Whole-target image capture apparatus and method

Abstract

A method for analysing a plurality of ultrasound images or volumes captured from a 3D target space comprises processing the images or volumes using an electronic processor and determining their relative positions and orientations within the 3D target space, using an artificial intelligence unit. To determine how much of a target has been scanned, segmented images are placed in 3-D space and a transformation is determined of an idealised 3-D model of the target which optimally overlaps the segmented images (for instance by maximising the sum of intersection over unions over all the images). As each scan is conducted, the method tracks which parts of the target have been seen, and which are unseen. Progress may be presented visually to the user in real-time, for example by displaying the percentage of the target that has been viewed, or as a 3d model where the “seen” portions are coloured differently. The target space may be an anatomical feature such as an organ of a body.

Description

The present invention relates to apparatus and a method for an image capture operation, and is concerned in particular, although not exclusively, with an apparatus and method in which a risk of failure to capture the whole of a target in an ultrasound scanning operation is minimised. As ultrasound is a cheap, portable, and safe imaging modality, it is an ideal tool for scanning a patient's organs to search for pathologies or acquire diagnostic images. For example, ultrasound can be used to identify and locate gall stones, kidney stones and liver lesions. Ultrasound can also be used to measure organs or other anatomical features, e.g., measuring the volume of liquid in the bladder. A key source of misdiagnosis when using ultrasound scanning to screen for pathologies is failure to fully examine the anatomical regions where they are present. For example, when assessing the presence of liver cancer, a false negative diagnosis might occur if the particular region of the liver in which the tumour is present has not been scanned. In a further example, the presence of spina bifida in a fetus may be misdiagnosed when the spine is not assessed in its entirety and from suitable orientations. It is therefore very important to ensure that the whole of the organ (or anatomical region) is seen to ensure that no pathologies have been missed. Scanning the whole of an organ is often done by establishing the extents of the organ and then "sweeping" from one side to the other by pivoting or sliding the probe. Scanning small organs such as the gallbladder is relatively simple. However, scanning larger organs such as the liver or a kidney is more difficult, as it may not be possible to view the entire organ from a single sweep point because artefacts such as bowel gas or ribs can obscure the view. In these situations, the sonographer must piece together a 3-D mental image of the organ and keep track of which parts of the organ have been seen and which have not. The sonographer must frequently use several different probe positions to ensure that whole of the organ has been viewed. In addition, ultrasound is often used to locate or monitor lesions, perhaps those found on other imaging modalities (x-ray, MRI etc.). When viewing suspected lesions under ultrasound, it is common practice to image the potential lesion from at least two separate viewpoints, since if the potential lesion is visible in at least two different planes, it is unlikely to be an artefact. However, determining suitable probe positions to provide views of a particular point in an organ can be difficult, especially if bowel gas or ribs are present, for example. Embodiments of the present invention aim to provide apparatus and a method of performing an ultrasound scan in which the aforementioned problems are addressed. 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 an apparatus for analysing a plurality of ultrasound images or volumes captured from a 3D target space, the apparatus comprising an image processor and an artificial intelligence unit, wherein the processor is arranged to process the images or volumes and the artificial intelligence unit is arranged to determine their relative positions and / or orientations within the 3D target space. A trained machined learning model may be used. Using the artificial intelligence unit, the apparatus is preferably arranged to determine the relative positions and orientations of at least two 2D ultrasound images in 3D space by assessing values of feature vectors obtained from a trained machine learning model while tracing along lines in each image, calculating similarity measures between the resultant feature vector trajectory signatures and assessing whether the similarity measures are below a threshold value. Alternatively or in addition, using the artificial intelligence unit, the apparatus may be arranged to determine the relative positions and orientations of at least two 3D ultrasound volumes in 3D space by employing feature detectors to identify common features within each of them and then solving for similarity transformations between the feature sets that optimally align the detected features. Preferably, the apparatus is arranged to segment anatomical structures of interest within the images or volumes being analysed. Segmentation of anatomical structures may be performed by applying a trained machine learning model. Preferably, the locations and orientations of segmented ultrasound images or volumes in 3D space are calculated, and an affine transformation of a 3D model of an anatomical structure is determined such that it optimally aligns it with those segmented images or volumes. The apparatus is preferably arranged to further determine from the segmented image portions and their intersection with the 3D model, which parts of the target anatomical structure have been scanned and / or from which orientations and / or which parts of the target anatomical structure have yet to be scanned from any given orientation. The anatomical structure (s) may comprise and / or may be comprised within, an organ of a body. Preferably each ultrasound image or volume is analysed using a trained machine learning model to assess whether pathology is present within it, and a numeric indicator calculated that summarises the proportion of the volume of an anatomical structure covered by images or volumes that were assessed to exhibit no pathology. The apparatus may be arranged in use to provide to the operator a / any subset of the following information: (i) which parts of the target anatomical structure (s) have been scanned, and from which directions (ii) The proportion of the volume of the anatomical structure (s) that has been scanned (iii) The proportion of the volume of the anatomical structure (s) that has been scanned from at least two different directions (iv) The proportion of the volume of the anatomical structure (s) that is covered by images or volumes that have been assessed to exhibit no pathologies. In a preferred arrangement, the apparatus is arranged in use to provide one or more prompts to position the ultrasound probe in particular location and / or orientation to best image any previously unseen portions of the target space. According to another aspect of the present invention, there is provided an apparatus for capturing an ultrasound image of a target, the apparatus comprising an ultrasound probe, an image processor and an artificial intelligence unit, wherein the ultrasound probe is arranged to scan a body containing a target to capture ultrasound images of the target, which images are processed by the image processor using the artificial intelligence unit to identify which portions of the target have been imaged and / or which portions of the body are yet to be imaged. The apparatus is preferably arranged in use to provide an indication to an operative as to which portions of the target have been imaged and / or which portions of the target are yet to be imaged. The apparatus may include a monitor for providing the indication (s) to the operative. In a preferred arrangement, the apparatus captures images of a target which comprises an organ of the body. The apparatus may be arranged in use to provide a prompt to the operative to position the ultrasound probe in a particular way to ensure that the whole of the target has been imaged. The artificial intelligence unit is preferably arranged in use to apply a deep learning model to the images captured by the probe to determine which features of the target have been imaged. Using the artificial intelligence unit, the apparatus is preferably arranged to assess values of feature vectors relating to features of the target during a scanning movement of the probe. The artificial intelligence unit preferably uses the feature vectors to determine a plurality of trajectory signatures which it then compares to determine whether a pair of trajectory signatures have a level of similarity greater than a threshold value. An image portion may then be segmented. In a preferred arrangement, the segmented image portions are placed in a representation of the target. The apparatus is preferably arranged to determine from the segmented image portions which parts of the target have been scanned and / or which parts of the target have yet to be scanned. According to another aspect of the present invention, there is provided a method for analysing a plurality of ultrasound images or volumes captured from a 3D target space, the method comprising processing the images or volumes using an electronic processor and determining their relative positions and / or orientations within the 3D target space, using an artificial intelligence unit. The method may use a trained machine learning model. Preferably the method comprises using the artificial intelligence unit to determine the relative positions and orientations of at least two 2D ultrasound images in 3D space by assessing values of feature vectors obtained from a trained machine learning model while tracing along lines in each image, calculating similarity measures between the resultant feature vector trajectory signatures and assessing whether the similarity measures are below a threshold value. The method may include using the artificial intelligence unit to determine the relative positions and orientations of at least two 3D ultrasound volumes in 3D space by employing feature detectors to identify common features within each of them and then solving for similarity transformations between the feature sets that optimally align the detected features. In a preferred arrangement, the method includes segmenting anatomical structures of interest within the images or volumes being analysed. Preferably the segmentation of anatomical structures is performed by applying a trained machine learning model. Preferably the locations and orientations of segmented ultrasound images or volumes in 3D space are calculated, and an affine transformation of a 3D model of an anatomical structure is determined such that it optimally aligns it with those segmented images or volumes. The method may further comprise determining from the segmented image portions and their intersection with the 3D model, which parts of the target anatomical structure have been scanned and / or from which orientations and / or which parts of the target anatomical structure have yet to be scanned from any given orientation. The anatomical structure(s) may comprise an organ of a body. In a preferred arrangement, each ultrasound image or volume is analysed using a trained machine learning model to assess whether pathology is present within it, and a numeric indicator calculated that summarises the proportion of the volume of an anatomical structure covered by images or volumes that were assessed to exhibit no pathology. The method preferably further includes providing to the operator a / any subset of the following information: (i) which parts of the target anatomical structure(s) have been scanned, and from which directions (ii) The proportion of the volume of the anatomical structure (s) that has been scanned (iii) The proportion of the volume of the anatomical structure (s) that has been scanned from at least two different directions (iv) The proportion of the volume of the anatomical structure (s) that is covered by images or volumes that have been assessed to exhibit no pathologies. In a preferred arrangement, prompts are provided to position the ultrasound probe in particular location and / or orientation to best image any previously unseen portions of the target space. The invention also includes a method of capturing an ultrasound image of a target, the method comprising using an ultrasound probe to scan a body containing a target to capture ultrasound images of the target and processing the images using an artificial intelligence unit to identify which portions of the target have been imaged and / or which portions of the target have yet to be imaged. The method preferably includes providing an indication to an operative as to which portions of the target have been imaged and / or which portions of the target are yet to be imaged. The indication (s) may be provided on a monitor. In a preferred arrangement, the method includes capturing images of a target which comprises an organ of the body. The method may include providing a prompt to the operative to position the ultrasound probe in a particular way to ensure that the whole of the target has been imaged. The method may include applying a deep learning model to the images captured by the probe to determine which features of the target have been imaged. Preferably, the method comprises using the artificial intelligence unit to assess values of feature vectors relating to features of the target during a scanning movement of the probe. The method may include using the artificial intelligence unit to assess the feature vectors to determine a plurality of trajectory signatures which are then compared to determine whether a pair of trajectory signatures have a level of similarity greater than a threshold value. An image portion may be segmented. In a preferred arrangement, the method comprises placing segmented image portions in a representation of the target. The method preferably includes determining from the segmented image portions which parts of the target have been scanned and / or which parts of the target have yet to be scanned. The method may comprise producing a numeric indicator which quantifies the likelihood that no pathology is present in the site of interest, more preferably in conjunction with pathology detection methods (implemented, for example, as additional deep learning models). Preferably, the method includes quantifying the fraction of the volume of the region of interest that is covered by ultrasound images / volumes that are assessed to contain no pathology. The value of this indicator is preferably reported to the operator in real-time as scanning progresses. 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 according to any statement herein. The invention also comprises a program for causing a device to perform a method according to 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, is an illustrative example of a pair of images in which corresponding features are identified; Figure 2 is an illustrative example of projective geometry for two images; Figure 3 depicts the field of view of an optical camera, with incoming photons hitting the camera frame. The distance of those objects from the frame is unknown; Figure 4 shows the field of view of a 2D ultrasound probe. There the distance from the probe is known; Figure 5 shows the field of view of a 3D ultrasound probe; Figure 6 shows a typical intersection of the fields of view of two optical cameras, the intersection being a 3D volume; Figure 7 shows a typical intersection of the fields of view of two 2D ultrasound probes, the intersection being a ID line segment; and Figure 8 illustrates how a line segment within a 2D ultrasound image may be mapped to a curve in a high dimensional representation space using a trained machine learning model. In some implementations the invention describes a system to help a sonographer determine how much of an anatomical structure, such as an organ, has been scanned, and from which positions. We define "anatomical structure" to broadly include organs, their constituent components, and other anatomical details within the body. An anatomical structure may have other anatomical structures within it. The system may also help the sonographer figure out how to image the remaining unseen portions of the organ, by suggesting scanning positions that should best image the missing sections. The system may also suggest additional ways for the sonographer to view a chosen section of the image, to allow suspected lesions to be imaged from multiple viewpoints . The system may achieve this by applying a combination of photogrammetry and deep learning to real-time ultrasound images collected during a scan. Photogrammetry is the science and technology of obtaining reliable information about physical objects and the environment through the process of recording, measuring, and interpreting photographic images and patterns of electromagnetic radiant imagery and other phenomena. A sub field of it relates to generating 3-D reconstructions of a scene from a collection of 2-D photographs taken of the scene from different angles . There are three initial key steps in doing this: 1. Via some algorithm (commonly used examples include SIFT and ORB), identify "landmark" features within each of the images; 2. Identify instances where the same feature appears in multiple images. This is done by: a. Characterising the features via a f eatur e descriptor" which describes characteristics of their local neighbourhoods. These feature descriptors typically take the form of vectors (sometimes over R (e.g., SIFT) and sometimes over {0, 1} (e.g., ORB)) b. Identifying correspondences between the detected features through use of a feature matching algorithm. Figure 1 shows a "textbook" pair of images that illustrates this. 3. Determine the positions of the detected features in 3D space using algorithms that derive from projective geometry and which aim to derive camera positions and 3D locations of the features that would result in them projecting onto all image planes in a way that optimally reproduces their actual locations within the images. Figure 2 gives an example of the projective geometry for two images and one feature (though in reality one feature is insufficient to determine the relative positions of the cameras at least seven points are needed for that). In practice, doing this is rather complicated due to false positives when matching and because physical cameras behave slightly differently to the theoretical models assumed in the algorithms, but the outcome is a 3D reconstruction of the object and the positions of the cameras in 3-D space around it. In the context of the present invention, photogrammetric techniques are required to derive understanding of the regions in 3D space to which the ultrasound images that have been acquired relate. The known prior art, as described above, cannot be directly applied to ultrasound images for two key reasons. Firstly, the feature detectors that exist for natural images are unlikely to work well for ultrasound images. Essentially, they aim to detect sharply defined corners within the image (regions where intensity gradients are high in two directions), which is a useful heuristic for identifying the boundaries of objects in natural images, but is less useful in ultrasound images, both because of their noisy nature and because the boundaries of anatomical features are typically smooth curves. While some prior work has been done on feature detection for ultrasound, those methods may not be sufficiently effective for use in photogrammetry. Deriving a new learned feature detector (via Deep Learning) is considered to be the most promising approach to generating features of sufficient quality to support robust pose estimation. There are several existing generic approaches to this that are operable for this purpose when trained on suitable ultrasound data. Examples include https : / / arxiv.org / pdf / 1905.035 61.pdf and https : / / proceedings.neurips.cc / paper / 2018 / file / f5496252609c 4 3 e b 8 a 3 d 14 7 a b 9 b 9 c 0 0 6 - P a p e r . p d f . Secondly, the geometric properties of ultrasound images are different from those of optical images. An optical camera measures light intensities of rays that hit its sensor, with the intensities corresponding to reflection of photons off objects at unknown and varying distances from the sensor. So, the image data contains information about objects over two dimensions out in a direction normal to the sensor plane but does not contain information about how far away the objects are in that normal direction. By contrast, a 2D ultrasound image measures changes in attenuation (caused by differences in tissue density) across a ID-sector defined by the sensor array and over a direction normal to it. So, the distance of objects from the sensor is known, but the sensor array itself is 1-dimensional rather than 2 dimensional. With a 3D probe, the region imaged is a 3-dimensional volume in front of the probe head, and again distances out from the sensor are known. These distinctions are illustrated in Figures 3-5. The geometric difference between 2-D optical and 2-D ultrasound images has a large implication for photogrammetry since the intersection of the field of view of two images is typically one-dimensional, rather than three-dimensional as it is for optical cameras, as illustrated in Figure 6. This means that the feature matching process described in (2) (b) above will generally not work, since it is very unlikely that any detected features would lie on the line of intersection, and so there would be none to match (see Figure 7) . Where the ultrasound images are 3-D volumes instead, the intersection of two FOVs will again be 3-D, the likelihood of encountering matching features is similar to the optical case and process (2)(b) can progress as before. The process that follows it in (3) would still be different though since rather than needing to solve for projective transformations and points in 3-D space that are consistent with the matched features, instead a similarity transformation (in the mathematical sense) is found which gives the best mapping between the two point sets. To determine whether and how two 2-D ultrasound images intersect in 3-D space an alternative procedure is possible. Rather than detecting a discrete, sparse set of features in the image along with their representations, in this approach, the values of feature-vectors derived from a deep learning model are assessed as a line of pixels in the image is traversed, yielding a "trajectory signature" in a high dimensional feature space (see Figure 8). If two different ultrasound images taken of the same object from different positions intersect the traj ectory signatures that are derived from tracing the line within each image corresponding to the intersection will be the same. We can measure how similar any two trajectory signatures are by computing the Euclidean distance in feature space as we traverse both lines. We do this by specifying a sampling frequency n, and generating n pairs of points (one point on each line) obtained by stepping along the line in units of Vn times its length. For each pair we compute the Euclidean distance between the projections of the points into feature space, sum the results and divide by n. We attempt to find a line within each image for which the differences in trajectory signatures are as small as possible. We do this by using a non-linear optimisation algorithm to solve for anchor coordinates and direction vectors of line segments within each image (the direction vector determining both the direction and length of the line segment) that yield the lowest possible value of the distance metric described above. The planes are only deemed to intersect if that value is below some threshold. Note that this process will only determine how the two images relate to one another up to a rotation about the intersection line but adding a third image and applying the process pairwise would suffice to locate exactly the relative position of all three images. A key benefit of this system in relation to other possible approaches is that it does not require the ultrasound probe to be fitted with external position or orientation trackers. The methods described also have application to the construction of large 3-D ultrasound volumes by coherently piecing together multiple smaller ones. This is a common task when developing ultrasound simulators and automating the process would generate time savings. To assess the portion of an anatomical structure, e.g., an organ, that has been scanned, in addition to being able to locate the ultrasound images in 3D space, it is necessary to evaluate whether they contain an organ of interest, and if so what portion of it is in view. To do this, a deep learning system analyses the ultrasound images to identity the extents of, and where relevant, internal structures within, the organ, via semantic segmentation, for example. To determine how much of the organ has been scanned, the system places the segmented images in 3-D space via the process described above and determines a transformation of an idealised 3-D model of the organ which optimally overlaps the segmented images (for instance by maximising the sum of intersection over unions over all the images). As each scan is conducted, the system may track which parts of the organ have been seen, and which are unseen. Progress may be presented visually to the user in real-time, for example by displaying the percentage of the organ that has been viewed, or as a 3d model where the "seen" portions are coloured differently. The system may also contain a database of scanning approaches for the selected organ. Each entry will contain a description of how to perform the approach (e.g., "sagittal subcostal sweep") and the likely portion of the idealised virtual model covered by that approach. By comparing the unseen portion of the model with the entries in the database, the system can make suggestions about which approaches will best cover the remaining unseen portion of the organ. Likewise, the same database can be used to provide suggestions about how to view a suspected lesion in more than one plane. In this mode, the user will indicate the point on the current image that they wish to view (or optionally based upon the semantic segmentation information, if it can detect lesions) . The system will then query the database to find approaches that also cover the selected point from the desired orientation. The approaches described above may be used in conjunction with pathology detection methods (implemented, for example, as additional deep learning models) to produce a numeric indicator which quantifies the likelihood that no pathology is present in the site of interest. This can be achieved by quantifying the fraction of the volume of the region of interest that is covered by ultrasound images / volumes that are assessed to contain no pathology. The value of this indicator is preferably reported to the operator in realtime as scanning progresses. In some implementations, the inventive method is designed to indicate to the operator how much of an anatomical structure has been seen. As anatomical structures may themselves be composed of other anatomical structures, this method, may be applied at multiple levels. As an example, the liver is commonly divided into eight functionally independent segments. So, as well as providing an indication of how much of the entire liver has been seen, the methods described herein could help the operator identify which of these segments have been visualised, and which have not. As a further example, the method could be applied to the various components making up a kidney, such as the cortex, renal pelvis, renal artery, and renal vein. The method would allow feedback to be given to the operator to indicate which of these had been seen (and how much of each) , and which had not. Whilst the descriptions contained within this document are generally framed in terms of scanning of human subjects, the skilled person will recognise that the methods and apparatus described are equally applicable to scanning animal subj ects. 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 5 shown in the drawings, whether or not particular emphasis has been placed thereon. 10

Claims

1. An apparatus for analysing a plurality of ultrasound images or volumes captured from a 3D target space, the apparatus comprising an image processor and an artificial intelligence unit, wherein the processor is arranged to process the images or volumes and the artificial intelligence unit is arranged to determine their relative positions and orientations within the 3D target space.

2. The apparatus according to Claim 1 wherein, using the artificial intelligence unit, the apparatus is arranged to determine the relative positions and orientations of at least two 2D ultrasound images in 3D space by assessing values of feature vectors obtained from a trained machine learning model while tracing along lines in each image, calculating similarity measures between the resultant feature vector trajectory signatures and assessing whether the similarity measures are below a threshold value.

3. The apparatus according to Claim 1, wherein using the artificial intelligence unit, the apparatus is arranged to determine the relative positions and orientations of at least two 3D ultrasound volumes in 3D space by employing feature detectors to identify common features within each of them and then solving for similarity transformations between the feature sets that optimally align the detected features.

4. The apparatus according to any of Claims 1-3, whereinthe apparatus is arranged to segment one or more anatomical structures of interest within the images or volumes being analysed.

5. The apparatus according to Claim 4, whereinsegmentation of one or more anatomical structures isperformed by applying a trained machine learning model.The apparatus accordingwherein the locationsto any of the preceding claims,and orientations of segmentedultrasound images or volumes in 3D space are calculated,andaffine transformation3D modelananatomical structure isdetermined such thatitoptimally aligns it withthose segmented imagesvolumes .

7. The apparatus according to Claim 6, wherein the apparatus is arranged to further determine from the segmented image portions and their intersection with the 3D model, which parts of the target anatomical structure have been scanned and from which positions and / or orientations and / or which parts of the target anatomical structure have yet to be scanned from any given orientation.

8. The apparatus according to Claim 7, wherein the anatomical structure(s) compris(e) an organ of a body.

9. The apparatus according to Claim 7 or 8, wherein each ultrasound image or volume is analysed using a trainedmachine learning model to assess whether pathology is present within it, and a numeric indicator calculated that summarises the proportion of the volume of an anatomical structure covered by images or volumes that were assessed to exhibit no pathology.

10. The apparatus according to any of Claims 7-9, further including providing to the operator any subset of the following information: (i) which parts of the target anatomical structure(s) have been scanned, and from which directions (ii) The proportion of the volume of the anatomical structure(s) that has been scanned (iii) The proportion of the volume of the anatomical structure (s) that has been scanned from at least two different directions (iv) The proportion of the volume of the anatomical structure(s) that is covered by images or volumes that have been assessed to exhibit no pathologies .

11. The apparatus according to any of the preceding claims, wherein the apparatus is arranged in use to provide one or more prompts to position the ultrasound probe in particular location and / or orientation to best image any previously unseen portions of the target space.

12. A method for analysing a plurality of ultrasound images or volumes captured from a 3D target space, the method comprising processing the images or volumes using an electronic processor and determining their relative positions and orientations within the 3D target space, using an artificial intelligence unit.

13. The method according to Claim 12 wherein, using theartificial intelligence unit,the relative positionsand orientations of at least two 2D ultrasound images in 3D space is determined by assessing values of feature vectors obtained from a trained machine learning model while tracing along lines in each image, calculating similarity measures between the resultant feature vector trajectory signatures and assessing whether the similarity measures are below a thresholdvalue .

14. The method according to Claim 12 or 13, wherein using the artificial intelligence unit, the method determines the relative positions and orientations of at least two 3D ultrasound volumes in 3D space by employing feature detectors to identify common features within each of them and then solving for similarity transformations between the feature sets that optimally align the detected features.

15. The method according to any of Claims 12-14, wherein anatomical structures of interest within the images or volumes being analysed are segmented.

16. The method according to Claim 15, wherein segmentation of anatomical structures is performed by applying a trained machine learning model.

17. The method according to any of Claims 12-16, wherein the locations and orientations of segmented ultrasoundimages or volumes in 3D space are calculated, and an affine transformation of a 3D model of an anatomical structure is determined such that it optimally aligns it with those segmented images or volumes.

18. The method according to Claim 17, wherein the method further comprises determining from the segmented image portions and their intersection with the 3D model, which parts of the target anatomical structure have been scanned and from which positions and / ororientations and / or which parts of the target anatomical structure have yet to be scanned from any given orientation.

19. The method according to Claim 18, wherein theanatomical structure(s) compris(e) an organ of a body.

20. The method according to any of Claims 17 to 19,wherein each ultrasound image or volume is analysed using a trained machine learning model to assess whether pathology is present within it, and a numeric indicator calculated that summarises the proportion of the volume of an anatomical structure covered by images or volumes that were assessed to exhibit no pathology.

21. The method according to any of Claims 17-20, further including providing to the operator any subset of the following information: (i) which parts of the target anatomical structure(s) have been scanned, and from which directions (ii) The proportion of the volume of the anatomical structure(s) that has been scanned (iii)The proportion of the volume of the anatomical structure (s) that has been scanned from at least two different directions (iv) The proportion of the volume of the anatomical structure(s) that is covered by images or volumes that have been assessed to exhibit no pathologies.

22. The method according to any of Claims 12-21, wherein prompts are provided to position the ultrasound probe in particular location and / or orientation to best image any previously unseen portions of the target space.

23. A computer programme product on a computer readable medium, comprising instructions that, when executed by a computer, cause the computer to perform a method according to any of Claims 12-22.

24. A program for causing a device to perform a method according to any of Claims 12-22.

Images