Ultrasound method

Abstract

A computer-implemented method for automatically identifying reference vessels in an ultrasound image, the method comprising: segmenting the image to identify an organ; locating predetermined anatomical landmarks specific to the organ; identifying appropriate vascular structures associated with the organ; and selecting reference vessels from the vascular structures that have dimensions within a predetermined dimensional range and are located within a predetermined distance from a predetermined interface.

Description

[Technical field]

[0001] The present invention relates to ultrasound imaging and measurement, particularly using power Doppler ultrasound techniques to measure blood flow.

[0002] background Power Doppler ultrasound is a technique that uses the amplitude of the reflected signal as a function of the frequency shift relative to the emitted signal to identify tissue motion, particularly blood flow, within a subject. Power Doppler (PD) ultrasound is quantified by integrating the amplitude of ultrasound scattering within an area of ​​interest (2D-PD) or volume of interest (3D-PD). This is achieved by summing the power Doppler signal within the region of interest, where the PD signal represents the local concentration of red blood cells (which scatter the ultrasound). In general, power Doppler ultrasound can only provide information about relative blood flow, since the 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 is strongly attenuating. Knowledge of the absolute blood flow would improve the diagnosis of various conditions. Known techniques for standardizing power Doppler signals include manually identifying a region of interest that serves as a standardization reference from which a normalization point is determined. The normalization point used is a "knee value"; power Doppler intensity values ​​above the normalization point are assigned a value of 1, and power Doppler intensity values ​​below the normalization point are scaled to the range 0–1 [1].

[0003] overview Techniques are desired to allow absolute measurements of blood flow.

[0004] In accordance with the present invention, there is provided a computer-implemented method for automatically identifying reference vessels in an ultrasound image, comprising the steps of: Segmenting the image to identify organs; Locating predetermined anatomical landmarks specific to the organ; Identifying an appropriate vascular structure for an organ; selecting a reference vessel from the vascular structure having a size within a predetermined size range and located within a predetermined distance from a predetermined anatomical landmark; A method is provided, comprising:

[0005] Preferably, the predetermined size range is greater than 3mm in diameter.

[0006] Preferably, the segmenting includes using a trained fully convolutional neural network.

[0007] Preferably, the localizing includes using a multi-class transfer learning model.

[0008] Desirably, the multi-class transfer learning model includes two independent passes, a first pass having parameters initialized using a segmentation model and a second pass having parameters initialized using variance scaling.

[0009] Preferably, the identifying includes using an automated multi-seed region growing based process.

[0010] Preferably, the identifying step further comprises using a 3D medial axis based thinning process on the output of the expansion based process.

[0011] Preferably, the organ is the placenta, and the anatomical landmark is preferably the uterus-placenta interface. In this case, preferably, the predetermined distance range is about 0.5 cm to about 1.5 cm.

[0012] In accordance with the present invention, there is provided a computer-implemented method for mapping blood flow in a power Doppler image of an organ, comprising the steps of: identifying a reference vessel in a power Doppler image using the method described above; determining a reference power Doppler value from a reference vessel; Scaling the power Doppler values ​​in the power Doppler image based on a reference power Doppler value to obtain a standardized blood flow image. A method is provided, comprising:

[0013] Desirably, determining includes fitting a model vessel profile to the reference vessel and determining standardization points that do not fall within high shear vessel peripheries.

[0014] Preferably, fitting the model vessel profile includes detecting local maximum intensity points and using a region growing method to obtain an initial vessel region where all interior intensities are above a threshold.

[0015] Preferably, an iterative gradient descent technique is applied to minimize the cost function.

[0016] According to the present invention there is provided a computer implemented method for determining fractional transfer blood volume in an organ, comprising the steps of: obtaining a power Doppler image of the organ; Mapping blood flow in a power Doppler image using the method described above to obtain a standardized blood flow image; Using standardized blood flow images to identify fractional transfer blood volume A method is provided, comprising:

[0017] According to the present invention there is provided a computer-implemented method for calculating risk of an adverse pregnancy outcome, such as fetal growth restriction and / or preeclampsia, comprising the steps of: determining a fractional transfer blood volume within the placenta of a subject in accordance with the method described above; Measuring the placental volume of the subject; calculating a risk score based at least in part on the fractional transfer blood volume and the placenta volume; A method is provided, comprising:

[0018] Thus, the present invention allows for fully automated comparison of quantitative blood flow estimates across multiple different patients, each with different tissue attenuation, by normalizing the absolute signal so that the integral of the signal across the region of interest is consistently baselined. To do this, a large vessel with "100% vascularity" is identified at a similar level of tissue depth as the region of interest, and the recorded value for the PD signal in that vessel is used as a "standardization point" [2]. When the PD signal across the entire region of interest is divided by the signal in that large vessel, other smaller vessels within this region of interest will consistently have the same proportional signal intensity in relation to the vessel with 100% vascularity. This standardization process produces a measure known as the Fractional Moving Blood Volume (FMBV), which is the only validated method for quantitatively measuring perfusion with ultrasound in both 2D and 3D.

[0019] The present invention provides a novel method for FMBV estimation that overcomes the limitations of the prior art and can increase the accuracy of the technique. For example, the present invention avoids the difficulties arising from the definition of a region of interest for analysis, where the normalization value is a function of the size of the region of interest and the nature of the tissue contained within the region of interest. Furthermore, the technique of the present invention is insensitive to the signal-to-noise ratio in the acquired images and is often highly insensitive to noise. Furthermore, the present invention is insensitive to the effect that medium-sized vessels within the region of interest have on the location of the "knee point" and the normalization value subsequently calculated.

[0020] The present invention allows for the automatic identification of power Doppler normalized values ​​from a single vessel to adjust for tissue attenuation of the ultrasound signal.Embodiments of the present invention provide a robust and efficient method for automatically identifying power Doppler normalized values ​​from a single vessel extracted from a complex vascular network imaged by standard 3D power Doppler ultrasound.

[0021] In the following the invention will be further explained with reference to exemplary embodiments and the accompanying drawings. [Brief description of the drawings]

[0022] [Figure 1] FIG. 1 illustrates the architecture of a fully convolutional neural network (fCNN) that can be used in embodiments of the present invention. [Diagram 2] FIG. 1 shows example images of placenta segmentation with 2D B-mode planes (left), semi-automated random walker segmentation (center), and segmentation using the fCNN from FIG. 1 (right). [Diagram 3] 3a and 3b are graphs comparing placental volumes calculated by the fCNN of FIG. 1 with the semi-automated Random Walker technique and the commercially available VOCAL® technique, respectively. [Figure 4] FIG. 1 illustrates how the utero-placental interface (UPI) can be identified once the amniotic-placental interface (API) and fetal-placental interface (FPI) are known. [Figure 5a] FIG. 1 illustrates a fully convolutional neural network architecture of a model used in one embodiment of the present invention, specifically, the Placenta Segmentation (PS) model. [Figure 5b] FIG. 1 illustrates a fully convolutional neural network architecture of a model used in one embodiment of the present invention, specifically, the Hybrid (HB) model. [Figure 5c] FIG. [Figure 6a] FIG. 1 shows a skeletonization of a vascular tree and an example of vascular centerlines showing cross-sectional profiles. [Figure 6b] FIG. 1 shows a vascular tree with separated vessel segments. [Figure 7a] FIG. 1 illustrates flash artifacts. [Figure 7b]FIG. 1 shows the average Doppler signal intensity across the scan plane. [Figure 8a] FIG. 1 shows a B-mode ultrasound scan superimposed with a power Doppler signal showing periodic noise. [Figure 8b] FIG. [Figure 8c] FIG. 2 illustrates a linear kernel. [Figure 9] FIG. 3D Fast Fourier Transform of 3D-US data showing periodic noise. [Figure 10] 2 is a flow chart of a method according to the present invention;

[0023] Like parts in the different drawings are designated by like reference numerals.

[0024] Exemplary embodiments Theory: Power Doppler signal profile of blood vessels Jannson et al. showed in a phantom model containing scattering particles that Doppler power is proportional to velocity over a range of powers and concluded that this is likely to be true for a variety of machines and settings. [3] In any elliptical cross-section of a vessel, the velocity or intensity I profile has the following shape: TIFF0007602015000001.tif20170, where a and b are the radii of the major and minor axes of the vessel cross-section, x and y are Cartesian coordinates defining the major and minor axes of the vessel, and I max is the true (noise-free) maximum intensity inside the vessel, and γ is a parameter that serves to “flatten” the velocity profile: γ=2 represents the classical Poiseuille flow profile in conduits with elliptical cross-sections, and γ greater than 2 allows for flattening of the profile due to flow pulsatility (especially in the aorta) [4].

[0025] The greatest risk factor for stillbirth is fetal growth restriction (FGR), which usually occurs following suboptimal placental implantation. Unfortunately, currently available methods for assigning risk of FGR perform very poorly, and many women who are determined to be at "low risk" are unaware that their fetus is not growing properly until they present with a stillbirth. If classified as "high risk," women undergo a series of ultrasound growth scans to identify the fetus as being growth restricted, with the goal of delivering the fetus before stillbirth occurs. Because current risk assessment methods perform so poorly, many women who are estimated to be at "high risk" give birth to well-growing, healthy infants after pregnancies filled with unnecessary stress and anxiety.

[0026] Robust and reliable first-trimester screening methods to assess the risk of FGR are essential, meaning that women at real risk of stillbirth following FGR will be referred to serial ultrasound developmental scans where available.

[0027] First Trimester Screening Exam First-trimester placental volume (PlVol) has long been known to correlate with term birth weight, and as early as 1981, it was suggested that PlVol measured using B-mode ultrasound could be used to screen for FGR. Since then, numerous studies have demonstrated that low PlVol between 11 and 13 weeks of gestation can predict adverse pregnancy outcomes, including small for gestational age (SGA) - an alternative marker to FGR, and preeclampsia. As PlVol has also been demonstrated to be independent of other biomarkers such as pregnancy associated plasma protein A (PAPP-A) and fetal nuchal translucency (NT), a recent systematic review concluded that PlVol could be successfully incorporated into future multivariate screening methods for FGR, similar to the "combined test" currently used to screen for fetal aneuploidies (including Down's syndrome) in the first trimester. Initial pilot data also demonstrated that combining PlVol with estimates of placental perfusion was able to distinguish between different adverse pregnancy outcomes, with low PlVol and normal vascularity being associated with an increased risk of FGR, and low PlVol and undervascularity increasing the risk of developing preeclampsia.

[0028] While there is solid background data supporting the value of PlVol and placental vascularity to make them useful imaging biomarkers, there is a need for reliable, operator-independent, real-time techniques to estimate PlVol and placental vascularity.

[0029] Until now, the only approach to estimate PlVol has been to annotate the placenta in three-dimensional ultrasound (3D-US) either manually (tracing the placenta's perimeter in all 2D slices that make up the 3D volume) or using semi-automated tools such as VOCAL (GE®, USA). VOCAL® requires the operator to "trace" the placenta's perimeter in some, but not all, 2D slices, and then use a rotational interpolation algorithm to estimate the final volume. Both of these strategies are too time-consuming and operator-dependent to be useful as part of a population-based screening study, and they fail to provide any other morphological indices such as shape or surface area. There is also significant debate about the repeatability and reproducibility of the volumetric results produced by VOCAL® due to the variability of the number of outlines used (degree of interpolation) and the organ of interest.

[0030] Products that allow automated measurement of validated vascular measurements, such as fractional moving blood volume (FMBV), are not currently available.

[0031] We have developed and validated a novel fully convolutional neural network (OxNNet) [5] (see https: / / insight.jci.org / articles / view / 120178) that can automatically identify placental volume from 3D-US. As discussed further below, the present invention provides a further development that allows for the automatic identification of relevant anatomical landmarks to facilitate fully automated vascular measurements.

[0032] To develop OxNNet, placentas were segmented from 2393 first-trimester 3D-US volumes using the gold-standard semi-automated random walker technique [6]. This was quality-controlled by three human operators to create a “ground truth” dataset. A fully convolutional neural network (fCNN), termed OxNNet, was created using the framework TensorFlow (v1.3) and a 3D architecture inspired by the 2D U-net [7] (see Figure 1). The impact of training set size on model performance was investigated by using samples of 100, 150, 300, 600, 900, and 1200 cases trained across 25,000 iterations.

[0033] To evaluate PlVol segmentation, 2-fold cross-validation was performed using 1097 placentas as a training set with 100 validation images. OxNNet was trained for 8 epochs, which took 26 hours to run. It was then tested on 1196 cases. The final computation of PlVol following training took an average of 11 seconds.

[0034] Each predicted segmentation was post-processed to remove disconnected parts less than 40% of the maximum region volume. Segments were binary dilated and eroded using a 3D kernel of radius 3 voxels and application of a 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 estimation of PlVol).

[0035] OxNNet produced fully automatic segmentation (median Dice Similarity Coefficient (DSC) = 0.84; see Figure 2) over conventional techniques when compared to the "ground truth" dataset. The impact of the training set size on OxNNet's performance demonstrated the need for large datasets (n>1000). This is important since most imaging tools generated by machine learning to date have used very small datasets. The clinical utility of PlVol estimates generated by OxNNet was tested by observing the prediction of small for gestational age (SGA) fetuses (an alternative marker to FGR). Receiver-operating characteristic curves demonstrated nearly identical results for prediction of SGA fetuses for PlVol generated by OxNNet (AUC 0.65 (95% CI; 0.61-0.69)) and the "ground truth" generated by the semi-automated random walker technique (AUC 0.65 (95% CI; 0.61-0.69): see Figure 3a). OxNNet outperformed the only commercially available tool, VOCAL® (AUC 0.57 (95% CI; 0.48-0.65): see Figure 3b).

[0036] The present invention proposes to identify the utero-placental interface (UPI) as an anatomical reference point to examine the vascularity of the placental bed (see FIG. 4). Although OxNNet can successfully segment the entire placenta, the present invention proposes to distinguish between two different placental surfaces: the UPI, where the placenta is implanted in the uterus, and the amniotic-placental interface (API), where the placenta is in contact with the amniotic fluid (which fills the sac that contains the fetus) (see FIG. 4). In some cases, the fetus may be lying in contact with the placenta, so the fetal-placental interface (FPI) must also be identified and included in the API.

[0037] One embodiment of the present invention uses machine learning to develop a robust multi-class segmentation technique to facilitate automatic identification of anatomical landmarks by discriminating between the placenta, fetus, and amniotic fluid from B-mode 3D ultrasound scans.

[0038] From the available 2393 3D-US volumes with existing placenta segmentation, 300 volumes were randomly selected for multi-class segmentation. Amniotic fluid and fetus were manually seeded by two operators and combined with the placenta seeding performed from a previous study. Since amniotic fluid and fetus have clearer boundaries than the placenta, initialization was much easier, but all cases where ambiguity existed were reviewed by a clinical expert with extensive experience in obstetric ultrasound. These three distinct classes were then segmented as a multi-class label map using the gold standard Random Walker algorithm.

[0039] The 300 multi-class (MC) cases were subdivided into 200 training cases, 40 validation cases, and 60 test cases. Four MC models were trained for 40 epochs each with a batch size of 30. First, the MC models were trained using an fCNN with an identical architecture to the original OxNNet (see Figure 5a). Next, a multi-class transfer learning (MCTL) model was trained using the same architecture but with weights and biases obtained from the original OxNNet training (effectively "transferring" what OxNNet "learned" from a much larger training dataset of 1097 cases compared to the 200-case simple MC model).

[0040] To overcome any possible performance degradation for the other two targets (fetus and amniotic fluid), two hybrid models were developed: the hybrid model (HB) and the hybrid model with exponential averaging (HBAV). The exponential moving average reduces noise by averaging the model weights throughout the training process, favoring more recent weight values, and also provides computational efficiency since not all weights need to be stored. Both hybrid models consist of two independent paths created as shown in Figure 5b. The architectures of the multi-class (MC) and multi-class transfer learning (MCTL) models are not shown here because they only differ from the placental segmentation (PS) model in having four channels instead of two.

[0041] Both the hybrid model (HB) and the hybrid model with exponential averaging (HBAV) consist of two independent passes as shown in Figure 5b. In the upper pass, the HB parameters were initialized using the PS model, and in the case of the HBAV model, the PSAV model. In the lower layer, the parameters were initialized using variance scaling and then trained on the MC data. For both the HB and HBAV models, the parameters were allowed to change in the lower pass and were fixed in the upper pass. The parameters of the Adam optimizer were identical to those used in the PS training.

[0042] P Background , P Placenta , P Aminotic , and P Fetus Let u, u, and u be the confidence that a voxel belongs to the background, placenta, amniotic fluid, and fetus, respectively. For a given voxel i, the softmax output of the upper pathway (T) has exactly two values ​​that sum to 1: TIFF0007602015000002.tif7170 and The image is given as TIFF0007602015000003.tif7170. In this case, the fetus and amniotic fluid have been included in the background. In the lower path (B), the final layer is softmaxed to sum to 1. A confidence for a given voxel's membership was generated with a scalar value of TIFF0007602015000004.tif13170. TIFF0007602015000005.tif13170 showed the confidence that a voxel is neither placenta, fetus, nor amniotic fluid. The loss function L is TIFF0007602015000006.tif13170 is defined by combining the outputs of the two paths, where M is a binary mask and m is the value of this binary mask. i is 0 if voxel i is within the placenta ground truth region and 1 if it is outside this region, and o i is the output of the lower pass, sl is the softmax cross entropy function, and n i was a normalization factor defined as 1 minus the output from the softmax layer in the upper path.

[0043] The loss function is masked over the placenta domain so that there is no contribution of the placenta in training the model. The final confidence vector in the HB model is The placenta segmentation of the HB model is therefore given by: For all voxels that are TIFF0007602015000008.tif13170, they will be identical to the PS model. The voxel in TIFF0007602015000009.tif7170 is classified as placenta by the HB model, while the remaining classes TIFF0007602015000010.tif13170, respectively If it has a value smaller than TIFF0007602015000011.tif13170, it will be classified as background by the PS model.

[0044] The optimizer parameters were identical to those used in the original OxNNet training. All four models provided useful results, but the hybrid model with exponential averaging was the best. Using these models, it is possible to automatically identify all possible placental morphological indices associated with anatomical landmarks such as the site and surface area of ​​UPI. This allows mapping of the vascularity of the placental bed, which improves predictability for adverse pregnancy outcomes such as preeclampsia.

[0045] Once the utero-placental interface (UPI) has been identified, the next step is to assess blood flow within the placenta.

[0046] Doppler ultrasound is known as the primary non-invasive imaging modality performed during pregnancy to assess blood flow inside the placenta and fetal organs. Power Doppler (PD) ultrasound is more useful for assessing tissue perfusion because it is less affected by the ultrasound incidence angle compared to color Doppler and is sensitive to multiple flow directions and low flow velocities. However, like all forms of ultrasound, PD ultrasound signals are attenuated by the tissues through which they pass. Standardization of the power Doppler signal is necessary to compensate for the attenuation and allow direct interpatient comparisons. The only validated method to estimate tissue perfusion from PD ultrasound is fractional moving blood volume (FMBV). This method has recently been validated in 3D against the gold standard of radioactive materials in a porcine model microspheres [8]. Using 3D-FMBV, we demonstrated in a small pilot study (n=143) that first trimester placental bed perfusion was significantly reduced in preeclamptic pregnancies (p=0.03) but not in normotensive pregnancies resulting in small for gestational age (SGA) infants (p=0.16).[9] This is consistent with conventional histopathological findings in preeclampsia of spiral artery maladaptation.

[0047] For the purpose of standardizing the power Doppler signal to determine the 3D-FMBV, it is proposed to identify a large reference vessel in the vicinity of the target of interest, which can be assumed to have 100% vascularity. The PD signal inside the reference vessel is then used to define a standardization value, which corrects the PD signal value recorded in the target, thereby adjusting for the attenuation of the individual signals to allow proper inter-patient comparison. The present invention aims to fully automate the estimation of 3D-FMBV, thereby facilitating large-scale testing of the present discovery and potentially improving the predictive ability of first trimester screening tests. The present invention can also be applied to measurements of vascularity in other organs and / or tumors, in particular using transfer learning to minimize the training required for other organs or tumors.

[0048] A fully automated technique for mapping and measuring utero-placental vasculature from 3D-PD ultrasound scans with automatic identification of anatomical landmarks as described above is described below. The technique moves away from defining regions of interest that may vary between studies and focuses on identifying consistent large vessels that can be automatically identified for normalization. Furthermore, because the procedure identifies and analyzes a single blood-filled vessel for normalization, the procedure provides an algorithm that is highly consistent over the range of signal-to-noise ratios typical for power Doppler and is not subject to normalization variability due to the presence of perfused parenchymal tissue or medium-sized vessels within the surrounding region of interest.

[0049] First, the entire uterine vasculature was extracted using an automated multi-seed region-growing-based method. This method was chosen because it is superior to commonly used thresholding methods, as it can preserve the representation of vascular connectivity. Next, a 3D medial axis-based thinning method

[10] was employed to extract the vascular plexus skeleton of the myometrium (see Fig. 6a). The vascular skeleton was then further divided into a number of segments by removing branching points. Using these centerline segments as markers, a 3D watershed algorithm

[11] was applied to divide the vascular network into multiple vascular segments (Fig. 6b). To find the vascular diameter, a local vascular cross-section, i.e., the intersection between the volume of the vascular segment and a plane perpendicular to the local vascular centerline, was drawn for each segment (Fig. 6a). Vascular segments showing a single peak intensity in this cross-section are identified as individual vessels to be considered as candidates for reference vessels. The reference vessels selected are vessels with a maximum diameter greater than 3 mm located inside the volume of interest. Alternatively, any individual vessels having a diameter greater than a pre-specified threshold, for example 3 mm, may be selected.

[0050] Combining this vessel identification technique with a grayscale OxNNet tool that can identify anatomical landmarks in the placenta, including the UPI, allows the selection of an appropriate reference vessel in the placental bed at a predefined distance from the UPI. The predefined distance may be about 1 cm, for example between 0.5 cm and 1.5 cm. If there are multiple vessels that meet the criteria, the largest one can be selected. This reference vessel can then be used to calculate a fully automated estimate of the 3D-FMBV (perfusion) throughout the UPI, or any volume of interest relative to the UPI, using the new method.

[0051] Estimation of 3D-FMBV using vessels known to have 100% vascularity is a way to normalize the numerical value of the power Doppler signal, which is affected by tissue attenuation. Such normalization allows quantitative inter-patient comparisons, since it corrects for the depth of the target organ and the type of tissue irradiated with ultrasound (e.g., correcting for fat mass).

[0052] The calculation of the normalized values ​​is independent of vessel dimensions greater than 3 mm and adapts to the vessel intensity profile shape. The calculation has been demonstrated to be accurate by comparison with synthetic data. Currently, there are no techniques to generate normalized points, either manually or automatically, from a single vessel.

[0053] In an embodiment, the normalization points are determined by automatically identifying high shear vascular perimeters for individual vessels based on vascular characteristics (intensity profile and diameter) and fluid dynamics simulations that indicate "constant" values ​​of blood flow velocity. A shear threshold can be applied, which allows for an accurate estimation of the numerical normalization value every time, as will be described in more detail below.

[0054] Locating vascular borders It is observed that vascular boundaries in PD ultrasound images have low intensity (close to the background intensity) and therefore low contrast. Thresholding is commonly used to segment vessels as an easy-to-operate method. However, due to the low contrast of the boundaries and the inevitable noise induced during imaging, simple thresholding does not provide accurate vessel segmentation. The polar active contour method has been shown to perform well in boundary detection of other forms of objects with roughly convex shapes

[12]

[13]

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

[0055] Active contours deform the object boundary to minimize a defined energy function to obtain a segmentation of the object

[14] . Let S be a given image containing a vessel of interest, defined in the domain Ω, and let C = {χ│φ(χ) = 0} be a closed contour, where φ is a signed distance function. The interior of C (the candidate vessel region) is an approximation of the smoothed Heaviside function: Defined by TIFF0007602015000012.tif26170.

[0056] The area outside of C, i.e., outside the target vessel, is 1-H φ(χ) Let E(C) be the energy function. The predicted vessel boundary is the contour C that minimizes the energy function.

[0057] In a polar coordinate system, the contour samples the boundary at multiple angles. In that case, the polar contour evolves only in the radial direction. To describe the energy in a polar coordinate system, we use the characteristic function TIFF0007602015000013.tif13170 is defined, where the vessel center is considered as a pole, χ and p are two points, and θ(·) is the angular coordinate of one point in the polar coordinate system. The function L(χ,p) is 1 if and only if p is on the same radial line as χ.

[0058] Next, the energy function is where δφ(χ) is the derivative of Hφ(χ). F predicts, for each contour point, the adherence (radial) of the local intensity to the best-fit intensity model at that point. According to equation (1), the radial intensity profile of the vessel is given by The file is TIFF0007602015000015.tif20170.

[0059] Note that when detecting small sized uterine vessels, which are observed to have a parabolic intensity profile, γ is set to 2. The estimated inner intensity χ is calculated by fitting equation (10) to the inner radial intensity of the vessel contour point χ. TIFF0007602015000016.tif20170 was obtained.

[0060] The outer average intensity is Calculated as TIFF0007602015000017.tif20170.

[0061] Next, the energy It is represented as TIFF0007602015000018.tif20170.

[0062] Finally, a regularization term penalizes local curve variance to smooth the contour. TIFF0007602015000019.tif13170 is added and weighted by the factor ξ. The final energy is Defined as TIFF0007602015000020.tif13170.

[0063] The method can be implemented in three steps: initialization, gradient descent, and detection of stopping conditions.

[0064] In the initialization, a local maximum intensity point is detected and considered as the vessel center. Using this point as a seed, a simple region growing method is applied to ensure that all interior intensities are within a threshold T ini The initial vascular area exceeding the initial contour C ini consists of a set of points on the boundary of the initial region.

[0065] Iterative gradient descent is applied to the energy E (a cost function) defined until an equilibrium stopping condition is met. total (φ) is iteratively minimized. The vessel center point is updated to be the center of gravity of the vessel region at each iteration.

[0066] An exemplary equilibrium stopping condition is that the variance in the vessel region is very small and the threshold T var It is smaller than that.

[0067] Determining the normalization point As discussed above, in a phantom model containing scattering particles, it was shown that Doppler power is proportional to velocity over the power range of 0-35 dB, and it was concluded that this is likely to be true for a variety of machines and settings, although it is machine setup dependent [3]. For any elliptical cross section of a vessel, the velocity or intensity I profile has the following shape: TIFF0007602015000021.tif20170, where a and b are the radii of the major and minor axes of the vessel cross-section, x and y are Cartesian coordinates defining the major and minor axes of the vessel, and I maxis the true (noise-free) maximum intensity inside the vessel, and γ is a parameter that serves to “flatten” the velocity profile: γ=2 represents the classical Poiseuille flow profile in conduits with elliptical cross-sections, and γ greater than 2 allows for flattening of the profile due to flow pulsatility (especially in the aorta) [5].

[0068] The classical cumulative intensity curve in power Doppler ultrasound imaging calculates the number of pixels / voxels in a region of interest that have an intensity less than or equal to a given intensity. The mathematical equivalent of this number in the elliptical cross section of the vessel is the area of ​​the vessel that has an intensity less than the given value. Transforming the coordinates into polar coordinates, defining x=ar cosnθ and y=br sinθ, the definition of intensity is: TIFF0007602015000022.tif13170. Then, (r>r * (I) is equal to the area of ​​an elliptical torus having * The area of ​​the vessel having It can be calculated as TIFF0007602015000023.tif13170.

[0069] Since the present invention is interested in pixel intensity and not vessel area, TIFF0007602015000024.tif13170 can be used instead. * The area of ​​the vessel having The result is TIFF0007602015000025.tif13170.

[0070] This value is normalized by the total cross-sectional area of ​​the vessel, πab, to obtain I * Cumulative probability of the intensity of Get TIFF0007602015000026.tif13170.

[0071] In the discrete sense, p(I * ) is calculated as ​​​TIFF0007602015000027.tif13170, 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 I * is the total number of pixels having a power intensity less than

[0072] Theoretically, if the region of interest is defined as a vessel cross-section, one can fit discrete data to this curve and obtain γ and I max It is possible to estimate both γ and I in the case of a parabolic velocity (or intensity) profile, i.e., γ = 2. Note that this is a straight line, independent of the aspect ratio of the vessel's elliptical cross section. Note that for many vessels within the utero-placental system, dimensions and hemodynamic considerations suggest that γ ≈ 2 (which is also confirmed by plotting the intensity profile through the center of the vessel cross section in a power Doppler image). That is, it is necessary to fit γ and I to the data as needed to accurately estimate the peak signal in the vessel. max Therefore, the present invention proposes an alternative "weighted" cumulative curve, which applies both γ and I to the vessel even though γ ≈ 2. max This allows both to be accurately and reliably applied.

[0073] The proposed cumulative curve is for intensity I * reflects the total signal in the vessel cross section, and in the discrete sense, TIFF0007602015000028.tif13170, where I(i) is the Doppler power intensity at pixel i, and I total is the cumulative total intensity over the entire vessel cross section.

[0074] (I total In integral form (without normalization by ), the proposed cumulative curve is ​The result is TIFF0007602015000029.tif58170.

[0075] lastly, When normalized by TIFF0007602015000030.tif13170, The file name is TIFF0007602015000031.tif20170.

[0076] This provides a suitable cumulative curve, which is then fitted to the maximum power intensity I max and the "flatness" γ.

[0077] Similar equations in polar coordinates can be derived for vessels of any cross-section.

[0078] The normalization point is determined by using the concept of a "shear threshold" to distinguish between the high shear periphery of the vessel (where red blood cells are scarce) and the vessel itself where red blood cells are more concentrated. Shear is specified as the gradient of the flow velocity at the vessel wall and its peak value. The intensity where shear is X% of the maximum value is found by calculating the gradient of equation (15): It is defined by using "radius" instead of intensity, as in TIFF0007602015000032.tif20170.

[0079] For a practical implementation, the following steps are performed:

[0080] 1. The intensity weighted Doppler power cumulative distribution is calculated using equation (21).

[0081] 2. γ and I max Equation (23) is fitted to the cumulative distribution to provide the best-fit value of

[0082] 3. These best fit values ​​are expressed as I normto calculate X = 85. The appropriate value of X may vary depending on the size and velocity profile in the normalized vessel (renal arteries are expected to have a flatter velocity profile than the utero-placental arteries). However, for smaller vessels such as the utero-placental arteries, where the flow profile is typically parabolic, X = 85 is an appropriate threshold value that keeps the vessel cross-section away from the high shear vessel wall.

[0083] Tests based on synthetic data have shown that for reference vessels with radii greater than 1 mm, it is possible to consistently and reliably predict normalized values ​​close to the theoretical ideal. An automated implementation of Rubin's method

[15] on the same images shows that the method is sensitive to both the reference vessel radius (relative to the ROI dimensions) and the SNR. This is not surprising, in part, because Rubin's method is designed to be applied to ROIs with tissue, distributed tissue vasculature, and large "reference vessels". Higher SNRs result in relatively large error bars, because the application of Rubin's method requires the presence of a discernible "knee" in the cumulative intensity curve. When the cumulative curve is relatively linear, the knee becomes difficult to identify (or does not exist) by the automated algorithm, and for this reason, Rubin's method may return artificially high or low normalized points.

[0084] However, there has been no approach to fully automate the identification of a single vessel within a tortuous vascular network such as that found in the placental bed. One embodiment can automatically identify the placenta from a 3D power Doppler Ultrasound scan (3D-PD USS) as described above, and can automatically identify relevant anatomical landmarks, including the utero-placental interface (UPI). With the 3D-PD signal superimposed, the vascular network of the placental bed is automatically mapped and measured. A large (diameter > 3 mm) vessel located in the myometrium approximately 1 cm (± 0.5 cm) from the UPI is identified from the vascular network within the placenta. Although it is possible to use a vessel in a different location, it is usually expected to find a suitable vessel at or near the above location. This suitable vessel should be the radial artery, since it is the largest vessel in this part of the myometrial vascular network at this anatomical location. The radial artery expands rapidly during pregnancy and at the placental stage becomes larger in diameter than the uterine artery that supplies it, a very unusual feature in the vascular network. This vessel is large enough that it can be assumed to have 100% vascularity and should therefore have a PD voxel value of approximately 250 if there is no or minimal tissue attenuation. The actual values ​​recorded for the voxels in this single vessel (normalized values) are used to correct for the unique amount of tissue attenuation experienced by the signal in each individual patient. This allows for proper quantitative comparison of organ perfusion across multiple women, despite different target organ depths, e.g. posterior / anterior placenta, different amounts of abdominal fat, or different tissue types, e.g. scar tissue or not.

[0085] This allows the present invention to perform a fully automated single vessel 3D-FMBV calculation. This described technique can be expanded to automatically identify other potential vascular markers (e.g., radial artery length or diameter) in the utero-placental vascular plexus that 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. Examples include the kidney, liver (including fetal stages), fetal brain, adult ovaries, endometrium during IVF, and any soft tissue tumor that can be examined by ultrasound. The fCNN used for landmark segmentation and identification can be trained using transfer learning to minimize the amount of manually labeled training data.

[0086] Figure 10 is a flow diagram summarizing a process according to one embodiment of the present invention, which includes: S1 - acquiring image data of a subject, for example by performing an ultrasound scan; S2 - Optionally, cleaning up the ultrasound data to remove noise and artifacts; S3 - Segmenting the ultrasound data to identify an organ of interest and distinguish the organ of interest from adjacent tissue; S4 - Locating a predetermined anatomical landmark, for example an interface between the organ of interest and a particular adjacent organ; S5 - Identifying vascular structures within and / or adjacent to an organ of interest; S6 - Selecting a reference vessel in or near the organ of interest; determining a reference power Doppler value or a standardized value from the S7-reference vessel; Applying a scaling function (e.g., linear) to the power Doppler values ​​in the ultrasound image based on the power Doppler values ​​of the S8-reference vessel; S9 - Calculating the fractional movement blood volume value from the scaled power Doppler value; Includes.

[0087] Eliminating unavoidable ultrasound artifacts The quality of Doppler US images is affected by many factors, e.g., parameter settings, cardiac cycle, and tissue motion. The cardiac cycle can induce periodic artifacts, resulting in truncation of vessel segments. "Flash" artifacts are caused by motion (often fetal movement) and can generate false signals of blood flow that lead to inaccurate estimation of perfusion. Detection and removal of those artifacts is a desirable preprocessing step to improve accuracy for estimating vascularity, including 3D-FMBV.

[0088] In the presence of large tissue or liquid movements or fast transducer movements, "flash" artifacts appear in stationary tissue. This is mainly due to fetal movements and appears as sudden bursts of random PD signal in some scan planes, while neighboring planes are unaffected (Fig. 7a). This is easily identifiable 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. It is proposed to automatically detect flash artifacts by comparing the average signal intensity of several different planes (marked with red circles in Fig. 7b). The identified plane can then be replaced, if appropriate, by the neighboring planes.

[0089] In the case of 2D US, an existing "flash" artifact suppression method is designed. The 2D method uses information from temporal serial frames to identify and remove artifacts, assuming that the signal intensity should be consistent between different time frames. Then, information from neighboring planes is used to suppress "flash" artifacts found in planes with abrupt changes in mean intensity (Figure 7b), based on the fact that the signal intensity should change smoothly between adjacent scan planes. In the case of 3D-PD US scanning, it is proposed to apply an interpolation from the nearest sharp neighboring planes to generate new planes and to replace these new planes with the "flash" artifacts. Perfusion estimates can be calculated before and after processing to evaluate the impact of such artifacts on the quantitative estimate of perfusion (3D-FMBV).

[0090] There is periodic noise in the 3D ultrasound volume as dark stripes in the PD signal (Fig. 8b). During the cardiac cycle, the blood flow velocity changes periodically. The mismatch between the sweep frequency of the scanning plane and the heart rate can cause dispersion of the signal intensity within the same vessel segment captured by multiple different scanning planes, resulting in the appearance of dark stripes in the 3D vessel.

[0091] Preliminary data demonstrated that performing linear smoothing between adjacent scan planes yielded promising results in removing periodic noise. A linear kernel was defined in toroidal space (Fig. 8c) and applied to the 3D-PD image in toroidal space. The filtered image can then be transformed to Cartesian space by our KretzConverter

[17] .

[0092] It is proposed to use 3D Fourier transform to identify and remove periodic noise caused by the cardiac cycle. The toroidal power Doppler image is transformed into the frequency domain by 3D FFT (Fast Fourier Transform). It is observed that periodic noise generates Fourier spikes in the frequency domain, as shown in Figure 9. A notch filter is applied to remove the detected spikes

[18] . The data is then transformed back into the spatial domain by a 3D inverse Fourier transform.

[0093] conclusion The method of the present invention can be implemented by a computer system including one or more computers. The computer used to implement the present invention can include one or more processors, including a general-purpose CPU, a graphics processing unit (GPU), a tensor processing unit (TPU), or other specialized processors. The computer used to implement the present invention can be physical or virtual. The computer used to implement the present invention can be a server, a client, or a workstation. The computers used to implement the present invention can be distributed and interconnected through a network, such as a local area network (LAN) or a wide area network (WAN). The individual steps of the method can be performed by a computer system, but not necessarily by the same computer system. The results of the method of the present invention can be displayed to a user or stored in any suitable storage medium. The present invention can be embodied as a non-transitory computer-readable storage medium that stores instructions for carrying out the method of the present invention. The present invention can be embodied in the form of a computer system including one or more processors and a memory or storage that stores instructions for carrying out the method of the present invention. The present invention can be incorporated into an ultrasound scanner or into a software update or add-on for such an apparatus.

[0094] Having described the invention, it will be understood that modifications may be made to the above embodiments which are not intended to be limiting. The invention has been described in relation to scanning female human subjects. It will be understood that the invention is also applicable to males and other animals. The invention is defined in the appended claims and their equivalents.

[0095] References [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,” vol. 205, pp. 757-765, 1997. [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, vol. 40, pp. 688-692., 2012. [3] T. Jansson, E. Hernandez-Andrade, G. Lingman and K. Marsal, “Estimation of fractional moving blood volume in fetal lung using power Doppler ultrasound: Methodological aspects.,” Ultrasound Med Biol, vol. 29, pp. 1551-1559, 2003. [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, vol. 62(3), pp. 990-1018, 2002. [5] P. Looney, G. N. Stevenson, K. H. Nicholaides, W. Plasencius, S. Natsis and S. L. Collins, “Fully automated, real-time 3D ultrasound segmentation to estimate first trimester placental volume using deep learning,” JCI insights 2018;3:e120178.. [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, Vols. 3182-93, p. 41, 2015. [7] O. Ronneberger, F. P and T. Brox, “U-Net: Convolutional Networks for Biomedical Image Segmentation,” in International Conference on Medical Image Computing and Computer-Assisted Intervention, 2015. [8] A. W. Welsh, J. B. Fowlkes, S. Z. Pinter and et al, “Three-dimensional US Fractional Moving Blood Volume: Validation of Renal Perfusion Quantification,” Radiology, vol. 293, pp. 460-68, 2019. [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, vol. 40, p. 119, 2012.

[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,” vol. 56, pp. 462-78, 1994.

[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, vol. 23, pp. 237-45, 2006.

[12] Z. Wangmeng et al, “Combination of polar edge detection and active contour model for automated tongue segmentation,” Third International Conference on Image and Graphics (ICIG’04), 2004.

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

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

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

[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, vol. 36, pp. 267-71, 2012.

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

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

Claims

1. 1. A computer-implemented method for automatically identifying reference vessels in an ultrasound image, comprising: Segmenting the image to identify organs; Locating predetermined anatomical landmarks characteristic of said organ; identifying an appropriate vascular structure for said organ; selecting a reference vessel from said vascular structure having a size within a predetermined size range and located within a predetermined distance from said predetermined anatomical landmark; A method comprising:

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

3. The method of claim 1 or 2, wherein the segmenting comprises using a trained fully convolutional neural network.

4. The method of claim 1 , wherein the localizing comprises using a multi-class transfer learning model.

5. The multi-class transfer learning model includes two independent paths: The first pass has parameters initialized using a segmentation model; The second pass has parameters initialized using variance scaling. The method of claim 4.

6. The method of claim 1 , wherein the identifying step comprises using an automated multi-seed region-growing based process.

7. The method of claim 6 , wherein the identifying step further comprises using a 3D medial axis-based thinning process on an output of the expansion-based process.

8. the organ is the placenta, Preferably, the anatomical landmark is the uterus-placenta interface.

8. The method according to any one of claims 1 to 7.

9. The method of claim 8, wherein the predetermined distance range is from about 0.5 cm to about 1.5 cm.

10. 1. A computer-implemented method for mapping blood flow in a power Doppler image of an organ, comprising: Identifying a reference vessel in the power Doppler image using a method according to any one of claims 1 to 9; determining a reference power Doppler value from the reference vessel; scaling power Doppler values ​​in the power Doppler image based on the reference power Doppler values ​​to obtain a standardized blood flow image; A method comprising:

11. The determining step comprises: fitting a model vascular profile to the reference vessel; Determining a standardized point that does not fall within the high shear vascular periphery; The method of claim 10, comprising:

12. Fitting the model vascular profile comprises: Detecting local maximum intensity points; using a region growing method to obtain an initial vascular region in which all internal intensities are above a threshold; The method of claim 11 , comprising:

13. The method of claim 12 , further comprising applying an iterative gradient descent technique to minimize the cost function.

14. 1. A computer-implemented method for identifying fractional moving blood volumes in an organ, comprising: obtaining a power Doppler image of the organ; Mapping blood flow in said power Doppler image to obtain a standardized blood flow image using a method according to any one of claims 10 to 13, Using the standardized blood flow image to identify a fractional displacement blood volume; A method comprising:

15. 1. A computer-implemented method for calculating risk of an adverse pregnancy outcome, comprising: Determining a fractional transfer blood volume within the placenta of a subject according to the method of claim 14; Measuring the placental volume of the subject; calculating a risk score based at least in part on said fractional transfer blood volume and said placental volume; A method comprising:

16. A computer program comprising instructions which, when executed by a computer system, instruct the computer system to perform a method according to any one of claims 1 to 15.

17. 16. An ultrasound scanner comprising a computer control system configured to carry out the method of any one of claims 1 to 15.

Images