Fetal assessment method and apparatus

Abstract

An ultrasound image 2010 of part of a foetus is obtained, and the foetus is assessed by automated analysis of the image using a deep learning model 2020 of an artificial intelligence unit. In an aspec

Description

The present invention relates to a method and apparatus for making assessments of a fetus, and is concerned particularly, although not exclusively, with estimating the likely date of delivery of a baby, and / or along with an estimate of its birth weight, based on ultrasound scanning performed during gestation. Obtaining an accurate estimate of the due date of a baby is an important aspect in planning for care of it, and the mother, during and immediately after birth. An estimate of birth weight is also valuable since research has shown, for example, that higher birth weighs are associated with greater risks during labour and poorer neonatal health outcomes. Current clinical practice estimates the due date of the baby via Gestational Age (GA) , and by assuming that all babies are delivered at a GA of 40 weeks (which is certainly not the case). GA is defined as the time elapsed since the start of the Last Menstrual Period (LMP). While it may be assessed by the mother's recollection of their last period date, doing so is problematic in general and so estimating it from ultrasound data at around 12 weeks of gestation is generally considered to be best clinical practice. Accurately estimating due date via this method suffers from these problems: • GA_ estimates become increasingly less accurate the later in gestation they are obtained. This transfers directly to uncertainty in estimated due date. • The method assumes that all babies are born at 40 weeks of gestation, but in reality there is variance around that. Embodiments of the present invention aim to address these problems by directly estimating the time to due date via automated analysis of ultrasound images using deep learning models . Another common facet of pregnancy care is estimation of Estimated Fetal Weight (EFW) which is obtained by taking measurements of features present in ultrasound images and applying formulas to them that yield the estimate. It is important to note that the nature of the approach means that the estimate that is produced is of the weight of the fetus at that point in time, rather than at birth. Despite the desirability of estimating the birth weight, converting to an estimate of that is made difficult by uncertainty in what the gestational age will actually be at birth. Embodiments of the present invention aim to address this difficulty by directly estimating birth weight via automated analysis of ultrasound images using deep learning models. Other details pertaining to the delivery of the baby may also be assessed by analysis of ultrasound images obtained earlier in gestation using deep learning models. These assessments may include any combination of the following: • the likelihood that Caesarean delivery will be required • the likelihood that induction of labour will be required the likelihood that an instrumental delivery (forceps or ventouse) will be required the likelihood of placenta previa the likelihood of placental abruption the likelihood of uterine rupture or scar dehiscence the likelihood of the fetus being in a non-cephalic (breech, transverse oblique or other) presentation after 36 weeks the likelihood of preterm birth the likelihood of rupture of the membranes the likelihood of chorioamnionitis or similar infection the likelihood of Intrauterine Growth Restriction or Small for Gestational Age the likelihood of pre-eclampsia in the mother • the likelihood of fetal distress, such as an abnormal fetal heart rate patter or acidosis • the likelihood of stillbirth • the likelihood of adverse outcome in the newborn such as the need for admission to a Special Care Baby Unit (SBCU), Local Neonatal Unit (LNU) or Neonatal Intensive Care Unit (NICU), low Apgar score, or Hypoxic Encephalopathy Based on the various estimates that are generated (and their confidence levels) the system may optionally present a recommendation for follow-up scanning and / or a physician appointment. The present invention is defined in the attached independent claims, to which reference should now be made. Further, preferred features may be found in the sub-claims appended thereto . According to one aspect of the present invention, there is provided a method for assessing a fetus, the method comprising obtaining at least one ultrasound image of at least a part of the fetus and calculating an estimate of the expected date of delivery of the fetus and / or expected birth weight by an automated analysis of the at least one ultrasound image using a deep learning model of an artificial intelligence unit. According to another aspect of the present invention, there is provided a method for assessing a fetus, the method comprising obtaining at least one ultrasound image of at least a part of the fetus and predicting one or more of: the likelihood that Caesarean delivery will be required the likelihood that induction of labour will be required the likelihood that an instrumental delivery (forceps or ventouse) will be required the likelihood of placenta previa the likelihood of placental abruption the likelihood of uterine rupture or scar dehiscence the likelihood of the fetus being in a non-cephalic (breech, transverse oblique or other) presentation after 36 weeks the likelihood of preterm birth the likelihood of rupture of the membranes the likelihood of chorioamnionitis or similar infection the likelihood of Intrauterine Growth Restriction or Small for Gestational Age the likelihood of pre-eclampsia in the mother the likelihood of fetal distress, such as an abnormal fetal heart rate patter or acidosis the likelihood of stillbirth • the likelihood of adverse outcome in the newborn such as the need for admission to a Special Care Baby Unit (SBCU), Local Neonatal Unit (LNU) or Neonatal Intensive Care Unit (NICU), low Apgar score, or Hypoxic Encephalopathy by an automated analysis of the at least one ultrasound image using a deep learning model of an artificial intelligence unit. The method may comprise calculating the expected date of delivery and / or expected birth weight of the fetus and associated confidence interval(s) from any 2D ultrasound image of the fetus. This is in contrast to standard clinical practice for estimating GA, for example, which requires 2D images from very specific planes to be acquired. The method does not require a 3D volume to be obtained, which requires expensive equipment and sufficient operator skill. The method may comprise calculating any of: the expected date of delivery, expected birth weight of the fetus, the likelihood of various delivery / labour outcomes as illustrated in the examples herein, from an ultrasound image obtained at any point during gestation. In a preferred arrangement, the method does not require an ultrasound machine operator to observe the or each ultrasound image . The method may incorporate data other than the ultrasound images, including, but not limited to, any of the following: • The real-world scale that pixels in the ultrasound images correspond to • Data on previous pregnancy outcomes • Maternal health history • Maternal biometric data, for example including blood pressure • Maternal haematological data Preferably, the method comprises calculating the estimate of the expected date of delivery and / or expected birth weight and corresponding confidence assessment (s) for those estimate(s) using one or more trained machine learning models . Preferably, the machine learning models are produced by training them on a set of representative ultrasound (US) images of fetuses, for each of which the birth date and weight are known. Preferably the dataset for training this model is curated to comprise only cases where the baby was delivered spontaneously (not requiring induction or Caesarean delivery, for example). Model training is preferably accomplished via a process known as supervised learning (which will be well known to a person skilled in the art), whereby the model is configured to achieve the strongest possible (whilst also robust and generalisable) association between each image and the corresponding time-to-birth and birth weight values. In the present invention, the model and supervised training method are preferably additionally configured to output a range of values in which estimates of each of time-to-birth and / or birth weight are expected to lie, with their actual values preferably falling outside of these ranges with low probability. Preferably, the supervised machine learning process takes the form of deep learning, in which the model is an artificial neural network. The supervised learning method may consist of optimising the parameters of this network, via stochastic gradient descent, in order to minimise a loss function. Preferably, the loss function shall be constructed such that for regression tasks (including for example estimates of each of time-to-birth and birth weight) it generates improved performance in terms of both the accuracy of the estimate and the likelihood that the range supplied contains the true value. There now follows, by way of example, a class of loss functions which have this property. The description is framed in terms of time-to-birth prediction, but the skilled person will recognise that it is operable for other regression tasks. Denote : x = an ultrasound image of at least part of a fetus y = the time-to-birth of the fetus = Two outputs of the neural network, representing respectively the predicted time-to-birth and the width of a confidence interval around it A class of functions that have the desired properties are described by L ^“reg + ^^conf Where : Lreg is any non-decreasing function of any semi-norm applied to y — y^ (or alternatively of Iny — y^ if the prediction task is performed in log-space) ^conf is any function having the following properties: For some non-decreasing function f: R -*R+, and for some function g which is a non-decreasing function of a seminorm applied to y— y^ (or Iny — y^) , and for some constant c e R ( _ if giy^y^ <fM conf [c + g^y^-fM otherwise a is an algorithm hyperparameter In a preferred arrangement, the loss function is defined as: L = (iny — 2 + a [max (o, |lny — y^| — f+(ya))] Where : |x| is used to denote the absolute value of x f+ denotes the softplus function f+(x) = ln(l + ex) a is an algorithm hyperparameter For classification tasks (including, for example, prediction of: The likelihood that Caesarean delivery will be required; The likelihood that induction of labour will be required; The likelihood that an instrumental delivery (forceps or ventouse) will be required; The likelihood of placenta previa; The likelihood of placental abruption; The likelihood of uterine rupture or scar dehiscence; The likelihood of the fetus being in a non-cephalic (breech, transverse oblique or other) presentation after 36 weeks; The likelihood of preterm birth; The likelihood of rupture of the membranes; The likelihood of chorioamnionitis or similar infection; The likelihood of Intrauterine Growth Restriction or Small for Gestational Age; The likelihood of pre-eclampsia in the mother; The likelihood of fetal distress, such as an abnormal fetal heart rate patter or acidosis; The likelihood of stillbirth; The likelihood of adverse outcome in the newborn such as the need for admission to a Special Care Baby Unit (SBCU), Local Neonatal Unit (LNU) or Neonatal Intensive Care Unit (NICU), low Apgar score, or Hypoxic Encephalopathy) it is also desirable that the model should both generate a prediction and quantify the uncertainty inherent in it. For such tasks, preferably, the loss function shall be constructed such that it generates improved performance in terms of both the accuracy of the estimate and quantification of the uncertainty inherent in it. We now describe, by way of example, a loss function which has this property. The description is framed in terms of predicting if the fetus will be stillborn, but the skilled person will recognise that it is readily operable for other categorical classification tasks. Denote : x — an ultrasound image of at least part of the content of a uterus containing one or more fetuses y = A one-hot encoded variable indicating whether a fetus was stillborn K (= 2 in this case) the number of categories in the dependent variable a = [alt aK] 6 (IR+)K the output of the neural network, representing concentration parameters of a Dirichlet distribution r(-) is the well-known Gamma function = £Mrw) is the digamma function The loss (for a single example) is defined as: L Lciass + ®-LConf Where : at = yt + (1 - y^ ViE{l,-,K} Lconf = ln n ^ai) - a7)) and a is an algorithm hyperparameter. The loss is, of course, summed over all examples in the dataset. The skilled person will recognise that many other loss formulations may be used to achieve the same objective, including formulations based on mean squared error or Bayes risk, for example, and also ones based on distributions other than Dirichlet. Having trained such a model it is possible to obtain from it point estimates of the class probabilities by computing, for example, the expectation of the resultant Dirichlet distribution: A predicted category may be obtained by identifying the category having the highest class probability: i* = argmax pj The uncertainty inherent in the predicted category may be estimated, for example, by computing the variance associated with it: Var(X^ Ml-Pt) 1 + Sf=i Preferably the neural network architecture will be a convolutional neural network, a vision transformer, or a variant thereof. Preferably the stochastic gradient descent will be performed using the Adam optimiser. The method preferably comprises determining a confidence value of the estimate or prediction. More preferably, the method may comprise determining a confidence value of the expected date of delivery and / or expected birth weight of the fetus. The confidence value may comprise an expression of the confidence that the estimated value is accurate. The method may comprise determining a confidence value of the estimate and / or prediction of the value expressed as a range, wherein the narrower the range the higher the confidence value . The method may comprise obtaining a plurality of ultrasound images of the fetus. In a preferred arrangement, the method comprises determining estimate(s) and corresponding confidence range(s) by applying a machine learning model to a plurality of ultrasound images, then applying numerical processing to the plurality of outputs obtained (including, though not limited to, averaging them). These images may be obtained during a single scanning session or over the course of several. The method may also incorporate other non-imaging data, including, for example any of those listed earlier. In a preferred arrangement, one or more of the steps of calculating the estimate(s) and / or predictions and / or determining their confidence value(s) are performed by an electronic processor. The image (or each of the images) may comprise an image of at least one anatomical feature of a fetus. The method may comprise producing a plurality of estimates and / or predictions, more preferably with a plurality of corresponding confidence values, and filtering the plurality of estimates / predictions to select only one or more estimates / predictions that meet a predetermined confidence value threshold. The method may comprise producing a plurality of estimates / predictions, more preferably with a plurality of corresponding confidence values, and ranking the estimates / predictions based on confidence value. The method may include directing an operative to obtain one or more specific images of the fetus. In one example, the operative may be instructed from the outset to acquire images of certain regions of the fetus that are known (whether through reasoning derived from clinical expertise or via data analysis) to yield accurate estimates / predictions of the quantities and outcomes of interest. In another example the operative may be dynamically instructed, while scanning, to acquire additional images in the event that the images already acquired have wide confidence ranges associated with them. The method may include arithmetically processing, for example averaging, the estimates / predictions obtained from a plurality of images. The images may be selected by an operative during a scan. Alternatively, or in addition, the images may comprise one or more sample frames captured / obtained automatically during a scanning operation. The method may include a method of estimating biometric measurement (s) including, for example, one or more of crown rump length (CRL) , head circumference (HC) , femur length (FL), abdominal circumference (AC) and trans-cerebellar diameter (TCD). The invention also includes a method of assessing a fetus, the method comprising obtaining at least one ultrasound image of at least a part of the fetus, and estimating the birthweight of the fetus by automated analysis of the at least one ultrasound image using a deep learning model of an artificial intelligence unit. According to another aspect of the present invention, there is provided apparatus for assessing a fetus, the apparatus comprising an ultrasound image capturing device for obtaining at least one ultrasound image of at least a part of the fetus and an electronic processing device arranged to process the at least one ultrasound image and to calculate an estimate of the expected date of delivery of the fetus and / or expected birth weight by an automated analysis of the at least one ultrasound image using a deep learning model of an artificial intelligence unit. According to another aspect of the present invention, there is provided apparatus for assessing a fetus, the apparatus comprising an ultrasound image capturing device for obtaining at least one ultrasound image of at least a part of the fetus and an electronic processing device arranged to process the at least one ultrasound image and predicting one or more of: the likelihood that Caesarean delivery will be required the likelihood that induction of labour will be required the likelihood that an instrumental delivery (forceps or ventouse) will be required the likelihood of placenta previa the likelihood of placental abruption the likelihood of uterine rupture or scar dehiscence the likelihood of the fetus being in a non-cephalic (breech, transverse oblique or other) presentation after 36 weeks the likelihood of preterm birth the likelihood of rupture of the membranes the likelihood of chorioamnionitis or similar infection the likelihood of Intrauterine Growth Restriction or Small for Gestational Age the likelihood of pre-eclampsia in the mother the likelihood of fetal distress, such as an abnormal fetal heart rate patter or acidosis the likelihood of stillbirth • the likelihood of adverse outcome in the newborn such as the need for admission to a Special Care Baby Unit (SBCU), Local Neonatal Unit (LNU) or Neonatal Intensive Care Unit (NICU), low Apgar score, or Hypoxic Encephalopathy by an automated analysis of the at least one ultrasound image using a deep learning model of an artificial intelligence unit. Preferably, the processing device is arranged to use a trained machine learning model to calculate the estimate and / or prediction and / or a confidence assessment. The apparatus is preferably arranged to determine a confidence value for the, or each, estimate and / or prediction. The confidence value may comprise an expression of the confidence that the estimate / prediction is accurate. The apparatus is preferably arranged in use to determine a confidence value of the estimate / prediction expressed as a range, wherein the narrower a range the higher the confidence value . The apparatus may be configured to provide real-time feedback to the user on the confidence range of the estimate / prediction that has been obtained, in order that they may direct scanning toward more suitable images when necessary. The processing may be implemented directly on the ultrasound apparatus or else on separate hardware which receives a video feed from the ultrasound apparatus and, optionally, displays results on a separate monitor. The image (or each of the images) may comprise an image of at least one anatomical feature of a fetus. In a further aspect, the invention provides a computer programme product on a computer readable medium, comprising instructions that, when executed by a computer, cause the computer to perform a method of assessing a fetus, the method being in accordance with any statement herein. The invention also comprises a program for causing a device to perform a method of assessing a fetus, the method being in accordance with any statement herein. The invention may include any combination of the features or limitations referred to herein, except such a combination of features as are mutually exclusive, or mutually inconsistent. A preferred embodiment of the present invention will now be described, by way of example only, with reference to the accompanying diagrammatic drawings, in which: Figure 1 shows, schematically, an ultrasound image of a human fetus as captured in accordance with an embodiment of the present invention; Figure 2 shows schematically a comparison of birth date estimation between a previously considered method and a method according to an embodiment of the present invention; Figure 3 shows schematically a comparison between errors in birth date estimation between a previously considered method and a method according to an embodiment of the present invention; Figure 4 shows, schematically a method for training a model for use with embodiments of the invention; Figure 5 shows, schematically, an overview of the method according to an embodiment of the invention; Figure 6 shows, schematically, a first embodiment of method according to the present invention; Figure 7 shows, schematically, a second embodiment of method according to the present invention; Figure 8 shows a first embodiment of apparatus according to the present invention; Figure 9 shows a second embodiment of apparatus according to the present invention; and Figure 10 shows a further embodiment of the present invention. In the developing fetus, different structures will develop at different rates and will change throughout gestation. In embodiments of the present invention, the way in which various structures change over time, optionally the way in which different structures change with respect to one another, can be used to estimate the expected date of delivery and / or expected birth weight of the fetus, and / or the likelihood of various delivery outcomes (including those noted above) using a trained model (or models). The current standard approach to estimating delivery date is via estimating Gestational Age (GA) by obtaining biometric measurements of the fetus, making use of correlations published in the clinical literature between those and GA, and then assuming that the baby will be delivered at a GA of 40 weeks exactly. The weight of the fetus is estimated via a similar process. The process suffers from several issues, as noted earlier. A particular difficulty is that the measurements must be taken from imaging planes that have particular properties, and acquiring such images and taking accurate measurements requires a high level of skill from the operator and extensive training. It would be highly beneficial, both to healthcare providers and patients, if it were possible to significantly de-skill the process of accurately estimating estimated due date and birth weight, such that it could be performed by a much broader population, after minimal training. This would enable assessment to be performed in a primary care setting, for instance. It would be further preferable if such a process could be performed without the operator to look at the ultrasound images at all. This is especially preferable where it is undesirable (or illegal - as in India) to reveal the sex of the fetus. The required level of operator skill may be reduced somewhat by the use of 3D ultrasound probes which, in principle at least, allows the burden of plane finding and measurement to be shifted somewhat from the operator to an algorithm. This is not a complete shift, since acquiring a suitable 3D volume still requires a higher level of skill than the average person has. Also, in practice, the plane finding and measurement problems are very difficult. All approaches that we are aware of are either less accurate than manual biometric measurement or impractical for use during routine scanning - where outputs are required in near real time - or both. Additionally, 3D ultrasound probes (and the types of ultrasound machines that they can work with) are significantly more expensive than 2D probes. It is therefore desirable to have a method of estimating GA which takes solely 2D ultrasound images as its inputs. Figure 1 shows schematically a part of an ultrasound image I of a fetus F obtained in accordance with an embodiment of the present invention. In some implementations, embodiments of the present invention use Deep Learning techniques to automatically generate reliable estimates of the expected date of delivery and / or expected birth weight of the fetus age from ultrasound images directly, rather than via proxy biometric measurements. The method and associated apparatus are preferably capable of generating accurate estimates from a broad range of images, without having to carefully acquire specific images of the fetus or images having particular properties or even for the operator to have to look at the image. Although the input ultrasound images do not contain information about the real-world scale of the anatomy present within them, they do contain information on the relative scales of different areas of it, which can inform on fetal growth, without requiring information about the real-world size of pixels in the image. However, where such information is available it may be integrated into the model to improve its predictive accuracy. A trained machine learning model is designed to produce an estimate of the expected date of delivery and / or expected birth weight from any image of the fetus, rather than requiring images that have certain properties relating to the position of the fetus and presence of certain anatomical features. Since some images presented to the model may be unsuitable for this purpose, the model also reports its confidence in the prediction that it makes, which may be in the form of a prediction interval - or range - optionally containing the actual age with a confidence expressed as a percentage. The prediction intervals are dynamic and based on the suitability of the input image for the task. On good, or highly suitable, images the intervals are narrow, and on bad ones they are wide. The model predictions can then be filtered to include only those which date the gestational age with sufficiently high precision. In some embodiments of the invention, the sonographer / operative may be directed to obtain certain images of the fetus that are known to be highly predictive of the quantity or outcome of interest (e.g., mid-sagittal view of the whole fetus, axial cross-section of the head), and these images combined as inputs to the model to improve its predictive accuracy. The model outputs may be averaged over a number of images (e.g., a collection selected by the sonographer during scanning, and / or a sample of frames obtained in real-time during scanning) where doing so is found to improve the accuracy of the predictions. This may include over images obtained during multiple scanning sessions, and with consideration of the trajectories of the outputs over time. A further purpose of the invention is to additionally output other predictions covering potential outcomes for the delivery of the baby, including, though not limited to: The likelihood that Caesarean delivery will be required; The likelihood that induction of labour will be required; The likelihood that an instrumental delivery (forceps or ventouse) will be required; The likelihood of placenta previa; The likelihood of placental abruption; The likelihood of uterine rupture or scar dehiscence; The likelihood of the fetus being in a non-cephalic (breech, transverse oblique or other) presentation after 36 weeks; The likelihood of preterm birth; The likelihood of rupture of the membranes; The likelihood of chorioamnionitis or similar infection; The likelihood of Intrauterine Growth Restriction or Small for Gestational Age; The likelihood of pre-eclampsia in the mother; The likelihood of fetal distress, such as an abnormal fetal heart rate patter or ac i do sis; The likelihood of stillbirth; The likelihood of adverse outcome in the newborn such as the need for admission to a Special Care Baby Unit (SBCU), Local Neonatal Unit (LNU) or Neonatal Intensive Care Unit (NICU), low Apgar score, or Hypoxic Encephalopathy. Outputs which depend upon an analysis of multiple frames of video may require a small adjustment to model configuration, as follows. Representation vectors, preferably obtained from a layer immediately prior to the prediction heads of the model, should be cached for each frame. The prediction heads should be adjusted to take as input a concatenation of some subset of these cached representation vectors. These adjustments are preferably made during both model training and in its application. The invention is operable for producing estimates and confidence values throughout the entire period of gestation. Figure 2 shows schematically a comparison of birth date estimation between a previously considered method and a method according to an embodiment of the present invention; Figure 3 shows schematically a comparison between errors in birth date estimation between a previously considered method and a method according to an embodiment of the present invention; Turning to Figure 4, this shows generally at 1000 a method for training a machine-learning model suitable for use with embodiments of the present invention. At Step 1010, a data batch is supplied to a processor which contains known values of time-to-birth which applies a loss function operation at Step 1020. At Step 1030, loss gradients are calculated with respect to selected model parameters. At Step 1050 the data is fed into the machine learning model, along with a batch of corresponding ultrasound images obtained at Step 1040. A machine learning model prediction (i.e. of the time-to-birth) is then obtained at Step 1060, before the loop is repeated for n iterations, with parameters updated at each cycle. Finally at Step 1070 the trained machine learning model is complete. Figure 5 shows generally at 2000 an overview of a model inference process. At Step 2010 an ultrasound image is obtained. At Step 2020 the image is supplied to the trained machine learning model. At Steps 2030 and 2040, which may be substantially simultaneous, the predicted time-to-birth value and corresponding confidence interval value are outputted. Figure 6 shows generally at 3000 a first embodiment of method according to the present invention. At Step 3010 an ultrasound image is acquired. The image is fed into the model at Step 3020 and a predicted time-to-birth value, together with an associated confidence interval are calculated at Step 3030. At Step 3040 it is determined whether the confidence interval is narrower than a value for the best confidence interval obtained. If not, the ultrasound operation is optionally directed to acquire an ultrasound image of a particular plane in the fetus at Step 3050. On the other hand, if the value is determined to be narrower than the best obtained at Step 3040, the predicted value and confidence interval are displayed at Step 3060 and a register for the value of the best confidence interval is updated at Step 3070. Figure 7 shows generally at 4000 a second embodiment of the method according to the present invention. At Step 4010 an ultrasound operator is provided with image acquisition instructions. Next, at Step 4020, a number of ultrasound images are captured by the operative. Each one is provided to the trained model at Step 4030 and predicted time-to-birth values together with associated confidence intervals are obtained at Step 4040. The values are aggregated at Step 4050 and the aggregated values are output at Step 4060. Figure 8 is a schematic representation of a first embodiment of apparatus according to the present invention. The apparatus is shown generally at 5000. An ultrasound scanning machine 5010 has an image receiving unit 5020 which receives acquired ultrasound images. The images are processed by the model in a processor 5030 and a time-to-birth prediction and confidence interval are supplied to a GUI pipeline 5040. A display 5050 of the machine 5010 then displays the prediction and confidence interval. Figure 9 shows schematically, generally at 6000 a second embodiment of apparatus according to the present invention. In this embodiment, an ultrasound scanning machine 6010 acquires ultrasound images which are then supplied to an external processing unit 6020, for example via an HDMI or other link. The external unit 6020 receives the images, processes them via the model and outputs a time-to-birth estimation and confidence interval which are then rendered on a separate display 6030 of the external unit. Figure 10 shows a further embodiment of the present invention, generally at 7000, in which an ultrasound scanning machine 7010 acquires ultrasound images which are then supplied via the camera interface C of a user device 7020, which in this case is a hand-held device such as a cell phone. The user device 7020 is optionally able to display the or each ultrasound image on its display 7030. The processor of the device 7020 uses software, such as may be provided in an app, to perform the tasks of receiving the or each ultrasound image 7040, processing the or each image 7050 and predicting the time-to-birth, along with a confidence calculation 7060. Whereas the embodiments described above use the examples of birth date estimation and birth weight estimation, it will be understood by the skilled person that there may, as an alternative or in addition, be assessments of different kinds, such as (but not limited to) : The likelihood that Caesarean delivery will be required; The likelihood that induction of labour will be required; The likelihood that an instrumental delivery (forceps or ventouse) will be required; The likelihood of placenta previa; The likelihood of placental abruption; The likelihood of uterine rupture or scar dehiscence; The likelihood of the fetus being in a non-cephalic (breech, transverse oblique or other) presentation after 36 weeks; The likelihood of preterm birth; The likelihood of rupture of the membranes; The likelihood of chorioamnionitis or similar infection; The likelihood of Intrauterine Growth Restriction or Small for Gestational Age; The likelihood of pre-eclampsia in the mother; The likelihood of fetal distress, such as an abnormal fetal heart rate patter or acidosis; The likelihood of stillbirth; The likelihood of adverse outcome in the newborn such as the need for admission to a Special Care Baby Unit (SBCU), Local Neonatal Unit (LNU) or Neonatal Intensive Care Unit (NICU), low Apgar score, or Hypoxic Encephalopathy. Whilst endeavouring in the foregoing specification to draw attention to those features of the invention believed to be of particular importance, it should be understood that the applicant claims protection in respect of any patentable feature or combination of features referred to herein, and / or shown in the drawings, whether or not particular emphasis has been placed thereon.

Claims

l.A method for assessing a fetus, the method comprisingobtaining at least one ultrasound image of at least apart of thefetus and calculating an estimate of theexpected date of delivery of the fetus and / or expectedbirth weight by an automated analysis of the at leastone ultrasound image using a deep learning model of anartificial intelligence unit.

2. A method for assessing a fetus, the method comprising obtaining at least one ultrasound image of at least a part of the fetus and predicting one or more of:• the likelihood that Caesarean delivery will be required• the likelihood that induction of labour will be required• the likelihood that an instrumental delivery (forceps or ventouse) will be required• the likelihood of placenta previa• the likelihood of placental abruption• the likelihood of uterine rupture or scar dehiscence• the likelihood of the fetus being in a non-cephalic (breech, transverse oblique or other) presentation after 36 weeks• the likelihood of preterm birth• the likelihood of rupture of the membranes• the likelihood of chorioamnionitis or similar infection• the likelihood of Intrauterine Growth Restriction or Small for Gestational Age• the likelihood of pre-eclampsia in the mother• the likelihood of fetal distress, such as an abnormal fetal heart rate patter or acidosis• the likelihood of stillbirth• the likelihood of adverse outcome in the newborn such as the need for admission to a Special Care Baby Unit (SBCU), Local Neonatal Unit (LNU) or Neonatal Intensive Care Unit (NICU), low Apgar score, or Hypoxic Encephalopathyby an automated analysis of the at least one ultrasound image using a deep learning model of an artificial intelligence unit.

3. A method according to Claim 1 or 2, wherein the estimate or prediction and / or a confidence assessment are produced by a machine learning model.

4. A method according to Claim 3, wherein the machine learning model is produced by training on a set of representative ultrasound images of fetuses, for each of which the at least one outcome is known.

5. A method according to Claim 3 or 4, wherein model training is accomplished via supervised learning, whereby the model is configured to achieve the strongest possible (whilst also robust and generalisable) association between each image and its corresponding outcome .

6. A method according to Claim 5, wherein the model and / or supervised training method are additionally configured to output a value range in which the estimate or prediction is expected to lie.

7. A method according to Claim 5 or 6, wherein the supervised machine learning process takes the form of deep learning, in which the model is an artificial neural network.

8. A method according to Claim 7, wherein the supervised learning method comprises optimising the parameters of the network, via stochastic gradient descent, in order to minimise a loss function.

9. A method according to Claim 8, wherein the loss function is constructed such that it generates improved performance for the accuracy of the prediction and the corresponding confidence assessment.

10. A method according to Claim 7, wherein the neural network architecture is a convolutional neural network, a vision transformer, or a variant thereof.

11. A method according to any of the preceding claims, wherein the method comprises determining at least one relative dimension of one or more anatomical features of the fetus.

12. A method according to any of the preceding claims, wherein the method comprises producing a plurality of estimates / predictions with a plurality of corresponding confidence values, and filtering the plurality ofestimates / predictions to select only one or moreestimates / predictions that meet a predeterminedconfidence value threshold.

13. A method according to any of the preceding claims, wherein the method comprises producing a plurality of estimates / predictions, with a plurality of corresponding confidence values, and ranking the estimates / predictions based on confidence value.

14. A method according to any of the preceding claims, wherein the method includes directing an operative to obtain one or more specific images of the fetus.

15. A method according to Claim 14, wherein the operative is instructed to acquire images of certain regions of the fetus that are known (whether through reasoning derived from clinical expertise or via data analysis) to yield accurate estimates / predictions.

16. A method according to Claim 14 or 15, wherein the operative is dynamically instructed, while scanning, to acquire additional images in the event that the images already acquired have wide confidence ranges associated with them.

17. A method according to any of the preceding claims, wherein the method includes arithmetically processing, for example averaging, the estimates / predictions obtained from a plurality of images.

18. A method according to any of the preceding claims, wherein the method includes calculating an estimate / prediction and a confidence value from any ultrasound image of the fetus.

19. A method according to any of the preceding claims,wherein the method includes calculating estimate / prediction and / or a confidence value from a fetal image obtained at any stage during gestation.

20. Apparatus for assessing a fetus, the apparatus comprising an ultrasound image capturing device for obtaining at least one ultrasound image of at least a part of the fetus and an electronic processing device arranged to process the at least one ultrasound image and to calculate an estimate of the expected date of delivery of the fetus and / or of expected birth weight by an automated analysis of the at least one ultrasound image using a deep learning model of an artificial intelligence unit.

21. Apparatus for assessing a fetus, the apparatus comprising an ultrasound image capturing device for obtaining at least one ultrasound image of at least a part of the fetus and an electronic processing device arranged to process the at least one ultrasound image and to predict one or more of:• the likelihood that Caesarean delivery will be required• the likelihood that induction of labour will be required• the likelihood that an instrumental delivery (forceps or ventouse) will be required• the likelihood of placenta previa• the likelihood of placental abruption• the likelihood of uterine rupture or scar dehiscence• the likelihood of the fetus being in a non-cephalic (breech, transverse oblique or other) presentation after 36 weeks• the likelihood of preterm birth• the likelihood of rupture of the membranes• the likelihood of chorioamnionitis or similar infection• the likelihood of Intrauterine Growth Restriction or Small for Gestational Age• the likelihood of pre-eclampsia in the mother• the likelihood of fetal distress, such as an abnormal fetal heart rate patter or acidosis• the likelihood of stillbirth• the likelihood of adverse outcome in the newborn such as the need for admission to a Special Care Baby Unit (SBCU), Local Neonatal Unit (LNU) or Neonatal Intensive Care Unit (NICU), low Apgar score, or Hypoxic Encephalopathyby an automated analysis of the at least one ultrasound image using a deep learning model of an artificial intelligence unit.

22. Apparatus according to Claim 20 or 21, wherein the processing device is arranged to use a trained machine learning model to calculate the estimate / prediction and / or a confidence assessment.

23. Apparatus according to any of Claims 20-22, wherein the apparatus is configured to provide real-time feedback to the user on the confidence range of the estimate / prediction that has been obtained, in order that they may direct scanning toward more suitable images when necessary.5 24. Apparatus according to any of Claims 20 to 23,wherein the processing is implemented directly on the ultrasound apparatus or else on a separate device whicheither receives a video feed from the ultrasound apparatus or captures a copy of one or more of the 10 ultrasound images via a camera of the device, andoptionally displays a result on a separate monitor.

Images