. After training, it finally took 11 seconds on average to calculate PlVol.

[0060] Each predicted segmentation was post-processed to remove disconnected parts with a maximum region volume < 40%. The segmentation was binary dilated and eroded using a 3D kernel with a radius of three voxels and an applied hole filling filter. These methods smoothed the placental boundary and filled any holes surrounded by placental tissue (cystic lesions occur in normal placentas and should be included in any PlVol estimate).

[0061] Compared with the "ground truth" dataset, OxNNet provides a state-of-the-art fully automated segmentation technique (median Dice similarity coefficient (DSC) of 0.84: see Figure 2) The impact of training set size on the performance of OxNNet indicates the need for large datasets (n > 1000). This is important because, to date, most imaging tools generated using machine learning have used relatively small datasets. The clinical utility of the PlVol estimates generated by OxNNet was tested by observing the prediction of small-for-gestational-age (SGA) infants (a surrogate marker for FGR). The receiver-operating characteristic curve showed that the prediction of PlVol estimates for SGA infants by OxNNet (AUC 0.65 (95% CI; 0.61 - 0.69)) and the "ground truth" generated by the semi-automatic random walk (RW) technique (AUC 0.65 (95% CI; 0.61 - 0.69): see Figure 3 the left figure in the middle) had almost the same results. OxNNet performed better than the only commercially available tool, VOCAL TM (AUC 0.57 (95% CI; 0.48 - 0.65: see Figure 3 the right figure in the middle).

[0062] To examine the vascular distribution of the placental bed, the present invention proposes to identify the uterine-placental interface (UPI) as an anatomical reference point (see Figure 4 ). OxNNet can successfully segment the entire placenta, but the present invention proposes to separate two different placental surfaces; the UPI - where the placenta implants into the uterus - and the amnio-placental interface (API) - where the placenta contacts the amniotic fluid (the amniotic fluid fills the sac containing the fetus; see Figure 4 ). Occasionally, the fetus contacts the placenta, so the fetal-placental interface (FPI) must also be identified and included in the API.

[0063] Embodiments of the present invention use machine learning to develop a robust multi-class segmentation technique to identify the placenta, fetus, and amniotic fluid from 3D B-mode ultrasound scans, thereby facilitating the identification of automatic anatomical landmarks.

[0064] From the available 2393 3D-US volumes that adopted existing placenta segmentation, 300 volumes were randomly selected for multi-class segmentation. The amniotic fluid and the fetus were manually seeded by two operators and combined with the placenta seeding performed in previous studies. The amniotic fluid and the fetus have more distinct boundaries than the placenta, so the initialization is much easier, but any cases with ambiguity were examined by clinical experts with extensive experience in obstetric ultrasound. Then, the three different classes were segmented into a multi-class label map using the gold-standard random walk algorithm.

[0065] 300 multi-class (MC) cases were subdivided into 200 training cases, 40 validation cases, and 60 test cases. Four MC models were trained 40 times each with a batch size of 30. First, an fCNN identical to the original OxNNet architecture was used to train the MC models (see Fig. 5(a)). Then, a multi-class transfer learning (MCTL) model was trained using the same architecture but with weights and biases taken from the original OxNNet training (effectively “transferring” what the OxNNet “learned” from a larger training dataset of 1097 cases rather than the 200 cases of the simple MC model).

[0066] To overcome any potential degradation in the performance of the other two targets (fetus and amniotic fluid), two hybrid models were developed, namely the hybrid model (HB) and the hybrid model with exponential averaging (HBAV). Since not all weights need to be stored, exponential moving average reduces noise by averaging the model's weights during training, favoring more recent values of the weights, and providing computational efficiency. Both hybrid models consist of two independent paths, as shown in Fig. 5(b). The architectures of the multi-class (MC) and multi-class transfer learning (MCTL) models are not shown here as they differ from the placenta segmentation (PS) model architecture in having four channels instead of two channels.

[0067] Both the hybrid model (HB) and the hybrid model with exponential averaging (HBAV) consist of two independent paths, as shown in Fig. 5(b). In the upper path, the HB parameters are initialized using the PS model, and the HB parameters of the HBAV model are initialized using the PSAV model. In the lower layer, the parameters are initialized using variance scaling and then trained on the MC data. The parameters in the lower path are allowed to vary and are fixed in the upper path for the HB and HBAV models. The parameters of the Adam optimizer are the same as those used in the PS training.

[0068] Let P Background ,P Placenta ,P Amniotic and P Fetus be the confidences that a voxel belongs to background, placenta, amnion, and fetus, respectively. For a given voxel i, the softmax output of the upper path (T) gives only two values and whose sum is 1. In this case, the fetus and amniotic fluid are included in the background. In the lower (B) path, the softmax of the last layer produces the confidences of the membership of the given voxel, with scalar values of and whose sum is 1. Represents the confidence that the voxel is not the placenta, fetus, or amniotic fluid. Combining the outputs of the two paths defines the loss function L as:

[0069] L = ∑ i∈M m i × sl(o i / n i ) (2)

[0070] where M is a binary mask, for voxel i within the placenta ground truth region, m i is 0, and outside this region m i is 1, o i is the output of the lower path, sl is the softmax cross-entropy function, and n i is the normalization factor, defined as 1 minus the output of the softmax layer of the upper path.

[0071] Since the loss function is masked in the placenta region, the placenta does not contribute to the training of the model. The final confidence vector in the HB model has scalar components as follows:

[0072]

[0073]

[0074]

[0075]

[0076] where the final segmentation of the voxel is taken as the maximum of the values defined in equations (3) to (6). Thus, for all voxels of , the placenta segmentation of the HB model will be the same as that of the PS model. The HB model classifies the voxels of as the placenta, while the PS model classifies them as background, and for the remaining classes, and both have the value of .

[0077] The parameters of the optimizer are the same as those used in the original OxNNet training. Although the exponentially averaged mixture model is the best, all four models provide useful results. Using these models, all possible placental morphometric measurements related to anatomical landmarks can be automatically identified, such as the location and surface area of the UPI, which makes it possible to map the vascular distribution of the placental bed, thus improving the prediction potential for adverse pregnancy outcomes such as preeclampsia.

[0078] After identifying the uterine-placental interface (UPI), the next step is to evaluate blood flow in the placenta.

[0079] Doppler ultrasound is considered the primary non-invasive imaging modality performed during pregnancy to evaluate blood flow within the placenta and fetal organs. Compared to color Doppler, power Doppler (PD) ultrasound is less affected by the ultrasound angle and is more sensitive to multiple flow directions and low flow velocities, and thus it is more helpful in evaluating tissue perfusion. However, like all forms of ultrasound, PD ultrasound signals are attenuated by the tissue through which they pass. To compensate for attenuation and allow direct inter-patient comparison, it is necessary to standardize the power Doppler signal. The only valid method for estimating tissue perfusion from PD ultrasound is the fractional moving blood volume (FMBV). This was recently validated in 3D against the radioactive gold standard in microspheres in a porcine model [8]. Using 3D-FMBV, the inventors have demonstrated in a small pilot study (n = 143) that perfusion in the early pregnancy placental bed is significantly reduced for pregnancies that develop preeclampsia (p = 0.03), but not for normotensive pregnancies that result in small-for-gestational-age (SGA) infants (p = 0.16) [9]. This is consistent with the classical histopathological findings of preeclampsia with poor spiral artery adaptation.

[0080] To standardize the power Doppler signal to determine 3D-FMBV, it is proposed to identify a large reference vessel close to the target of interest, which can be assumed to have 100% vascular distribution. Then, the PD signal within the reference vessel is used to define a normalization value that corrects the PD signal value recorded in the target, thus adjusting for individual signal attenuation and allowing appropriate inter-patient comparison. The present invention aims to fully automate the estimation of 3D-FMBV, thereby facilitating large-scale testing of this finding and potentially improving the predictive ability of early pregnancy screening tests. The present invention can also be applied to the measurement of vascular distribution in other organs and / or tumors, particularly using transfer learning to minimize the training required for other organs or tumors.

[0081] Now described is a fully automated technique for mapping and measuring the uterine-placental vasculature from 3D-PD ultrasound scans, which already has the automatically identified anatomical landmarks described above. This avoids defining regions of interest that may vary between studies and instead focuses on identifying consistent, automatically identifiable large blood vessels for standardization. Additionally, when the process identifies and analyzes individual blood-filled vessels for standardization, an algorithm is provided that is highly consistent within the typical signal-to-noise ratio range of power Doppler and does not suffer from variability in standardization due to the presence of perfused soft tissue or medium-sized blood vessels in the surrounding area of interest.

[0082] First, a method based on automatic multi-seed region growing is used to extract the entire uterine vasculature. This method is chosen because it is superior to the commonly used threshold segmentation method as it can preserve the presentation of vascular connections. Then, a thinning method based on the 3D medial axis

[10] is employed to extract the myometrial vascular plexus skeleton (see Fig. 6(a)). Then, the vascular skeleton is further separated into several segments by removing the bifurcation points. These central line segments are used as markers to apply a 3D watershed algorithm

[11] to divide the vascular network into vascular segments (Fig. 6(b)). A local vascular cross-section is drawn for each segment, which is the intersection surface between the vascular segment body and the plane perpendicular to the local vascular centerline, to determine the vascular diameter (Fig. 6(a)). Vascular segments showing a unimodal intensity in the cross-section are identified as individual vessels, which are considered candidates for reference vessels. The selected reference vessel is a vessel with a maximum diameter > 3 mm within the volume of interest. Optionally, any single vessel with a diameter greater than a predetermined threshold (e.g., 3 mm) can be selected.

[0083] By combining this vascular identification technique with the gray-scale OxNNet tool that can identify anatomical landmarks in the placenta including the UPI, an appropriate reference vessel in the placental bed located at a predefined distance from the UPI can be selected. The predefined distance can be about 1 cm, e.g., between 0.5 cm and 1.5 cm. If there are multiple vessels meeting the criteria, the largest one can be selected. Using this reference vessel, a fully automatic estimation of the 3D-FMBV (perfusion) of the entire UPI or any volume of interest relative to the UPI can be calculated using the new method.

[0084] Using a vessel with a known 100% vascular distribution to estimate the 3D-FMBV is a way to standardize the values of power Doppler signals affected by tissue attenuation. Performing this standardization enables quantitative inter-patient comparison as it corrects for the depth of the target organ and the type of tissue injected (e.g., corrects for the amount of fat).

[0085] The calculation of the standardized value is independent of the vascular size above 3 mm but applies to the shape of the vascular intensity distribution. Its accuracy is demonstrated by comparison with synthetic data. There is currently no technique for manually or automatically generating standardized points from individual vessels.

[0086] In an embodiment, the standardized point indicating the "constant" value of the blood flow velocity is determined by the automatic identification of the high-shear vascular edge of an individual vessel based on vascular characteristics (intensity distribution and diameter) and hydrodynamic simulation. A shear threshold can be applied. This enables an accurate estimation of the digital standardized value each time. This will be described in detail below.

[0087] Localizing the Vascular Boundary

[0088] It is observed that the intensity of the blood vessel boundary in the PD ultrasound image is low (close to the background intensity), so the contrast is low. Threshold segmentation, as an easy-to-operate method, is usually used to segment blood vessels. However, due to the low-contrast boundary and the inevitable noise generated during the imaging process, simple threshold segmentation will not provide accurate blood vessel segmentation. It has been demonstrated that the polar active contour method performs well in the boundary detection of other forms of objects with generally convex shapes

[12]

[13]

[14] . The present invention proposes to use a similar blood vessel segmentation method to achieve robust and highly accurate blood vessel segmentation.

[0089] The active contour deforms the object boundary to minimize the defined energy function in order to obtain the segmentation of the object

[14] . Let S be a given image defined in the Ω domain, which contains the object blood vessels, and C = {χ|φ(χ) = 0} is a closed contour. φ is the signed distance function. The interior of C (the candidate blood vessel region) is defined by approximating the following smooth Heaviside function:

[0090]

[0091] The exterior of C, that is, the region outside the target blood vessel, is defined as Let E(C) be the energy function. The expected blood vessel boundary is the contour C that minimizes the energy function.

[0092] In the polar coordinate system, the contour samples the boundary at multiple angles. Then the polar contour only evolves radially. To describe the energy in the polar coordinate system, we define the following characteristic function:

[0093]

[0094] where the blood vessel center is regarded as the pole, χ and p are two points, and θ(·) is the angular coordinate of a point in the polar coordinate system. The function is 1 only when p and χ are on the same radial line. The function is 1.

[0095] Then, the energy function is expressed as:

[0096]

[0097] where δφ(χ) is the derivative of. The local intensity (radial) is predicted for each point following the best intensity model at that point. According to Equation (1), the blood vessel radial intensity distribution is:

[0098]

[0099] Note that for the detection of uterine blood vessels, which are small in size and observed to have a parabolic intensity distribution, γ is set to 2. By fitting equation (10) to the radial intensity within the vascular contour points χ, we obtain the estimated internal intensity:

[0100]

[0101] The external average intensity is calculated as follows:

[0102]

[0103] Then, the energy is expressed as:

[0104]

[0105] Finally, a regularization term is given This algorithm penalizes the variance of the local curve, smooths the contour, and is weighted by a factor ξ. The final energy is defined as:

[0106] E total (φ) = E(φ) + ξE cur (φ) (14)

[0107] This method can be implemented in three steps: initialization, gradient descent, and detection stop condition.

[0108] For initialization, the local maximum intensity points are detected and regarded as the vascular center. Using this point as a seed, a simple region growing method is applied to obtain an initial vascular region where all internal intensities are greater than the threshold T ini The initial contour C ini is composed of a set of points on the boundary of the initial region.

[0109] The iterative gradient descent technique is iteratively applied to iteratively minimize the defined energy E total (φ) (which is an example of a cost function) until the equilibrium stop condition is met. In each iteration, the vascular center point is updated to the centroid of the vascular region.

[0110] An exemplary equilibrium stop condition is that the variance in the vascular region is very small and less than the threshold T for 10 iterations var .

[0111] Determine the normalization points

[0112] As described above, it has been shown that [3] in a phantom model containing scattering particles, the Doppler power is proportional to the velocity in the power range of 0 dB to 35 dB. Although this depends on the machine settings, the conclusion is that it may be the case for a range of machines and settings. In any elliptical blood vessel cross-section, we can expect the distribution of the flow velocity or intensity I to have the following shape:

[0113]

[0114] where a and b are the major and minor axis radii of the blood vessel cross-section, and x and y are Cartesian coordinates that define the major and minor axes of the blood vessel, and I max is the true (noise-free) maximum intensity within the blood vessel, and γ is a parameter used to "flatten" the velocity distribution. γ = 2 represents the classical Poiseuille flow distribution in a pipe with an ellipsoidal cross-section. An increased γ > 2 enables the distribution to be flattened due to the pulsatile nature of blood flow (especially in large arteries) [5].

[0115] The classical cumulative intensity curve in power Doppler ultrasound imaging calculates the number of pixels / voxels in the region of interest with an intensity less than or equal to a given intensity. An elliptical blood vessel cross-section is mathematically equivalent to the area of the blood vessel with an intensity less than a given value. If we define the coordinate transformation as polar coordinates x = ar cosθ and y = br sinθ, then the definition of the intensity becomes:

[0116] I(x, y) = I max [1 - r γ (16)

[0117] We can calculate the area of the blood vessel where I < I * (equivalent to the area of the elliptical ring where r > r * ) as:

[0118]

[0119] We are interested in the pixel intensity rather than the blood vessel area, so we can substitute Therefore, the area of the blood vessel where I < I * is:

[0120]

[0121] We normalize this value by the total cross-sectional area of the blood vessel to obtain the cumulative probability of the intensity where I < I * as:

[0122]

[0123] In the discrete sense, calculating p(I *)Is exactly the same as calculating the cumulative Doppler power distribution function using the following formula:

[0124]

[0125] where n(i) is the number of pixels with power intensity i, and N total is the total number of pixels in the region of interest, and N(I * ) is the total number of pixels with power intensity < I * .

[0126] Theoretically, if the region of interest is defined as the vascular cross-section, the discrete data can be fitted to this curve to estimate both γ and I max . It should be noted that for a parabolic velocity (or intensity) distribution, γ = 2, which is a straight line and is independent of the aspect ratio of the elliptical vascular cross-section. We note that in many blood vessels within the uterine-placental system, size and hemodynamic considerations imply that γ ≈ 2 (which is also confirmed by plotting the intensity distribution through the center of the vascular cross-section in a power Doppler image). This means that fitting both γ and I max to the data as required to accurately estimate the peak signal in the blood vessel will be error-prone. Therefore, we propose an alternative "weighted" cumulative curve that can accurately and reliably fit γ and I max of the blood vessel even when γ ≈ 2.

[0127] The proposed cumulative curve reflects the variation of the total signal across the vascular cross-section with intensity I < I * , and is defined in a discrete sense as:

[0128]

[0129] where I(i) is the Doppler power intensity at pixel i, and I total is the cumulative total intensity across the entire vascular cross-section.

[0130] In integral form (not normalized by I total ) it is:

[0131]

[0132] Finally, we normalize by to obtain:

[0133]

[0134] The above provides a suitable cumulative curve with which we can fit two parameters, the maximum power intensity I max and the "flatness" γ.

[0135] For blood vessels of any cross-section, a similar polar equation can be derived.

[0136] The normalization point is determined by separating the high-shear vessel edge (with a small number of red blood cells) and the blood vessel itself with a dense concentration of red blood cells using the concept of a "shear threshold". Shear is determined as the gradient of the blood flow velocity and its peak at the vessel wall. Shear is defined as the intensity at X% of its maximum value by calculating the gradient of Equation (15) and replacing "radius" with intensity, i.e.,

[0137]

[0138] For practical implementation, the following steps are performed:

[0139] 1. Calculate the intensity-weighted Doppler power cumulative distribution using Equation (21).

[0140] 2. Fit Equation (23) to the cumulative distribution to provide the best-fit values of γ and I max

[0141] 3. Substitute these best-fit values into Equation (24) to calculate I norm . The appropriate value of X can vary depending on the size and velocity distribution in the normalized vessel (the renal artery is expected to have a flatter velocity distribution than the uterine-placental artery). However, for relatively small vessels such as, for example, the uterine-placental artery, the blood flow distribution is typically parabolic, so X = 85 is an appropriate threshold to keep the vessel cross-section away from the high-shear vessel wall.

[0142] In tests based on synthetic data, the above method was able to reliably and consistently predict the normalized values close to the theoretical ideal values of reference vessels with a radius > 1 mm. The automatic implementation of Rubin's method

[15] on the same images showed that the method is sensitive to both the reference vessel radius (relative to the ROI size) and the SNR. This is to some extent because Rubin's method is designed to be applied to regions of interest (ROIs) with tissue, a distributed tissue vasculature, and large "reference vessels", so this is not an unexpected result. When the SNR is high, a relatively large error band appears because for Rubin's method to be applied, there must be a recognizable "inflection point" in the cumulative intensity curve. When the cumulative curve is relatively straight, the inflection point becomes difficult to determine (or does not exist at all) for an automatic algorithm, so Rubin's method can artificially return high or low normalization points.

[0143] However, until now, there has been no way to fully automatically identify individual blood vessels in the intricate vascular network seen in the placental bed. As described above, embodiments of the present invention can automatically identify the placenta from 3D power Doppler ultrasound scans (3D-PD USS) and identify relevant anatomical landmarks including the uteroplacental interface (UPI). With the superposition of 3D-PD signals, the vascular network of the placental bed is automatically mapped and measured. From the vascular network in the placental bed, large blood vessels (diameter > 3 mm) located in the myometrium at a distance of about 1 cm (+ / - 0.5 cm) from the UPI can be identified. Blood vessels at different locations can also be used, but it is generally desirable to find suitable blood vessels at or near this location. These blood vessels should be radial arteries because they are the largest blood vessels in this part of the myometrial vascular network at this anatomical location. Radial arteries rapidly dilate during pregnancy and the placental period, and their diameter is larger than that of the uterine arteries supplying them, which is a very rare feature in the vascular network. Since this blood vessel is large enough, it can be assumed that it has 100% vascular distribution. Therefore, in the absence of tissue attenuation or with minimal tissue attenuation, this blood vessel should have a PD voxel value of approximately 250. The actual numerical value (normalized value) of the voxels in this individual blood vessel is used to correct for the unique amount of tissue attenuation experienced by the signal in each individual patient. This enables appropriate quantitative comparison of organ perfusion between women, despite differences in the depth of the target organ (e.g., posterior / anterior placenta), the amount of abdominal fat, or the type of tissue (e.g., presence of scar tissue).

[0144] Therefore, the present invention is capable of performing fully automated single-vessel 3D-FMBV calculations. The techniques described can be extended to automatically identify other potential vascular markers in the uteroplacental vascular plexus (e.g., the length or diameter of the radial artery), which are expected to be useful imaging biomarkers for adverse pregnancy outcomes. Embodiments of the present invention can also be adapted to measure any vascular organ or lesion that can be segmented, such as the kidney, liver (including the fetus), fetal brain, adult ovary, endometrium during IVF, and any soft tissue tumor that can be examined by ultrasound. The fCNN for segmentation and identification of markers can be trained using transfer learning to minimize the amount of manually labeled training data.

[0145] Figure 10 is a flowchart outlining the process according to an embodiment of the present invention. The process includes:

[0146] S1 - Obtaining image data of a subject, for example, by performing an ultrasound scan

[0147] S2 - Optionally cleaning the ultrasound data to remove noise and artifacts

[0148] S3 - Segment the ultrasound data to identify the organ of interest and distinguish it from adjacent tissues

[0149] S4 - Locate predefined anatomical landmarks, such as the interface between the organ of interest and specific neighboring organs

[0150] S5 - Identify the vasculature within and / or near the organ of interest

[0151] S6 - Select a reference vessel within or near the organ of interest

[0152] S7 - Determine a reference power Doppler value or a normalized value from the reference vessel

[0153] S8 - Apply a scaling function (e.g., linear) to the power Doppler values in the ultrasound image based on the power Doppler value of the reference vessel

[0154] S9 - Calculate the moving blood volume fraction value based on the scaled power Doppler values.

[0155] Remove inevitable ultrasound artifacts

[0156] The quality of Doppler US images is affected by various factors, such as parameter settings, cardiac cycle, tissue motion, etc. The cardiac cycle can induce periodic artifacts, resulting in the disconnection of vascular segments. "Flash" artifacts are caused by motion (usually fetal movement), which can generate false blood flow signals, leading to inaccurate perfusion estimation. The detection and removal of these artifacts are ideal preprocessing steps to improve the accuracy of vascular distribution estimation, including 3D - FMBV.

[0157] If there is significant tissue, fluid motion, or rapid movement of the transducer, "flash" artifacts will appear in stationary tissues. This is mainly caused by fetal movement and appears as a sudden burst of random PD signals in a few scan planes, while adjacent planes are not affected (Figure 7(a)). It is easily recognizable by the naked eye. In the literature, images showing "flash" artifacts are usually classified as low - quality images and excluded from the dataset for further analysis. An automatic detection of flash artifacts by comparing the average signal intensity of different planes is proposed (marked by red circles in Figure 7(b)). Then, the identified plane can potentially be replaced by adjacent planes.

[0158] Existing "flash" artifact suppression methods have been designed for 2D US. The 2D methods use information from a time series of frames to identify and remove artifacts while ensuring consistent signal intensity across different time frames. Then, based on the fact that signal intensity should vary smoothly between adjacent scan planes, information from adjacent planes is used to suppress the "flash" artifacts seen in planes where the average intensity changes abruptly (Figure 7(b)). For 3D-PD US scans, it has been proposed to apply interpolation from the nearest clear adjacent planes to generate new planes to replace those with "flash" artifacts. Estimates of perfusion can be calculated before and after processing to evaluate the impact of such artifacts on the quantitative estimation of perfusion (3D-FMBV).

[0159] Periodic noise exists in 3D ultrasound volumes in the form of dark stripes in the PD signal (Figure 8(b)). During the cardiac cycle, blood flow velocity varies periodically. A mismatch between the scan frequency of the scan plane and the heart rate results in variations in signal intensity within the same vascular segment captured by different scan planes, leading to the appearance of dark stripes in 3D vessels.

[0160] Preliminary data indicate that performing linear smoothing between adjacent scan planes has good results in removing periodic noise. In the annular space, a linear kernel is defined (Figure 8(c)) and applied to 3D-PD images in the annular space. The filtered image can be transformed to Cartesian space by our KretzConverter

[17] .

[0161] It has been proposed to use three-dimensional Fourier transform to identify and remove periodic noise caused by the cardiac cycle. The 3D Fast Fourier Transform (FFT) is used to transform the annular power Doppler image to the frequency domain. Periodic noise produces Fourier spikes in the frequency domain, as Figure 9 shown. A notch filter is applied to remove the detected spikes

[18] . Then, the data is transformed back to the spatial domain by the inverse three-dimensional Fourier transform.

[0162] Conclusion

[0163] The method of the present invention can be executed by a computer system including one or more computers. The computer for implementing the present invention may include one or more processors, and the one or more processors include a general-purpose CPU, a graphical processing unit (GPU), a tensor processing unit (TPU), or other dedicated processors. The computer for implementing the present invention can be physical or virtual. The computer for implementing the present invention can be a server, a client, or a workstation. Multiple computers for implementing the present invention can be distributed and interconnected through a network such as a local area network (LAN) or a wide area network (WAN). Each step of the method can be executed by a computer system, but not necessarily the same computer system. The result of the method of the present invention can be displayed to the user or stored in any suitable storage medium. The present invention can be embodied in a non-transitory computer-readable storage medium storing instructions for executing the method of the present invention. The present invention can be embodied in a computer system that includes one or more processors and a memory or storage for storing instructions for executing the method of the present invention. The present invention can be incorporated into an ultrasound scanner or into a software update or add-on component of such a device.

[0164] After describing the present invention, it will be understood that changes can be made to the above embodiments that are not intended to be limiting. The present invention has been described in relation to the scanning of female human subjects. It should be understood that the present invention can also be applied to male humans and other animals. The present invention is defined in the appended claims and their equivalents.

[0165] References

[0166] [1] J.Rubin et al. “Normalizing fractional moving blood volume estimate with power Doppler US: Defining a stable intravascular point within the cumulative power distribution function”. Radiology, 1997, 205: 757 - 765.

[0167] [2] A. Welsh et al. “Inapplicability of fractional moving blood volume technique to standardize Virtual Organ Computer-aided AnaLysis indices for quantified three-dimensional power Doppler”. Ultrasound Obstet Gynecol, 2012, 40: 688 - 692.

[0168] [3] T. Jansson, E. Hernandez-Andrade, G. Lingman and K. “Estimation of fractional moving blood volume in fetal lung using power Doppler ultrasound: Methodological aspects”. Ultrasound Med Biol, 2003, 29: 1551 - 1559.

[0169] [4] N. A. Smith and P. J. Hunter. “An anatomically based model of transient coronary blood flow in the heart”. SIAM Journal on Applied mathematics, 2002, 62(3): 990 - 1018.

[0170] [5] P. Looney, G. N. Stevenson, K. H. Nicholaides, W. Plasencius, S. Natsis, S. L. Collins. “Fully automated, real-time 3D ultrasound segmentation to estimate first trimester placental volume using deep learning”. JCI insights, 2018, 3: e120178.

[0171] [6] G.N. Stevenson, S.L. Collins, J. Ding, L. Impey and J.A. Noble. “3-D Ultrasound Segmentation of the Placenta Using the Random Walker Algorithm: Reliability and Agreement”. Ultrasound Med Biol, 2015, 3182(93): 41.

[0172] [7] O. Ronneberger, F.P and T. Brox. “U-Net: Convolutional Networks for Biomedical Image Segmentation”. 2015 International Conference on Medical Image Computing and Computer-Assisted Intervention, 2015.

[0173] [8] A.W. Welsh, J.B. Fowlkes, S.Z. Pinter et al. “Three-dimensional US Fractional Moving Blood Volume: Validation of Renal Perfusion Quantification”. Radiology, 2019, 293: 460 - 68.

[0174] [9] S.L. Collins, G.N. Stevenson, J.A. Noble and L. Impey. “Utero-placental interface vascularity in early pregnancy estimated by 3D fractional moving blood volume(3D FMBV) predicts fetal growth restriction”. Ultrasound Obstet Gynecol, 2012, 40: 119.

[0175]

[10] T.C. Lee, R.L. Kashyap and C.N. Chu. “Building Skeleton Models via 3-D Medial Surface Axis Thinning Algorithms”. CVGIP: Graphical Models and Image Processing, 1994, 56: 462 - 78.

[0176]

[11] A. Videla, C.-L. Lin and J.D. Miller. “Watershed Functions Applied to a 3D Image Segmentation Problem for the Analysis of Packed Particle Beds”. Particle & Particle Systems Characterization, 2006, 23: 237 - 45.

[0177]

[12] Zuo Wangmeng, et al. “Combination of polar edge detection and active contour model for automated tongue segmentation”. The Third International Conference on Image and Graphics, 2004.

[0178]

[13] C.I. Collewet. “Polar snakes: A fast and robust parametric active contour mode”. The 16th IEEE International Conference on Image Processing, 2009.

[0179]

[14] T.F. Chan and L.A. Vese. “Active contours without edges”. IEEE Transactions on Image Processing, 2001, 10(2): 266 - 277.

[0180]

[15] J.M. Rubin, R.S. Adler and J.B. Fowlkes, et al. “Fractional moving blood volume: estimation with power Doppler”. US. Radiology, 1995, 197: 183 - 90.

[0181]

[16] P.-L. Yen, H.-K. Wu, and H.-S. Tseng et al. “Vascular morphologic information of three-dimensional power Doppler ultrasound is valuable in the classification of breast lesions”. Clinical Imaging, 2012, 36: 267 - 71.

[0182]

[17] P. Looney, G. N. Stevenson, and S. L. Collins. “3D ultrasound file reading and coordinate transformations”. Journal of Open Source Software, 2019, 4: 1063.

[0183]

[18] F. Sur. “An a-contrario approach to quasi-periodic noise removal”. 2015 IEEE International Conference on Image Processing, 2015.

Claims

1. A computer-implemented method for automatically identifying a reference blood vessel in an ultrasound image, the method comprising: Segmenting the ultrasound image to identify an organ; Localizing a predetermined anatomical landmark specific to the organ; Identifying the vasculature near the organ; And Selecting a reference blood vessel from the vasculature having a size within a predetermined size range and located at a predetermined distance range from the predetermined anatomical landmark, wherein the blood vessel boundary of a candidate reference blood vessel is located by determining a blood vessel boundary profile that minimizes an energy function of the blood vessel boundary profile.

2. The method according to claim 1, wherein The diameter of the predetermined size range is greater than 3 mm.

3. The method according to claim 1, wherein The segmentation includes using a trained fully convolutional neural network.

4. The method according to claim 1, wherein The localization includes using a multi-class transfer learning model.

5. The method according to claim 4, wherein The multi-class transfer learning model includes two independent paths: a first path and a second path, the first path having parameters initialized using a segmentation model, and the second path having parameters initialized using variance scaling.

6. The method according to any one of claims 1 to 5, wherein, The identification includes using a method based on automatic multi-seed region growing.

7. The method according to claim 6, wherein The identification further includes using a 3D-medial axis based refinement method on the output of the method based on automatic multi-seed region growing.

8. The method according to any one of claims 1 to 5, wherein The organ is the placenta.

9. The method according to claim 8, wherein, The predetermined anatomical landmark is the uterine-placental interface.

10. The method according to claim 8, wherein, The predetermined distance range is from 0.5 cm to 1.5 cm.

11. A computer-implemented method for plotting blood flow in a power Doppler image of an organ, the method comprising: Using the method according to any one of claims 1 to 10 to identify a reference blood vessel in the power Doppler image; Determining a reference power Doppler value from the reference blood vessel; And Scaling the power Doppler values in the power Doppler image based on the reference power Doppler value to obtain a standardized blood flow image.

12. The method according to claim 11, wherein, The determining includes fitting a model blood vessel distribution to the reference blood vessel and determining standardized points that do not fall within the high shear blood vessel margins.

13. The method according to claim 12, wherein, The fitted model blood vessel distribution includes detecting local maximum intensity points and using a region growing method to obtain an initial blood vessel region where all internal intensities are greater than a threshold.

14. The method according to claim 13, the method further comprising applying an iterative gradient descent technique to minimize a cost function.

15. A computer-implemented method for determining a moving blood volume fraction in an organ, the method comprising: Obtaining a power Doppler image of the organ; Plotting blood flow in the power Doppler image using the method according to any one of claims 11 to 14 to obtain a standardized blood flow image; and Determining the moving blood volume fraction using the standardized blood flow image.

16. A computer-implemented method for calculating the risk of adverse pregnancy outcomes, comprising: Determining the moving blood volume fraction in the placenta of a subject according to the method of claim 15; Measuring the placenta volume of the subject; And Calculating a risk score based at least in part on the moving blood volume fraction and the placenta volume.

17. A computer storage medium comprising instructions which, when executed by a computer system, direct the computer system to perform the method according to any one of claims 1 to 10.

18. A computer storage medium comprising instructions which, when executed by a computer system, direct the computer system to perform the method according to any one of claims 11 to 14.

19. A computer storage medium comprising instructions which, when executed by a computer system, direct the computer system to perform the method according to claim 15.

20. A computer storage medium comprising instructions which, when executed by a computer system, direct the computer system to perform the method according to claim 16.

21. An ultrasound scanner comprising a computer control system configured to perform the method according to any one of claims 1 to 10.

22. An ultrasound scanner comprising a computer control system configured to perform the method according to any one of claims 11 to 14.

23. An ultrasound scanner comprising a computer control system configured to perform the method according to claim 15.

24. An ultrasound scanner comprising a computer control system configured to perform the method according to claim 16.

Images