FIELD OF INVENTION

The invention relates to the fields of artificial intelligence and computer assisted surgery. In particular, the invention relates to a system and a method for determining a spatial position, 3D orientation, and 3D representation of an object based on an X-ray projection image. The method may be implemented as a computer program executable on a processing unit of the device.

BACKGROUND OF THE INVENTION

In a case in which a bone is fractured, the pieces of the bone may be stabilized by an implant like an intramedullary nail, which may be inserted into the medullary canal of the bone, or a bone plate, which may be affixed to the surface of the bone, as support for the healing of the fracture. The surgical procedure for implanting such implants may be minimally invasive and may require repeated acquisition of X-ray images to enable the surgeon to correctly place the implant. The implant may also be connected to one or more sub-implants, e.g., a screw or a blade.

Steps in the surgical procedure may require determining the locations, shapes, and dimensions of objects (e.g., surgical tools, implants, or bone structures) depicted in an X-ray image. For example, the surgeon may need to determine the length of a blade for

implantation by measuring an anatomical structure shown in an X-ray image. However, a reliable measurement of dimensions based on an X-ray image requires calibration.

It is suggested in the art to utilize a reference body mounted to the X-ray imaging device in order to calibrate acquired X-ray images. Furthermore, it is suggested in the art to use at least one further reference body mounted to an instrumentation. Such a reference body may assist in determining the 3D position of an implant or surgical instrumentation. This allows determining and displaying the relative position of an implant with respect to a bone in a computer-assisted surgery system. Moreover, a reference body may also be required for matching images depicting the femoral head acquired from different directions in order to provide a 3D representation of the femoral head.

For example, US 2014/0343572 A1 describes a method for computer assisted determination of values of correction parameters for positioning a medical device relative to a target structure or to a reference base. US 2009/0209851 A1 describes a computer assisted surgical system which processes two or more two-dimensional images of a region of interest, taken at different angles, to produce three-dimensional information associated with the region of interest.

SUMMARY OF THE INVENTION

The invention as described herein suggests combining knowledge about the X-ray image generation process with artificial intelligence (in the form of so-called deep morphing and/or the utilization of a neural net) instead of any reference body for providing information needed when performing a treatment of, for example, a fractured bone. It may thus be seen as an object of the invention to provide a device and/or a method enabling a 3D representation and localization of at least one object at least partially visible in an X-ray projection image with respect to a coordinate system (given by, e.g., an image plane). Here, an object may be any object visible in an X-ray image, e.g. an anatomical structure, an implant, a surgical tool, and/or a part of an implant system.

The term“3D representation” may refer to a complete or partial description of a 3D volume or 3D surface, and it may also refer to selected geometric aspects, such as a radius, an axis, a plane, or the like. The present invention may allow the determination of complete 3D information about the 3D surface or volume of an object, but methods that determine only selected geometric aspects are also considered disclosed with this application.

Throughout this application, the terms“localize” and“localization” are intended to encompass a determination of a 3D orientation of an object with respect to a coordinate system and a determination of a 2D spatial position of the projection of that object onto an image plane or projection plane. The imaging depth (which is the distance of the imaged object from the image/projection plane), on the other hand, is estimated based on a priori information about typical constellations of objects in the operating room, e.g., relative positions of implant, patient, and imaging device. In the context of this invention, such an estimated imaging depth is sufficient. It is nevertheless possible to determine the imaging depth more precisely in certain cases, as discussed further below.

The invention does not require any reference body, which simplifies product development (e.g., if employing a new implant), is more cost-effective, and allows an operating room workflow that more closely resembles the typical workflow.

A 3D representation and localization of related objects like anatomical structures, implants, surgical tools, and/or parts of implant systems, may even be provided if not or only partially visible in the X-ray image. As an example, even if the projection image does not fully depict the femoral head, it may still be completely reconstructed.

It is noted that the image data of the processed X-ray image may be received directly from an imaging device, for example from a C-arm based 2D X-ray device, or alternatively from a data base. Aspects of the invention may also be used to process medical images acquired using other imaging modalities, such as ultrasound or magnetic resonance imaging.

The system suggested in accordance with the invention comprises at least one processing unit configured to execute a computer program product including sets of instructions causing the device (i) to receive an X-ray projection image whose characteristics depend on imaging parameters, (ii) to classify at least one object in the X-ray projection image, (iii) to receive a model of the classified object, and (iv) to determine a 3D representation of the classified object. Further, the classified object may be localized with respect to a coordinate system, by matching a virtual projection of the model to the actual projection image of the classified object. This process may consider the characteristics of the X-ray imaging method. In particular, the fact may be considered that the intercept theorem applies, as discussed in the examples later.

The X-ray projection image may represent an anatomical structure of interest, in particular a bone. The bone may for example be a bone of a hand or foot, but may in particular be a long bone of the lower extremities, like the femur and the tibia, and of the upper extremities, like the humerus. The image may also include an artificial object like a bone implant being already inserted into or affixed to the imaged anatomical structure of interest. In the context of the invention, it will be differentiated between an "object" and a "model".

The term "object" will be used for a real object, e.g., for a bone or part of a bone or another anatomical structure, or for an implant like an intramedullary nail, a bone plate or a bone screw, or for a surgical tool like a sleeve, k-wire, or aiming device, which may be connected to an implant. An“object” may also refer to only part of a real object (e.g., a part of a bone), or it may be an assembly of real objects and thus consist of sub-objects. For instance, an assembly of implant and aiming device may be considered to consist of the sub-object “aiming device” and the sub-object“nail”. As another example, the proximal part of a femur may be considered an object, and the distal part of the same femur may be considered another object.

On the other hand, the term "model" will be used for a virtual representation of an object. For example, a data set defining the shape and dimensions of an implant may constitute a model of an implant. As another example, a 3D representation of an anatomical structure as generated for example during a diagnostic procedure may be taken as a model of a real anatomical object. It should be noted that a“model” may describe a particular object, e.g., a particular nail, or it may describe a class of objects, such as a femur, which have some variability. In the latter case, such objects may for instance be described by a statistical shape or appearance model. It may then be an aim of the invention to find a 3D representation of the particular instance from the class of objects that is depicted in the acquired X-ray image. For instance, it may be an aim to find a 3D representation of the femur depicted in an acquired X-ray image based on a general statistical shape model of femurs. It may also be possible to use a model that contains a discrete set of deterministic possibilities, and the system would then select which one of these best describes an object in the image. For instance, there could be several nails in a database, and an algorithm would then identify which nail is depicted in the image (if this information is not provided by a user beforehand).

Since a model is actually a set of computer data, it is easily possible to extract specific information like geometrical aspects and/or dimensions of the virtually represented object from that data. The model may include more than one part of an imaged object, with at least one part not being visible in the X-ray projection image. For example, a model of an implant may include a screw intended to be used with an implant, but only the implant is already introduced into an anatomic structure and thus only the implant is visible in the X-ray projection image.

It is also noted that a model may not be a complete 3D model of a real object, in the sense that it only describes certain geometrical aspects of an object, such as the fact that the femoral head can be approximated by a ball in 3D and a circle in the 2D projection image.

The appearance of an object in a projection image may be affected by the X-ray imaging procedure. For example, imaging parameters like the imaging direction relative to gravity, a zoom, the radiation intensity, and/or a presence of a magnetic field may influence the appearance of an object in a projection image. Those or further imaging parameters may cause characteristic changes in the projection image like deformations of the projected object due to a pillow effect, mechanical bending of a C-arm imaging device depending on the imaging direction, a curvature, noise, and/or distortion. Here, those changes are denoted as "image characteristics".

It will be understood that it may be possible to determine those image characteristics in a projection image. For example, a position of a structure shown in an edge region of the image may be more affected by a pillow effect than a structure in the center of the image. In consequence, the characteristics of a pillow effect may be determined with a sufficient precision based on a structure of known shape that spans from an edge region to a central region. Image characteristics determined for a region in the 2D X-ray may be extrapolated to the entire image.

The effect of the earth’s magnetic field on image distortion depends on whether the image intensifier of the C-arm is positioned horizontally or vertically relative to gravity (due to the effect the magnetic field has on the electron beam inside the image intensifier, cf. M. Dotter, “Fluoroskopiebasierte Navigation zur intraoperativen Unterstiitzung orthopadischer

Eingriffe,” dissertation, TU Munich, 2004). If the patient positioning (e.g., lateral, left, prone, or supine) is known (e.g., provided as input by a user) and after the imaging direction has been calculated by the system (e.g., using the invention by Blau filed as patent application on 23.8.2018), it may be determined whether the C-arm is positioned horizontally or vertically relative to gravity. This provides an a priori information about the effect that the earth’s magnetic field has on the distortion in an X-ray image. Another example of a priori information that may be used by the system is the bending of the C-arm imaging device due to gravity.

Another approach to exploit the position of the C-arm may be to obtain an initial

measurement of the distortion of images with typical views (e.g., AP and ML) with a respective C-arm position, which can be performed before a patient is opened and can then be saved by the system. If the position of the C-arm is known (determined using the approach in the previous paragraph), the respective previously determined and saved image

characteristics may then be applied to the current image and object match.

Taking into account image and object characteristics as well as the effects of X-ray attenuation, absorption, and deflection, a virtual projection of a model may be deformed and/or distorted like the object is deformed and/or distorted in the X-ray projection image. Such a virtual projection may then be matched to the projection seen in the X-ray image. It will be understood that the matching of the object in the X-ray projection image to the model may include an adaptation of image characteristics of the X-ray projection image to image characteristics of the virtual projection of the model and/or an adaptation of image characteristics of the virtual projection of the model to image characteristics of the X-ray projection image. It is also understood that a match in the 3D projection volume, by minimizing distances in 3D, may also be possible.

The physical dimensions of an object are related to the dimensions of its projection in an X- ray image through the intercept theorem (also known as basic proportionality theorem) because the X-ray beams originate from the X-ray source (the focal point) and are detected by an X-ray detector at the image plane. The precise imaging depth (which is the distance of the imaged object from the image plane) is not generally required in the context of this invention. However, if an object is sufficiently large, the imaging depth may be determined through the intercept theorem, and the larger the object, the more precise this determination will be. Alternatively, the imaging depth may also be determined if the size of the X-ray detector and the distance between image plane and focal point are known. If, for some application, a more precise imaging depth is required, the size of the X-ray detector and the distance between image plane and focal point may be provided by user input or database. The former may also be determined by a calibration step by placing a known object (e.g., nail, k- wire, etc.) directly on the image intensifier/X-ray detector.

According to an embodiment, a deep neural net (DNN) may be utilized for a classification of an object in an X-ray projection image (e.g., proximal part of femur, distal part of femur, proximal part of nail, or distal part of nail, etc.). It is noted that a DNN may classify an object without determining its position (see, e.g., Krizhevsky, A., Sutskever, L, and Hinton, G. E. ImageNet classification with deep convolutional neural networks. In NIPS, pp. 1106-1114, 2012). A neural net may also be utilized for a rough classification of the imaging direction (e.g., AP vs. ML, cf. the paper: Aaron Pries, Peter J. Schreier, Artur Lamm, Stefan Pede, Jiirgen Schmidt: Deep morphing: Detecting bone structures in fluoroscopic X-ray images with prior knowledge, 2018, available online at https://arxiv.org/abs/1808.04441). Such a classification of an object and/or of an imaging direction may be used to select an appropriate model for following processing steps.

According to an embodiment, the outline of the classified object may be detected in the X-ray image. Lor objects with variable shape such as anatomical structures, this may proceed by using a“deep morphing” approach as described in the above cited paper by Pries et al.

(2018). This paper proposes an approach based on a deep neural network to detect bone structures in fluoroscopic X-ray images. The technique specifically addresses the challenges in the automatic processing of fluoroscopic X-rays, namely their low quality and the fact that typically only a small dataset is available for training the neural network. The technique incorporates high-level information about the objects in the form of a statistical shape model. The technique consists of a two-stage approach (called deep morphing), where in the first stage a neural segmentation network detects the contour (outline) of the bone or other object, and then in the second stage a statistical shape model is fit to this contour using a variant of an Active Shape Model algorithm (but other algorithms can be used as well for the second stage). This combination allows the technique to label points on the object contour. Lor instance, in the segmentation of a femur, the technique will be able to determine which points on the contour in the 2D X-ray projection image correspond to the lesser trochanter region, and which points correspond to the femoral neck region, etc. Objects described by a deterministic model (e.g., a nail) may also be detected by deep morphing, or simply by a neural segmentation network, as in the first stage of deep morphing. In a further step, taking into account image and/or object characteristics as well as the effects of X-ray attenuation, absorption, and deflection, a virtual projection of the model may be adjusted to match the appearance of the object in the X-ray projection image. According to an embodiment, for objects described by a deterministic model, this matching may proceed along the lines described in the paper: Lavallee S., Szeliski R., Brunie L. (1993) Matching 3- D smooth surfaces with their 2-D projections using 3-D distance maps. In: Laugier C. (eds) Geometric Reasoning for Perception and Action. GRPA 1991. Lecture Notes in Computer Science, vol. 708. Springer, Berlin, Heidelberg. In this approach, image characteristics and objects characteristics as well as the effects of X-ray attenuation, absorption, and deflection may be accounted for by introducing additional degrees of freedom into the parameter vector or by using a suitably adjusted model.

According to an embodiment, for objects described by a statistical shape or appearance model, the matching of virtual projection to the actual projection may proceed along the lines of the following paper: V. Blanz, T. Vetter (2003) Face Recognition Based on Fitting a 3D Morphable Model, IEEE Transactions on Pattern Analysis and Machine Intelligence. In this paper, a statistical, morphable 3D model is fitted to 2D images. For this, statistical model parameters for contour and appearance and camera and pose parameters for perspective projection are determined. It requires an initialization with key points, which may be provided by the result of the deep morphing step. Another approach may be to follow the paper: L. Gu and T. Kanade, 3D alignment of face in a single image, in Proc. of 2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition. This is a variant of an Active Appearance Model for 2D-3D matching, which also determines position of camera and statistical model parameters.

A neural net may be trained based on a multiplicity of data that is comparable to the data on which it will be applied. In case of an assessment of bone structures in images, a neural net should be trained on the basis of a multiplicity of X-ray images of bones of interest. It will be understood that the neural net may also be trained on the basis of simulated X-ray images. Simulated X-ray images may, for example, be generated from 3D CT data, as described in the appendix of the paper: Aaron Pries, Peter J. Schreier, Artur Lamm, Stefan Pede, Jiirgen Schmidt: Deep morphing: Detecting bone structures in fluoroscopic X-ray images with prior knowledge, available online at https://arxiv.org/abs/1808.04441. The realism of simulated X- ray images may be enhanced by using so-called generative adversarial networks (GANs). In this approach, a fully convolutional neural network may receive a simulated image as its input and then return a refined image. The refinement is learned with an adversarial loss function that rewards a realistic refinement. This is described in the paper: Ashish

Shrivastava, Tomas Pfister, Oncel Tuzel, Josh Susskind, Wenda Wang, Russ Webb: Learning from Simulated and Unsupervised Images through Adversarial Training, available online at https://arxiv.org/abs/1612.07828.

According to an embodiment, more than one neural network may be used, wherein each of the neural nets may specifically be trained for a sub-step necessary to achieve a desired solution. For example, a first neural net may be trained to evaluate X-ray image data so as to classify an anatomical structure in the 2D projection image, whereas a second neural net may be trained to detect the location of that structure in the 2D projection image. A third net may be trained to determine the 3D location of that structure with respect to a coordinate system.

It is also possible to combine neural networks with other algorithms, including but not limited to, Active Shape Models. It is noted that a neural net may also learn to localize an object or to determine an imaging direction without the need to first detect the outline of the object in the 2D X-ray image. It is also noted that a neural net may also be utilized for other tasks, e.g., a determination of one or more image characteristics like a pillow effect.

According to an embodiment, an object may also be manually classified and/or identified in the X-ray projection image. Such a classification or identification may be supported by the device by automatically referring to structures that were recognized by the system.

According to an embodiment, the system may compute geometrical aspects of an object (e.g., an axis, an outline, a curvature, a center point), and dimensions of an object (e.g., a length, a radius or a diameter, a distance). This may be accomplished due to the correspondence between the model and the virtual projection that has been matched to the projection seen in the X-ray image.

When displaying the X-ray projection image, geometrical aspects and/or dimensions may be shown as an overlay in the projection image. Alternatively and/or additionally, at least a portion of the model may be shown in the X-ray image, for example as a transparent visualization or 3D rendering, which may facilitate an identification of structural aspects of the model and thus of the imaged object by a user. It will be understood that sizes may also define a scale suitable for measuring dimensions of and between objects visible in the X-ray projection image. For example, a diameter of a matched object (called“Object A”) as indirectly provided by model data, may be utilized for determining a size of other depicted and potentially unknown objects, at least for those objects that can be assumed as positioned at a sufficiently close imaging depth. It may even be possible to calculate a size of a different object (called“Object B”) at a different imaging depth based on the intercept theorem if the imaging depth of Object A is known (e.g., because Object A is sufficiently big or because the size of the X-ray detector and the distance between image plane and focal point is known) and if there is information about the differences in imaging depths between Objects A and B (e.g., based on anatomical knowledge).

The present invention may also provide for a 3D reconstruction and localization of an object whose shape and appearance have some variability. Such objects may for instance be described by a 3D statistical shape or appearance model. This can be done based on a single X-ray image or multiple X-ray images.

Based on one image of an anatomical object, the model is deformed in such a way that its virtual projection matches the actual projection of the object in the X-ray image. Doing so allows a computation of an imaging direction (which describes the direction in which the X- ray beam passes through the object). As an additional plausibility check, the computed imaging direction may then be compared with the imaging direction for the same object that is determined based on the methodology using the invention by Blau filed as patent application on 23.8.2018, which may take the entire X-ray image (and not just the particular object) into account. This enables a plausibility check because these two imaging directions should not differ.

If multiple X-ray images are acquired, the information from them may be fused (or registered) to increase the accuracy of 3D reconstruction and/or determination of spatial positions or orientations. It is preferable if these X-ray images are acquired from different imaging directions because this may help resolve ambiguities. The more different the imaging directions are (e.g., AP and ML images), the more helpful the images may be in terms of a determination of 3D information. For the registration of a pair of X-ray images showing the same object (e.g.“Object C”) from two different imaging directions, it may be possible to increase the accuracy of the registration process by taking into account the 3D angle (which may, for instance, be represented by the Euler angles or another suitable representation) between the imaging directions. One way of determining this angle would be to determine the imaging directions using the invention by Blau filed as patent application on 23.8.2018 for each X-ray image and to compute their difference. Another way may be to utilize another object (e.g.“Object D”), also shown in both images, whose model is deterministic (e.g., a nail connected to an aiming device). By matching the virtual projection of Object D to its actual projection in each X-ray image, the imaging directions for Object D may be determined. This requires that either (i) there be no movement of Object C relative to Object D between acquiring the two X-ray images, or (ii) in case there is a movement, it can be determined as discussed later. This procedure may obviously be extended to register more than two images. This allows a determination of the relative 3D position and orientation between multiple objects depicted in the same X-ray image, even though the imaging depth has not been determined.

The appearance of an object in an X-ray image depends inter alia on attenuation, absorption, and deflection of X-ray radiation, which depend on the object’s material. The more material the X-ray beam must pass through, the less X-ray radiation is received by the X-ray detector. This affects not only the appearance of the object within its outline, but it may also change the shape of the outline itself in the X-ray projection image, in particular in areas where the object is thin.

The strength of this effect also depends on the X-ray intensity and the amount of tissue surrounding the object that the X-ray beam must pass through. The latter depends on the body mass index of the patient and the imaging direction (e.g., using the invention by Blau filed as patent application on 23.8.2018). The amount of soft tissue surrounding the object could be derived from a database, which considers, e.g., ethnicity, gender, body mass index, age.

"Object characteristics", as already mentioned above, may for example be the material of an intramedullary nail, wobbling of connections, or a fracture of a bone. Those object characteristics may be determined automatically, manually, or semi-automatically, wherein the device may support the manual determination for example by indications of structural aspects which would not be expected when considering only "image characteristics".

On the other hand, object characteristics that are known a priori may be utilized to increase accuracy of localization and determination of 3D representation. An example would be wobbling of connections, e.g., between an implant and an aiming device, which may be described by a model. It is understood that when object characteristics are used in a computation of localization and determination of 3D representation, some characteristics may be weighted higher than others.

It is noted that a processing unit may be realized by only one processor performing all the steps of the process, or by a group or a plurality of processors, which need not be located at the same place. For example, cloud computing allows a processor to be placed anywhere. For example, a processing unit may be divided into (i) a first sub-processor on which a first neural net is implemented assessing the image data including a classification of anatomical structures like a bone surface, (ii) a second sub-processor on which a second neural net is implemented specialized for determining an imaging direction of the classified anatomical structure, and (iii) a further processor for controlling a monitor for visualizing results. One of these or a further processor may also control movements of, for example, a C-arm of an X- ray imaging device.

According to an embodiment, the system may further comprise storage means providing a database for storing, for example, X-ray images. It will be understood, that such storage means may also be provided in a network to which the system may be connected, and that data related to the neural net may be received over that network.

Furthermore, the system may comprise an imaging unit for generating at least one 2D X-ray image, wherein the imaging unit may be capable of generating images from different directions.

The system may further comprise input means for manually determining or selecting a position or part of an object in the X-ray image, such as a bone outline, for example for measuring a distance in the image. Such input means may be for example a computer keyboard, a computer mouse or a touch screen, to control a pointing device like a cursor on a monitor screen, which may also be included in the system.

A computer program may preferably be loaded into the random access memory of a data processor. The data processor or processing unit of a system according to an embodiment may thus be equipped to carry out at least a part of the described process. Further, the invention relates to a computer-readable medium such as a CD-ROM on which the disclosed computer program may be stored. However, the computer program may also be presented over a network like the World Wide Web and can be downloaded into the random access memory of the data processor from such a network. Furthermore, the computer program may also be executed on a cloud-based processor, with results presented over the network.

It has to be noted that embodiments are described with reference to different subject-matters. In particular, some embodiments are described with reference to method-type claims (computer program) whereas other embodiments are described with reference to apparatus- type claims (system/device). However, a person skilled in the art will gather from the above and the following description that, unless otherwise specified, any combination of features belonging to one type of subject-matter as well as any combination between features relating to different subject-matters is considered to be disclosed with this application.

The aspects defined above and further aspects, features and advantages of the present invention can also be derived from the examples of the embodiments to be described hereinafter and are explained with reference to examples of embodiments also shown in the figures, but to which the invention is not limited.

BRIEF DESCRIPTION OF THE DRAWINGS

Fig. 1 shows an example for a 3D registration of AP and ML images.

Fig. 2 shows an example for a 3D registration of AP and ML images and illustrates the effect of an incorrectly estimated C-arm width.

Fig. 3 compares the situations of Figs. 1 and 2.

Fig. 4 shows an example for a 3D registration of AP and ML images and illustrates the effect of a zoom.

Fig. 5 compares the situations of Figs. 1 and 4. Fig. 6 shows an example for a 3D registration of AP and ML images and illustrates the effect of the X-ray receiver size.

Fig. 7 shows an undistorted proximal AP X-ray of a femur with nail and aiming device.

Fig. 8 shows a distorted proximal AP X-ray of a femur with nail and aiming device.

Fig. 9 compares Figs. 7 and 8.

Fig. 10 shows an undistorted proximal ML X-ray of a femur with nail and aiming device.

Fig. 11 shows a distorted proximal ML X-ray of a femur with nail and aiming device.

Fig. 12 compares Figs. 10 and 11.

Fig. 13 illustrates the effect that ignoring distortion has on the localization of the nail and aiming device.

Fig. 14 shows an incorrect 2D match of the outline of nail and aiming device.

Fig. 15 is a zoom of part of Fig. 14.

Fig. 16 is a zoom of part of Fig. 14.

Fig. 17 shows an example for a 3D registration of AP and ML images and illustrates the effect of receiving a PA image instead of an AP image.

Fig. 18 illustrates the effect of receiving a PA image instead of an AP image in the projection plane.

Fig. 19 illustrates the effect of receiving an LM image instead of an ML image in the projection plane.

Fig. 20 shows an example for a possible workflow.

Throughout the drawings, the same reference numerals and characters, unless otherwise stated, are used to denote like features, elements, components, or portions of the illustrated embodiments. Moreover, while the present disclosure will now be described in detail with reference to the figures, it is done so in connection with the illustrative embodiments and is not limited by the particular embodiments illustrated in the figures.

DETAILED DESCRIPTION OF THE EMBODIMENTS

3D reconstruction of an anatomical object based on one image

The above cited paper on Deep Morphing by Pries et al. (2018) proposes a method that enables a system to detect (in a 2D projection image) the outline/contour of a bone and label points on the contour. For instance, in the segmentation of a femur, the technique is able to determine which points on the contour in the 2D X-ray projection image correspond to the lesser trochanter, and which points correspond to the femoral neck, etc. Given a 3D statistical shape or appearance model of the same anatomical structure, this model can then be deformed in a way that its virtual projection matches the actual projection in the X-ray image, hence leading to a 3D reconstruction of the anatomical structure and allowing a localization of the object and determination of the imaging direction.

Being able to provide a 3D reconstruction of anatomy based on one X-ray image only is an advantage over the state-of-the-art, which requires the acquisition of at least two images from different viewing directions (typically an AP and an ML image). Moreover, the present invention is also able to localize an implant or instrument relative to anatomy based on one X-ray image.

This allows addressing surgical procedures where acquiring a second image from a different imaging direction is not feasible, e.g., because a fracture is not stable enough.

The accuracy of the 3D reconstruction of an anatomical object may be improved upon by a registration of multiple images acquired from different viewing angles (to be discussed in the next section) or by using essential a priori information for the above described method based on one image only. This a priori information may be the size or the gender of the patient, but also more anatomy- specific information like geometric information about the anatomical object to be reconstructed. For the example of a 3D reconstruction of a proximal femur based on an ML image, this information may include the length of the femoral neck or the CCD angle. Such information significantly reduces ambiguities in the 3D reconstruction. However, because such information may not be determined with sufficient precision in typical ML images, it may be extracted from AP images that are routinely acquired earlier in the course of the surgery on the proximal femur. The more information is used from earlier images, the more accurate a 3D reconstruction may be. Another way of describing this procedure would be to say that, based on an AP image, a 3D reconstruction of the proximal femur may be performed with typical remaining uncertainties (such as the width of the femoral neck in AP direction), and this 3D reconstruction serves as a priori information or as a starting point for a 3D reconstruction based on a later ML image. New registration process of two or more X-ray images from different directions

Depending on the bone shape there may be a remaining ambiguity or matching error in the 3D reconstruction based on one image only. This may be alleviated by acquiring a series of images, potentially from different viewing directions, which may be routinely acquired during surgery anyway (e.g. when repositioning the fracture). In general, additional images from different imaging directions are more helpful, and the more different the imaging directions are (e.g., AP and ML images), the more helpful additional images may be in terms of a determination of 3D information. However, even adding images from only slightly different viewing angles, which may be more easily acquired during surgery instead of changing to completely different view (AP to ML or vice versa), may be beneficial.

Finally, another way to increase precision in the matching procedure may be to use preoperative imaging to generate information about the 3D shape of the patient’s specific bone, instead of working with general statistical models that describe the variability of bones.

The invention allows to register multiple X-ray images of at least one common object taken from different directions. This is important because 3D registration allows a determination of relative 3D positions between multiple objects without an explicit determination of the imaging depth.

For the 3D reconstruction of an object of variable shape (typically an anatomical structure described, e.g., by a statistical shape or appearance model and called“Object F” in this section) based on two or more X-ray images, the procedure outlined above for one image may be extended to two or more images. That is, Deep Morphing may be used to detect the contour of Object F and label points on its contour in each 2D X-ray image. Given a 3D statistical shape model of Object F, this model can then be deformed in a way that its virtual projections simultaneously match the actual projections of Object F in two or more X-ray images as closely as possible. This procedure does not need a priori information about the imaging directions because it implicitly determines the imaging direction for each X-ray image.

As an alternative for the registration of a pair of X-ray images taken from two different imaging directions, it may be possible to increase the accuracy of the registration process by taking into account the 3D angle between the imaging directions, which may be determined using two different procedures. The more precisely this angle can be determined, the more precise the 3D registration may be.

One way of determining this angle would be to determine the imaging directions using the invention by Blau filed as patent application on 23.8.2018 for each X-ray image and to compute their difference. Another way may be to utilize another object in the X-ray image (called“Object G”) whose model is deterministic (e.g., a nail connected to an aiming device). By matching the virtual projection of Object G to its actual projection in each X-ray image, the imaging directions for Object G may be determined.

If the 3D angle between imaging directions computed from Object G differ significantly from the 3D angle between imaging directions for Object F, this could indicate, for instance, that Object G’s position relative to Object F has changed between acquiring the images (e.g., there has been a rotation around the nail axis). It may be possible to automatically correct for such a movement; if this is not possible, no 3D registration should be attempted. If the angles determined by the two procedures are sufficiently close, the results could be weighted and averaged (with a weighting depending, e.g., on image quality, detection quality, visibility, 3D reconstruction uncertainties), and a 3D registration may be performed. This procedure may obviously be extended to register more than two images.

In the following, the influence of C-arm width, size of image detector, zoom, etc. on a 3D registration will be illustrated with examples. It is shown that in all of these examples determination of imaging depth is not required.

Influence of C-arm width: Figure 1 depicts the left femur (denoted LF) and the nail implant with attached aiming device (denoted NAD). Furthermore, it shows the AP X-ray image (denoted l.AP) and ML X-ray image (denoted l.ML) and their corresponding focal points (denoted l.FP.AP and l.FP.ML). The 3D ball approximates the femoral head (denoted FH), and the dashed white circles are its 2D approximated projections in the images (denoted l.FH.AP and l.FH.ML). The C-arm has a width (here defined as the distance between focal point and image plane) of 1000 mm. The cones indicate the part of the X-ray beam passing through the femoral head. It is noted that throughout this application, we follow the convention to call images taken in a posterior-anterior direction“AP” images, and images taken in an anterior-posterior direction“PA” images. Similarly, we call images taken in lateral-medial direction“ML” images, and images taken in medial-lateral direction“LM” images.

In Fig. 2, instead of the true 1000 mm, the C-arm width was incorrectly estimated as 900 mm. Hence, all objects in the image, including the femoral head (FH), appear smaller in the X-ray images than they should. Therefore, it seems as if the objects were shifted towards the AP image plane (denoted 2.AP) as well as towards the ML image plane (denoted 2.ML). The corresponding focal points are denoted 2.FP.AP and 2.FP.ML. A 3D reconstruction of the femoral head (FH) based on the 2D projections of the approximated femoral head (white circles 2.FH.AP and 2.FH.ML) remains unchanged compared to Fig. 1. The only parameter that is changed is the apparent imaging depth. The imaging depth, however, is not relevant in this scenario because the relative 3D position of femoral head and nail has not changed.

In order to illustrate that the only difference between Fig. 1 and Fig. 2 is the apparent imaging depth, Fig. 3 shows both scenarios simultaneously.

Influence of zoom: If one of the images was captured with a zoom factor, the objects appear bigger than without zoom. For Fig. 4, the AP image (denoted 4.AP) was captured with a zoom factor of 1.5. Hence, all objects in the image, including the femoral head (FH), seem as if they had been moved towards the focal point in AP (denoted 4.FP.AP). As before, a 3D reconstruction of the femoral head (FH) based on the 2D projections of the approximated femoral head (dashed white circles 4.FH.AP and 4.FH.ML) remains unchanged compared to Fig. 1. The only parameter that is changed is the apparent imaging depth. The imaging depth, however, is not relevant in this scenario because the relative 3D position of femoral head and nail has not changed. Analogous comments apply when both images have a zoom. Figure 5 compares the situation with zoom (as in Fig. 4) and without zoom (as in Fig. 1).

Influence of size of X-ray detector: If the assumed size of the X-ray detector is 12” instead of the true 9”, the objects appear bigger in the image, and it seems as if the objects had been moved towards the focal points in both images. This is shown in Fig. 6, where:

• 6.AP.9” refers to the AP image with 9” X-ray detector with focal point denoted

6.FP.AP.9” • 6.AP.12” refers to the AP image with 12” X-ray detector with focal point denoted 6.FP.AP.12”

• 6.ML.9” refers to the ML image with 9” X-ray detector with focal point denoted 6.FP.ML.9”

• 6. ML.12” refers to the ML image with 12” X-ray detector with focal point denoted 6.FP.ML.12”

The effect is equivalent to a zoom factor that is applied to both images. Hence, the same conclusions as in the case of zoom may be drawn.

Influence of gravity on ML images: When capturing ML images with a C-arm based X-ray imaging device, the focal point moves towards the floor by 8 mm according to the dissertation by Dotter (2004). With a given C-arm width of 1000 mm, this corresponds to a rotation of the C-arm by approximately 1 degree. This rotation causes a vertical shift and a small rotation of the objects depicted in the image. Since the shift and rotation is detected during the reconstruction of the scene (based on the estimated location of the implant), neglecting this effect causes a small error with respect to the center of the reconstructed femoral head. Moreover, it is possible to consider the gravity effect during reconstruction to avoid this error.

Measuring a feature of a classified object

The current invention does not require an a priori calibration. Measurements may be performed in mm if there is a known object in the image located in the vicinity of (at a similar depth as) the structure to be measured. Since the known object has known

dimensions, it can be used for calibrating measurements. This is similar to the procedure proposed by Baumgaertner et al. to determine a TAD value (cf. Baumgaertner MR, Curtin SL, Lindskog DM, Keggi JM: The value of the tip-apex distance in predicting failure of fixation of peritrochanteric fractures of the hip. J Bone Joint Surg Am. 1995, 77: 1058- 1064.).

Example: A nail has been inserted, and an AP image is available. The nail has been identified and localized. Since the nail is located in the middle of the shaft and thus at a similar imaging depth as the depicted lateral cortex of the shaft, the known nail geometry can be used for calibration. This allows to provide a scaling for determining the distance between the nail axis and the lateral cortex of the shaft.

Handling image distortion for the example of an intramedullary nail

In general, there are two ways of handling distortion of images:

1. Deemphasizing regions in the X-ray image where distortion is known to be strong (e.g., border of images), placing more emphasis on regions less affected by distortion

2. Determining distortion and accounting for it

These will now be illustrated at the example of an AP image of a femur with inserted nail.

Re 1. The following letters are used in the labeling of Fig. 9. The solid line is the contour of a nail and aiming device as seen in a distorted X-ray image. The white dashed line shows the hypothetical outline of the nail and aiming device as they would be shown in an image without distortion.

9.D: Distal part of intramedullary nail

9.C: Central part of nail, including hole for neck screw

9.P: Proximal part of intramedullary nail

9. A: Aiming device

Typically, 9.D is located in a more distorted region of the X-ray image. Moreover, the precise location of 9.D is not as important when forecasting a trajectory for a screw inserted through the hole at 9.C. Thus, in a forecast of a screw trajectory, the locations of 9.C and 9.P may receive a higher weighting than 9.D, where the exact weighting may be determined based on their visibility and reliability of detection. A higher weighting on 9.C and 9.P may also be justified because these regions are closer to the region of interest (the screw hole and femoral head). Moreover, the appearance of 9.C carries information about the rotation of the nail around its axis.

Re 2. Distortion in an image may be determined by:

a) surgeries performed earlier (could be learned for a specific C-arm) b) calibration before surgery: a known object (e.g., nail, k-wire, etc.) could be placed directly on the image intensifier/X-ray detector at a known distance to the image plane. This may also be used for determining the size of the X-ray detector and the distance between focal point and image plane.

c) images acquired earlier (could be learned by an algorithm during a surgery)

d) a database with typical distortion effects (e.g., typical pillow effect, earth’s magnetic field, for typical C-arm positions). The device may use the knowledge that digital X-ray machines do not distort.

If such information is available, it may be utilized when matching a virtual projection of a model to a projection in the X-ray image. The distortion may be applied to the entire image, or specifically to the shape that is being matched.

The distortion depends strongly on the position of the C-arm imaging device in physical space, due to the earth’s magnetic field. In a typical operating room setting, the position of the C-arm imaging device may be obtained from the imaging direction if the positioning of the patient (e.g., prone or supine) is known. The imaging direction in turn may be determined, e.g., using the invention by Blau filed as patent application on 23.8.2018.

In the following, an example of the pillow effect in the case of an AP image of the proximal femoral nail is discussed. Figure 7 depicts a typical AP image without distortion, as can be seen from the dotted rectangular grid.

Figure 8 shows the same scene with a simulated pillow effect distortion. This is the distortion type that has the biggest influence on AP images, according to the dissertation by Dotter (2004). The distortion can be seen from the distorted dotted grid. It is clearly visible that pixels in the vicinity of the image borders are more affected than pixels in the image center.

Figure 9 again shows the same distorted image as in Fig. 8, but without the grid. The solid line is the contour of the nail and aiming device as seen in the X-ray image (denoted by 9. NAD). The white dashed line shows the hypothetical outline of the nail and aiming device, as they would be shown in an image without distortion.

In the following, an example of sinusoidal distortions is discussed in the case of an ML image of the proximal femoral nail. According to the dissertation of Dotter (2004), ML images suffer from sinusoidal distortion (due to the earth’s magnetic field) in addition to the pillow effect. The pillow effect may be determined based on an AP image, and then a correction may be applied to an ML image. Assuming that this correction is correct, only a sinusoidal distortion would remain in the ML image.

Figure 10 depicts an ML image without distortion, as can be seen from the dotted rectangular grid. Figure 11 shows the same scene with sinusoidal distortion, as can been seen from the distorted grid.

Figure 12 again shows the same distorted image as in Fig. 11, but without the grid. The solid line is the contour of the nail and aiming device as seen in the X-ray image (denoted by

12. NAD). The white dashed line shows the hypothetical outline of the nail and aiming device, as they would be shown in an image without distortion.

Figure 13 shows the consequences of localizing the nail and aiming device based on a distorted X-ray AP image of the proximal femur. The X-ray projection image 13.XRAY is the same as the one depicted in Fig. 9. If distortion is not taken into account, this would lead to a localization of the nail and aiming device at the incorrect position indicated by

13.NAD.IC, whereas the correct location would be at the position indicated by 13.NAD.C. While the incorrectly determined imaging depth is not relevant in the context of this invention, the incorrect 3D orientation on the other hand is problematic. In the shown situation, there is an error of approximately 4.5 degrees with respect to the nail axis.

However, Figure 14 shows that the match of the 2D outline in the corresponding X-ray image (the solid white line) that leads to the incorrect localization in Fig. 13 is itself only a poor match of the true outline. This can be more clearly seen when zooming into the image, as is shown in Fig. 15 and Fig. 16. For instance, in Fig. 15, there are clearly several gaps between the nail appearing in black and the determined outline (the solid white line, denoted 15.0), some of which are indicated by 15. GAP. Moreover, in Fig. 16, there is also clearly a gap visible between the nail appearing in black and the determined outline (the solid white line, denoted 16.0), in particular at the nail hole, as indicated by 16. GAP. The gap 16.GAP is explained by an incorrectly determined nail rotation. By combining the two methods described above (focus on less distorted region around the nail hole, and taking into account the type of the expected distortion especially in the border of the image where the aiming device is located) the accuracy of localization may significantly be improved. The present invention allows for an interaction between dewarping and matching process, leading to more precise results.

Handling an exchange of the positions of X-ray source and receiver

Because X-ray imaging devices allow mirroring of images and this invention does not use a calibration reference body attached to the image detector throughout the surgery, an exchange of the positions of X-ray source and receiver may not be detected, even if the treatment side (left or right bone) is known. A user could be required to provide information about whether or not the mirroring function is activated. If this is not desired, in the following, situations are discussed how a correct 3D reconstruction of the scene is nevertheless possible.

In case of a proximal AP image of the femur or humerus, an exchange of the positions of X- ray source and receiver cannot be detected, but such an exchange has no significant effect on a correct 3D reconstruction due to symmetries of proximal femur or humerus. Figure 17 shows the scene from Fig. 1, where the original imaging direction is AP with focal point 17.FP.AP, leading to AP image denoted 17.AP. In addition, Fig. 17 also shows a PA image, denoted 17. PA, which is an approximate mirror image of 17.AP (it is to be noted that in the figure, the image 17. AP is shown from the front, and the image 17. PA from the back). The image 17. PA is generated by the C-arm if its focal point is at location 17.FP.PA. In the 3D scene, the PA configuration is obtained from the AP configuration by mirroring with respect to the plane defined by femur shaft axis and femoral head center. It can be seen that both AP and PA images will lead to the same 3D reconstruction of the femoral head (FH).

The explanation is that the nail implant looks very similar from either AP or mirrored PA direction when keeping it at a fixed distance from the image plane. Furthermore, the proximal femur has a certain symmetry. For instance, if a mirrored image is received, it would be concluded that the other side is being treated (e.g., right instead of left) if the current C-arm orientation is kept. However, the projected 2D outlines of both bones (i.e., left and right) are quite similar. Figure 18 illustrates this statement. It shows a typical AP imaging scenario for a left femur. It also depicts the right femur at the same imaging depth. For both left and right bones, the outline of the nail implant in the 2D projection plane (denoted 18.PROJ) would be essentially identical. Moreover, the 2D bone outlines in the 2D projection plane differ only slightly, as can be seen by comparing the dashed and solid lines in the plane 18.PROJ.

Since in a typical ML image the nail has a lower distance to the image plane than the aiming device, these assumptions no longer hold when switching from ML to LM with an additional mirroring. Figure 19 depicts the outlines of aiming device and nail in the projection plane (denoted 19.PROJ) for an ML image (thick solid line) and the corresponding mirrored LM image (thick dashed line). It can be seen that these outlines differ significantly. Figure 19 also depicts the outlines of a femur for an ML view (thin solid line) and a mirrored LM view (thin dashed line) in the projection plane 19.PROJ. Within the field of view of the X-ray receiver (the rectangle denoted 19.XRAY), the outlines of the bones are almost identical. However, the outlines of aiming device and nail differ sufficiently within the rectangle 19.XRAY such that it is easily possible to distinguish between an ML image and a mirrored LM image.

Example for a potential processing workflow treating a hip fracture

The flow-chart in Fig. 20 illustrates the principle of the steps performed in accordance with an embodiment of the disclosed invention. It will be understood that the steps described are major steps, wherein these major steps might be differentiated or divided into several sub steps. Furthermore, there might be also sub-steps between these major steps. It will also be understood that only part of the whole method may constitute the invention, i.e. steps may be omitted.

Step 1: User provides information about procedure. This could include information about, for example, which procedure is attempted (e.g., trochanteric nail implantation or retrograde femur nail implantation), which implant will be used (e.g., nail manufacturer, type, and material), which side of the patient will be treated (left or right), how the patient is positioned, the patient’s size and body mass index etc. Based on these inputs, the system selects the appropriate model database.

Step 2: The system waits until a new X-ray image has been acquired. A new X-ray image is detected by performing an image difference analysis.

Step 3: Image viewing direction (AP or ML) is determined by applying a neural net to the depicted anatomy (using the invention by Blau filed as patent application on 23.8.2018). Step 4: Objects in image are classified. For this example, the device determines that image shows a nail assembly (nail, aiming device, and interface) and a proximal femur. Appropriate models are selected in the database.

Step 5: A deep-morphing approach detects the outline of the nail assembly (nail, aiming device, and interface) in the X-ray image, taking into account the viewing direction (AP or ML).

Step 6: A neural net computes a range of possible values for

a. image characteristics such as pillow effect and earth’s magnetic distortion function

b. object characteristics such as wobbling connection between nail and aiming device,

based on the appearance of the nail assembly in the X-ray image and any available a priori information, including information gained from processing images acquired earlier (if any exist).

Step 7: A matching algorithm matches a virtual projection of a 3D model of the nail assembly to the actual projection of the nail assembly in the 2D X-ray. This matching procedure finds the best tradeoff between a good visual fit on the one hand and using plausible values for the image characteristics and object characteristics, as determined in Step 6 on the other hand. The matching procedure also considers X-ray attenuation, absorption, and deflection.

Step 8: A deep-morphing approach detects and labels the outline of the proximal femur. A matching algorithm matches a virtual projection of a 3D statistical model of the proximal femur to the actual projection of the femur in the 2D X-ray. This considers all the information gathered from having processed the nail in Steps 5 through 7, in particular, extrapolated image characteristics.

Step 9: The correspondence between the 3D model of the nail assembly and its virtual projection is used to calibrate the image (i.e., determine the pixel size) at the imaging depth where the nail is located. This may be used for displaying a virtual scale in the X-ray image. Step 10: A visualization of the 3D model of the nail assembly and/or the proximal femur may be shown in the X-ray projection image. This could be the outline of their 2D projection or a rendering of the 3D model. It may also include additional parts that are not visible in the image, e.g. the axis of a neck screw not yet implanted.

Step 11: Steps 2 through 10 may be repeated. When processing a new image, the system considers any information (in particular, about image and object characteristics) it has gathered from previously processed images. Step 12: As soon as one AP and one ML image are available, the system attempts to register these images by utilizing the nail assembly as a reference in each image and taking into account the geometry of the imaging procedure (cf. Section“New registration process of two or more X-ray images from different directions”). The system computes the 3D angle between the imaging directions of these two images.

Step 13: The 3D angle computed in Step 12 may be compared to the 3D angle that is obtained by computing the difference between the imaging directions determined in Step 3 based on anatomy. If the 3D angle computed in Step 12 differs significantly from the angle computed based on anatomy, the system concludes that there has been a movement of the nail relative to the anatomy in between the acquisition of the two images, and it will thus inform the user and deny registration.

Step 14: If the registration is considered successful, it enables the fusion of information from these two images, leading to enhanced accuracy of the 3D representation of the proximal femur (e.g., because ambiguities in a single image can be resolved by the other image). The system may now provide a more accurate 3D reconstruction of anatomy and other objects (such as, e.g., nail or screw) and display these in the X-ray either as an outline or a 3D rendering.

Benefits of being able to determine the imaging direction based on different approaches

As mentioned before, the present invention and using the invention by Blau filed as patent application on 23.8.2018 allow a determination of the imaging direction based on

independent approaches utilizing different information. This may be used (i) to cross-validate the results obtained by the different approaches and/or (ii) to enhance the precision of computing an imaging direction by appropriately fusing the results obtained by the different approaches. In the following, examples showing the benefits of this approach are shown.

Example 1: A critical question that may arise in the processing of a series of X-ray images is whether there has been a movement of an object relative to another object also depicted in the X-ray image, e.g., nail relative to anatomy. If there has been a movement, it may be necessary to compute it. A complicating factor is that, when an object moves in the X-ray projection image, it may be due only to a movement of the C-arm imaging device. The present invention allows differentiating between a movement of the C-arm imaging device and a movement of an object. In the following, it is explained how a movement of the nail relative to the patient can be detected. The present invention addresses this problem by determining the viewing direction onto anatomy (e.g., using the invention by Blau filed as patent application on 23.8.2018) together with the location of the nail. Hence, it is possible to determine at which angle the C-arm views the nail, and at which angle it views the anatomy. Based on two subsequent X-ray images taken in AP direction, it is possible to detect whether the nail was rotated around its axis, even if there was a (slight) C-arm rotation between the images. Such detection is important because it may happen in the surgical procedure that when adjusting the entry depth of the nail, it may also inadvertently be rotated.

Example 2: An existing system requires that when attempting a 3D registration there be no movement of the implant relative to anatomy when switching between AP and ML views. Ignoring this requirement may result in an incorrect 3D registration. This example refers to the example for a potential processing workflow. In this processing workflow, the 3D angle between two viewing directions is computed based on two approaches (cf. Steps 3, 12, and 13), one based on anatomy (using the invention by Blau filed as patent application on 23.8.2018) and one utilizing the implant assembly as a reference. If both calculations are considered correct (such determination may be made based on viewing direction, visibility of anatomy, image quality, etc.) yet differ more than a threshold, the system concludes that there has been a movement of the nail relative to the anatomy in between the acquisition of the two images, and it will thus inform the user. It may also be possible to automatically correct for a movement (cf. Example 1).

If a 3D registration procedure is successful, it may lead to higher accuracy in localization and 3D reconstruction. For instance, it may help resolve ambiguities that may exist when processing a single image only.

Exemplary embodiments related to the invention

1. A system for determining a 3D representation and localization of an object in an X- ray projection image with respect to a coordinate system, the system comprising a processing unit configured to execute a computer program product including sets of instructions causing the system

to receive an X-ray projection image,

to classify an object in the X-ray projection image, to receive a model of the classified object,

to determine a 3D representation of the classified object, and

to localize the classified object with respect to the coordinate system, by applying the model to match the classified object in the X-ray image.

2. The system of example 1, wherein the match of the model with the classified object takes into account image characteristics of the X-ray projection image, wherein the image characteristics depend on at least one imaging parameter and include at least one out of the group consisting of a pillow effect, a curvature, noise, distortion, and the X-ray imaging generation method.

3. The system of any one of examples 1 or 2, where the match of the model with the classified object takes into account the effects of at least one out of the group of X-ray attenuation, absorption, and deflection.

4. The system of any one of example 1 to 3, wherein at least part of the outline of the classified object is detected in the X-ray projection image.

5. The system of any one of example 1 to 4, wherein the system is further caused to localize another object, wherein that object may even be not visible or only partially visible in the X-ray image.

6. The system of any one of example 1 to 5, wherein the computer program further includes sets of instructions causing the system to transfer geometrical aspects from the model to the object.

7. The system of example 6, wherein the computer program further includes sets of instructions causing the system to display the X-ray projection image together with those geometrical aspects.

8. The system of example 6, wherein the geometrical aspects include sizes, wherein the sizes define a scale suitable for measuring dimensions in the X-ray projection image. 9. The system of any one of examples 2 to 8, wherein at least one of the imaging characteristics is determined by using an object of known shape for calibration, by automatic detection of imaging direction, based on information from a database.

10. The system of any one of examples 4 to 9, wherein at least one of the steps of classifying an object in the X-ray projection image, detecting the outline of the classified object in the X-ray projection image, determining a 3D representation of the classified object, localization of the classified object, determining image characteristics, determining object characteristics, determining X-ray attenuation, absorption, deflection, is assisted by user input.

11. The system of any one of examples 4 to 10, wherein at least one of the steps of classifying an object in the X-ray projection image, detecting the outline of the classified object in the X-ray projection image, determining a 3D representation of the classified object, localization of the classified object, determining image characteristics, determining object characteristics, determining X-ray attenuation, absorption, deflection, is performed automatically.

12. The system of example 11, wherein at least one of the steps of classifying an object in the X-ray projection image, detecting the outline of the classified object in the X-ray projection image, determining a 3D representation of the classified object, determining image characteristics, localization of the classified object, determining object characteristics, determining X-ray attenuation, absorption, deflection, is performed by a neural net.

13. The system of any one of examples 1 to 12, wherein the matching of the object in the X-ray projection image with the model includes an adaptation of image characteristics of the X-ray projection image to image characteristics of a virtual projection of the model and/or an adaptation of image characteristics of the virtual projection of the model to image

characteristics of the X-ray projection image.

14. The system of any one of examples 1 to 13, wherein the matching of the object in the X-ray projection image takes into account object characteristics, which include at least one out of the group of wobbling of mechanical interfaces, material of implants and tools, bone fractures. 15. The system of any one of examples 1 to 14, wherein the model is a statistical shape or appearance model. 16. The system of any one of examples 1 to 15, wherein the computer program further includes sets of instructions causing the system to receive at least one further X-ray projection image of the same object, wherein the imaging parameters and/or imaging direction of the X-ray projection images differ from each other, and to determine three- dimensional aspects of the object by combining information from the plurality of projection images.

17. The system of any one of examples 1 to 16, wherein the computer program further includes sets of instructions causing the system to receive at least one further X-ray projection image depicting the same object from a different imaging direction, wherein the 3D representation of the object is used to register the images.

18. The system of example 17, wherein the registration of the images is used to determine a 3D representation and localization of at least one further object depicted at least partially in all of those X-ray images.