COMPUTER-IMPLEMENTED METHOD FOR DETERMINING BONE INFORMATION DESCRIBING THE SKELETON MATURE OF A PATIENT, X-RAYS, DETECTION DEVICE, COMPUTER PROGRAM AND ELECTRONICALLY READABLE DATA CARRIER

Description

[0001] Computer-implemented method for determining bone information describing a patient's skeletal maturity, X-ray equipment, diagnostic equipment, computer program and electronically readable data carrier

[0002] The invention relates to a computer-implemented method for determining bone information describing the skeletal maturity of a patient, an X-ray device, a detection device, a computer program and an electronically readable data carrier.

[0003] Scoliosis is a condition in which the spine does not follow its natural, straight line in the frontal plane. In other words, it is a lateral deviation of the spine from its longitudinal axis, which can involve rotation of the vertebrae and / or torsion of the vertebral bodies. Idiopathic adolescent scoliosis (AIS) is the most common form of scoliosis, affecting approximately 1–4% of adolescents. In AIS, as well as in other bone diseases, physicians must determine the most appropriate treatment for each individual patient. It is crucial to avoid both under-treatment, such as delayed surgical intervention, and overtreatment, such as unnecessary surgery.In AIS, experts assess the risk of worsening scoliosis, particularly the risk that, without treatment, a currently non-critical scoliotic deformity will progress to a critical scoliotic deformity; see, for example, the article by Sabrina Donzelli et al. "Predicting scoliosis progression: a challenge for researchers and clinicians", EClinicalMedicine 18:100244, 2020.

[0004] To assess the risk of scoliosis progression, imaging techniques, as well as biomarkers not obtainable through imaging, can be used. Common biomarkers obtainable through imaging include the Cobb angle and skeletal maturity. The Cobb angle describes the degree of curvature based on the angle of lines drawn through the vertebrae that are most twisted above and below the apex. Various approaches exist for describing skeletal maturity (also known as bone maturity). Skeletal maturity refers to the degree of maturation of a child's bones. For example, an X-ray of the left hand, fingers, and wrist can be taken to assess skeletal maturity. The visible bones can then be compared with a standard atlas, such as Greulich and Pyle.Another approach regarding the left hand involves the Sanders maturity scale (SMS). Further indications can also be obtained through X-ray imaging of the ulna or the proximal humerus. A frequently used approach, particularly when assessing scoliosis, is the determination of Risser's sign, which describes ossification at the iliac crest (specifically the ilium), since the pelvis is often included in imaging studies of the spine. In this regard, reference should also be made to the article by Maximilian Lenz et al., "Scoliosis and Prognosis - a systematic review regarding patient-specific and radiological predictive factors for curve progression," European Spine Journal, doi:10.1007 / s00586-021-06817-0, March 26, 2021.

[0005] The Risser sign is frequently used to assess skeletal maturity for evaluating spinal growth potential and the risk of scoliosis worsening. Additional X-rays are usually not required, as the pelvic region is typically included in a standard spinal X-ray examination for scoliosis evaluation. A score from 0 to 5 is assigned based on the ossification of the iliac crest. However, studies, such as the article by Jae Hyuk Yang et al., "Evaluation of accuracy of plain radiography in determining the Risser stage and identification of common sources of errors," Journal of Orthopaedic Surgery and Research 2014, 9:101, have shown that estimating the Risser sign using plain radiography is not always accurate. Variations in ossification of the ilium apophysis and misinterpretation of apophyseal fusion are the main sources of error.

[0006] Such uncertainties can also arise when considering other signs and skeletal maturity measurements. As already mentioned, various approaches are presented in the cited article by Maximilian Lenz et al.

[0007] Skeletal maturity measurements, particularly the Risser sign, are currently mostly determined visually based on simple radiographs. To improve the determination of the Risser sign, Houda Kaddioui et al., "Convolutional Neural Networks for Automatic Risser Stage Assessment," Radiology: Artificial Intelligence, 2(3), 2020, proposed evaluating standard radiographs using artificial intelligence. However, this does not always lead to reliable and accurate results. Martin Thaler et al., "Radiographic versus ultrasound evaluation of the Risser Grade in adolescent idiopathic scoliosis: a prospective study of 46 patients," European Spine Journal, 17(9): 1251-1255, suggested using ultrasound imaging instead of X-ray imaging, although this requires additional examination time.

[0008] From the publication by VAN HOUWELINGEN, Ilva. Automated Bone Age Assessment based on DXA scans for a variety of ethnicities using Deep Transfer Learning. 2021., an automated method for determining bone age based on DXA data using a trained evaluation function is known.

[0009] The invention is based on the objective of providing a method for determining bone information, in particular a measure of skeletal maturity, that is as robust, accurate and cost-effective as possible.

[0010] This problem is solved according to the invention by a computer-implemented method, an X-ray device, a detection device, a computer program, and an electronically readable data carrier according to the independent claims. Advantageous embodiments are described in the dependent claims.

[0011] A computer-implemented method according to the invention for determining bone information describing the skeletal maturity of a patient comprises the following steps: Provision of input data, comprehensive spectrally resolved X-ray data of an area of ​​interest in the patient to be used for assessing skeletal maturity, with at least two X-ray image datasets related to different X-ray spectra, application of a trained evaluation function to the input data to obtain the bone information as output data, and output of the bone information.

[0012] Spectral X-ray imaging (often also called multi-energy X-ray imaging) is fundamentally known in the prior art. This technique does not produce just a single X-ray image data set, but rather generates several different X-ray image data sets, each corresponding to a different X-ray spectra, particularly different average X-ray energies. This can be achieved by successively driving an X-ray tube to emit different radiation spectra, for example, by using different tube voltages, and sequentially acquiring the X-ray image data sets.However, within the scope of the present invention, it is preferred, as will be discussed in more detail below, to use detector-based spectral X-ray imaging, since the X-ray detector is spectrally selective and can therefore distinguish between different X-ray energies and generate X-ray image data sets assigned to different X-ray spectra in a single, common acquisition process, thus avoiding motion artifacts and reducing acquisition time. Within the scope of the present invention, dual-energy X-ray imaging is preferably used, so that two X-ray image data sets, each showing different X-ray spectra, are acquired, in particular a high-energy X-ray image data set and a low-energy X-ray image data set.

[0013] In contrast to current best practices, this approach proposes automatically determining bone information using spectrally resolved X-ray data, employing artificial intelligence in the form of a trained evaluation function. The underlying idea is that this not only provides an excellent description of bone morphology, which can be derived, for example, from high-energy X-ray image datasets, but also includes information on bone density and bone mineral content (BMC). This, in turn, is based on the understanding that skeletal maturation is influenced by the bone growth process, which itself depends on the local bone mineral content. Bone mineral content is the mineral component of bone in the form of hydroxyapatite, the assessment of which is possible using multi-energy imaging, particularly dual-energy X-ray imaging.

[0014] Therefore, a particularly preferred embodiment of the present invention provides that a combined image dataset highlighting bone minerals, derived from the X-ray image datasets, is also used as input data, in particular by weighted linear combination of the X-ray image datasets. For example, with regard to osteoporosis assessment, it has already been proposed in the prior art to determine a relative measure of bone mineral content by combining X-ray image datasets assigned to different X-ray spectra. In this context, dual-energy X-ray absorptiometry (DXA) has been proposed in particular. In this context, reference is made in particular to the IAEA publication "Dual Energy X-ray Absorptiometry for Bone Mineral Density and Body Composition Assessment", 2010, available at https: / / www-pub.iaea.org / MTCD / Publications / PDF / Pub1479_web.pdf.

[0015] It can therefore be said that the different multi-energy X-ray image datasets contain complementary information, so that at least two relevant characteristics for skeletal maturity can be derived from the same patient scan. The use of artificial intelligence in the form of a trained evaluation function makes it possible in this context to utilize all relevant features for determining bone information, thus also taking into account aspects that would not be visually apparent to the observer from the X-ray image datasets. In this way, a high-quality, robust, accurate, and reliable determination of bone information describing skeletal maturity is enabled.

[0016] As already mentioned, the invention prefers the use of detector-based spectral X-ray imaging. A preferred embodiment of the present invention therefore provides that the X-ray data are acquired using a spectrally selective X-ray detector, in particular a multilayer detector having several layers measuring different energy ranges. Such a multilayer detector can, for example, be a dual-layer detector. A dual-layer detector is configured to decompose the spectrum of the incident X-ray radiation into a low-energy and a high-energy component. For this purpose, the dual-layer detector is constructed from two layers, wherein a detector layer facing the X-ray source measures X-ray photons of the incident X-ray radiation that have low energy. The measured signals are assigned to a first (low-energy) X-ray image data set.The first detector layer is transparent to high-energy X-rays, allowing X-ray photons with higher quantum energies to be measured in the second detector layer located below or behind it, i.e., facing away from the X-ray source. These measured signals are then assigned to a second (high-energy) X-ray image dataset. For example, both detector layers can include a scintillator, making the two-layer detector an indirect-converting detector. In particular, crystals such as cesium iodide, cadmium tungstate, or ceramic materials, such as gadolinium oxysulfide, can be used as detector materials.

[0017] Another type of spectrally selective X-ray detector that can be used within the scope of the present invention is a photon-counting X-ray detector, which, for example, does not use a scintillator but can use a semiconductor that directly detects X-ray photons. Therefore, such detectors are also called direct-converting detectors. X-ray photons incident on the detector surface can be assigned to specific energy bands or energy bins depending on their quantum energy.

[0018] Advantageously, at least two two-dimensional and / or at least two three-dimensional X-ray image datasets can be provided for the different spectra. Three-dimensional X-ray image datasets, which are preferably acquired by means of 3D tomosynthesis within the scope of the present invention, have the advantage that the accuracy of the determination of bone information can be further improved, since overlapping structures can be reduced compared to two-dimensional X-ray image datasets.

[0019] It is entirely possible to use both two-dimensional and three-dimensional X-ray image datasets as input data for the trained evaluation function. The two-dimensional X-ray image datasets are preferably acquired using slot scanning. This involves repeatedly taking narrow images of adjacent areas using a slot-like collimator, which can be moved perpendicular to the longitudinal direction of the slot across the detector area of ​​the X-ray detector. Thus, only a narrow section is captured at any given time.This has the advantage that less scatter radiation reaches the X-ray detector, thus reducing scatter radiation effects and increasing image quality and therefore the quality of bone information. Furthermore, it provides an acquisition geometry in which the X-ray radiation, with the X-ray tube appropriately moved, strikes the area being imaged more or less perpendicularly. Regarding three-dimensional X-ray image datasets, it is proposed to acquire them using tomosynthesis, a long-established technique that employs a reduced number of projection directions, thus advantageously minimizing the X-ray exposure for the patient.

[0020] The invention provides that the X-ray data are acquired at least partially as part of, or within the scope of, an examination procedure covering a larger imaging area than the area of ​​interest. In particular, the imaging area can encompass the entire spine, for example, for the examination of scoliosis, while the area of ​​interest is the pelvic region. This has the advantage that no additional examination procedure is required for the X-ray data. In particular, X-ray images acquired for a different purpose, such as the diagnosis of scoliosis, can also be reused as X-ray data. This means, in particular, that X-ray images acquired during the examination procedure covering the larger imaging area can also be used as X-ray data. The multiple use of this X-ray data can therefore also reduce the radiation dose for the patient.Furthermore, there are advantages regarding the overall examination time. For example, if the imaging area includes the spine, the X-ray images from the entire examination can be used to assess the presence and severity of scoliosis, for instance, by determining the Cobb angle. Simultaneously, the pelvic region, which is also typically imaged in this context, can be used to automatically derive bone information regarding skeletal maturity using the trained evaluation function and, for example, to consider this information for a risk assessment regarding the progression of scoliosis. Although the main example discussed here may relate to the spine and pelvic region, particularly with regard to the Risser sign as a measure of skeletal maturity, the concept described here can be applied more generally.For example, the additional energy information regarding bone morphology and bone density / bone mineral content can also be used for areas of interest or recording areas in the humerus, radius, ulna and / or hand.

[0021] Specifically, it is provided that at least one two-dimensionally recorded X-ray image of the examination procedure and / or received user information for localizing the area of ​​interest are evaluated, wherein the evaluation result is used for subsequent or interruptive acquisition of three-dimensional X-ray data relating only to the area of ​​interest and / or for selecting two-dimensional X-ray data from the X-ray images to be used as input data. According to the invention, it is preferred to determine in the X-ray images during or after the examination procedure when the area of ​​interest has been reached or where it is located in relation to the recording area.However, it is also conceivable to perform the localization at least partially based on received user information. For example, at least one X-ray image could be displayed to the user, allowing them to mark the location of the area of ​​interest or confirm an image-based marker. The localization information obtained in this way can then be used, for instance, to select portions of the X-ray images to be used as two-dimensional X-ray data and pass them on to the trained evaluation function. Alternatively, it is also conceivable to interrupt the acquisition of at least one two-dimensional X-ray image and, particularly through tomosynthesis, to acquire three-dimensional X-ray data of the area of ​​interest once it has been reached. This data can then be used as input.Of course, it is also possible to record the three-dimensional X-ray data only after the examination process has been completed with regard to the recording area, then preferably as tomosynthesis.

[0022] In connection with the interruption, a particularly advantageous embodiment of the present invention provides that a two-dimensional synthetic image of the area of ​​interest is derived from the three-dimensional X-ray data and fused with the at least one X-ray image. No two-dimensional X-ray images need to be acquired for this area. In other words, a synthetic radiographic X-ray image—the synthetic image—can be derived from the three-dimensional X-ray data and fused with the at least one acquired two-dimensional X-ray image, preferably acquired by a slit scan. In this way, X-ray dose can be further reduced, since a synthetic image can be calculated and used for the area of ​​interest.

[0023] The following section describes three specific methods for acquiring X-ray data. In a first method, particularly using a spectrally selective X-ray detector, at least one spectral X-ray image of the imaging area can be acquired. This is preferably done by slit scanning. For example, for an imaging area relating to the spine, three two-dimensional X-ray images can be acquired by three consecutive slit scans and fused into a single image. Preferably automatically, but also manually or based on a prior setting, the location of the area of ​​interest, for example, the pelvic region, can be determined. From the combined image, the two-dimensional X-ray data, as two two-dimensional X-ray image datasets, can be selected that show the area of ​​interest and used as input data for the trained evaluation function.Preferably, three X-ray image datasets can be used, namely a high-energy X-ray image dataset, a low-energy X-ray image dataset and a combination image dataset, which relates in particular to the bone mineral content.

[0024] In a second variant, a two-dimensional scan of the entire imaging area is first performed, particularly by slit scanning. Following this examination, the at least one X-ray image, especially the previously described overall image, is analyzed to locate the area of ​​interest. For this area of ​​interest, a tomosynthesis acquisition of three-dimensional X-ray data can then be performed—preferably with the same X-ray system—preferably using the spectrally selective X-ray detector. Between the acquisition of the two-dimensional X-ray images and the three-dimensional X-ray data, the patient can remain in their position, and the acquisition status can optionally be displayed to the patient. For the acquisition of the three-dimensional X-ray data by tomosynthesis, the collimation is then preferably adjusted with respect to the localized area of ​​interest.For example, after confirmation by the user, the tomosynthesis scan can then be performed.

[0025] It should be noted that, depending on the area of ​​interest and the clinical question, the area of ​​interest may be located off-center with respect to the preceding two-dimensional scan. In some configurations, it is also conceivable to provide an additional camera image with skeletal markers, for example, if at least partial manual marking of the area of ​​interest is required. Alternatively, the area of ​​interest for the acquisition of the three-dimensional X-ray data could be pre-selected by the user, for example, using camera data.

[0026] Overall, in this second variant, it is conceivable to use only the three-dimensional X-ray data as input data for the trained evaluation algorithm, but it is preferred to use both the two-dimensional X-ray data and the three-dimensional X-ray data of the area of ​​interest as input data.

[0027] In a particularly preferred third embodiment, at least one two-dimensional X-ray image of the imaging area is acquired in a first step, again preferably by slit scanning, whereby, as described above, a complete image can be obtained. However, during the acquisition of the at least one two-dimensional X-ray image of the imaging area, image-based monitoring is preferably performed to determine whether the area of ​​interest, for example, the pelvic region, is reached within the imaging area. If this is the case, the acquisition of the at least one X-ray image is aborted, and the acquisition mode is adapted to acquire three-dimensional X-ray data of the area of ​​interest, particularly with the spectrally selective X-ray detector. Here, a tomosynthesis scan is again preferably used to acquire the three-dimensional X-ray data.In other words, in this variant, the two-dimensional scan is automatically paused as soon as the area of ​​interest is automatically detected. Optionally, the area of ​​interest can be displayed to a user, for example, for confirmation or adjustment. The collimation is adjusted with respect to the area of ​​interest, and the three-dimensional X-ray data is acquired via a tomosynthesis scan. Particularly preferably, a synthetic image of the area of ​​interest can be generated, especially as synthetic radiography. This synthetic image is then fused with the at least one two-dimensional X-ray image, acquired particularly by slit scanning. If a portion of the scan area is still missing, the 2D scan, i.e., the acquisition of the at least one X-ray image, can then be continued.Generally speaking, it is possible to use only three-dimensional X-ray data as input. However, two-dimensional X-ray data, particularly of the synthetic image, are preferably also used as input for the trained evaluation algorithm. Of course, other variations are conceivable within the scope of the invention.

[0028] In a particularly advantageous embodiment of the present invention, it can be provided that, when bone information relates to a specific anatomical feature, in particular the iliac crest, the area of ​​interest is selected to include at least one additional, adjacent anatomical feature providing supplementary information, in particular the proximal femoral epiphyses. Studies on the use of artificial intelligence, as well as those conducted by experienced radiologists, have shown that other anatomical features can also be expediently considered when determining bone information. For example, the Risser sign is defined with respect to the iliac crest, in particular the iliac apophysis, and radiologists often consider other anatomical features when assigning a Risser score, i.e., a category of the Risser sign, such as the proximal femoral epiphyses.The trained evaluation function can also mimic this cognitive behavior of the radiologist if given the opportunity to do so, in this advantageous embodiment of the present invention by using a specifically selected area of ​​interest that is larger than just the anatomical feature underlying the definition and also contains additional anatomical features that can provide supplementary information. In training evaluation functions, it has already been shown that the use of such supplementary information does not necessarily only relate to anatomical features consciously considered by the manual examiner, but also to other relationships that may escape human observation and are likewise related to the evaluation of the anatomical feature relevant to the bone information.Therefore, such an extension of the area of ​​interest to adjacent, additional anatomical features provides a further improvement in the accuracy and reliability of determining bone information.

[0029] The present invention utilizes artificial intelligence within the framework of a trained evaluation function. Generally, a trained function models cognitive functions that humans associate with other human brains. Through training based on training data (machine learning), the trained function is able to adapt to new circumstances and detect and extrapolate patterns.

[0030] Generally speaking, the parameters of a trained function can be adjusted through training. Specifically, supervised learning, semi-supervised learning, unsupervised learning, reinforcement learning, and / or active learning can be used. Furthermore, representational learning (also known as feature learning) can be employed. The parameters of the trained function can be adjusted iteratively through multiple training steps.

[0031] A trained function can, for example, comprise a neural network, a support vector machine (SVM), a decision tree, and / or a Bayesian network, and / or the trained function can be based on k-means clustering, q-learning, genetic algorithms, and / or assignment rules. Specifically, a neural network can be a deep neural network, a convolutional neural network (CNN), or a deep CNN. Furthermore, the neural network can be an adversarial network, a deep adversarial network, and / or a generative adversarial network (GAN).

[0032] In the present invention, a trained evaluation function comprising a Convolutional Neural Network (CNN) is preferably used. In a specific embodiment, for example, a DenseNet can be used. A DenseNet is a CNN that uses dense connections between layers, in which, in particular, all layers with suitable input data dimensions are directly connected to each other.

[0033] Specifically, bone information can be used to determine a measure of skeletal maturity, in particular a Risser sign or a category of the Risser sign, often also referred to as "Risser grades." In this case, an automated image analysis based on machine learning is proposed to classify skeletal maturity according to the Risser sign classification scheme. Although the Risser sign is mentioned as the primary example, the approach is, as stated, also transferable to other measures of skeletal maturity, such as bone morphology, bone density, and bone mineral density (SMS), since the utilization of bone morphology, bone material content, and bone density information in general is highly beneficial in these cases as well.

[0034] In one embodiment of the invention, it may also be provided that the method further comprises: Providing training datasets, each comprising spectral X-ray data as training input data with an associated training bone information, training an evaluation function using the training datasets and providing the trained evaluation function.

[0035] A standalone training method is of course also conceivable within the scope of the present invention, but this is not claimed here. The evaluation function, in particular comprising a CNN, can therefore be trained by using spectral X-ray data, which has been classified by human experts with regard to bone information, especially skeletal maturity, as training data. For example, human experts can assign Risser character categories, usually from 0 to 5, to the training input data in order to identify and use training datasets. During inference, a corresponding Risser character category from 0 to 5 is then assigned to unseen spectral X-ray data.It should be noted that, of course, within the scope of the present invention, it is also conceivable to use intermediate values ​​as the output of the trained evaluation function, thus allowing not only the usual categories 0, 1, 2, 3, 4 and 5, which are defined as no bone formation of the iliac apophysis, < 25% bone formation, 25-50% bone formation, 50-75% bone formation, 75-100% bone formation and complete bone formation of the apophysis, but in particular to allow continuous values ​​for the Risser sign, for example 3, 4.

[0036] The described method according to the invention is preferably implemented by the control unit of an X-ray device, particularly with regard to the specific possibilities for acquiring the X-ray data. Therefore, the present invention also relates to an X-ray device comprising an X-ray source and, in particular, a spectrally selective X-ray detector, as well as a control unit designed for carrying out a method according to the invention. All aspects relating to the method according to the invention can be applied analogously to the X-ray device according to the invention, with which the aforementioned advantages can therefore also be obtained.

[0037] The spectrally selective X-ray detector is advantageously a two-layer detector or a photon-counting detector. The X-ray system can be a two-arm robotic unit, with both the X-ray source and the X-ray detector being mounted on one of the arms to allow for different imaging geometries, for example, to enable tomosynthesis of the area of ​​interest. The X-ray system also preferably includes a collimator, for example, to perform slit scanning. The imaging process is controlled by the control unit. This unit is also configured to use the trained evaluation function to directly provide bone information along with the acquired X-ray data.The control unit can, for example, store examination programs tailored to different applications, allowing the specific application of the method according to the invention to various clinical questions. For example, with regard to an examination of scoliosis, the examination program can be designed such that a 2D scan is performed in an area of ​​interest relating to the spine, preferably with the addition of a 3D scan, in particular a tomosynthesis, in an area of ​​interest relating to the pelvic region, in order to determine three-dimensional X-ray data as input data for the evaluation function, which can then, for example, output a Risser sign category as a measure of skeletal maturity. An implementation according to the third variant described above is particularly preferred.

[0038] In order to carry out the method according to the invention, the X-ray device can in particular comprise a more generally defined detection device according to the invention for determining bone information describing the skeletal maturity of a patient. This device has: A first interface for providing input data, comprising spectrally resolved X-ray data of an area of ​​interest to be used for assessing skeletal maturity of the patient with at least two X-ray image datasets related to different X-ray spectra, an evaluation unit for applying a trained evaluation function to the input data to obtain the bone information of the output data, and a second interface for outputting the bone information.

[0039] The X-ray data are acquired at least partially as part of or within the framework of an examination procedure covering a larger recording area than the area of ​​interest, and at least one two-dimensionally acquired X-ray image of the examination procedure and / or received user information for localizing the area of ​​interest is evaluated, whereby the evaluation result is used for the subsequent or interrupting acquisition of three-dimensional X-ray data relating only to the area of ​​interest and / or for the selection of input data from the X-ray images.

[0040] All descriptions regarding the X-ray equipment and the method can be applied analogously to the investigation equipment. With regard to the control unit of the X-ray equipment, this can additionally include a recording unit for controlling the recording operation, which specifically controls the recording operation and, if the specific implementation of the method according to the invention includes steps relating to the recording process, can also be part of the investigation equipment.

[0041] A computer program according to the invention can be directly loaded into a storage medium of an investigative device or a control unit of an X-ray device according to the invention and comprises program means for carrying out the steps of a method according to the invention when the computer program is executed on the computing device, in particular the investigative device or the control unit. The control unit or the investigative device may, for this purpose, include a processor. The computer program can be stored on an electronically readable data carrier according to the present invention, which therefore includes control information stored thereon, comprising at least one computer program according to the invention and, when the data carrier is used in a computing device, enabling it to carry out the steps of a method according to the invention.The data carrier can be, in particular, a non-transient data carrier, such as a CD-ROM.

[0042] Further advantages and details of the present invention will become apparent from the exemplary embodiments described below and from the drawing. The drawings show: Fig. 1 an embodiment of an artificial neural network, Fig. 2 an embodiment of a convolutional neural network, Fig. 3 a flowchart of a first embodiment for providing input data, Fig. 4 a flowchart of a second embodiment for providing input data, Fig. 5 a flowchart of a third embodiment for providing input data, Fig. 6 a flowchart for determining bone information, Fig. 7 a view of the pelvic region of a human, Fig. 8 a sketch to explain categories of the Risser symbol, Fig. 9 a schematic diagram of an X-ray device according to the invention, and Fig. 10 the functional structure of a control unit of the X-ray device.

[0043] Fig. 1 Figure 1 shows an embodiment of an artificial neural network 1. English terms for the artificial neural network 1 are "artificial neural network", "neural network", "artificial neural net" or "neural net".

[0044] Artificial neural network 1 comprises nodes 6 to 18 and edges 19 to 21, where each edge 19 to 21 is a directed connection from a first node 6 to 18 to a second node 6 to 18. Generally, the first node 6 to 18 and the second node 6 to 18 are distinct nodes 6 to 18; however, it is also conceivable that the first node 6 to 18 and the second node 6 to 18 are identical. For example, in Fig. 1 Edge 19 is a directed connection from node 6 to node 9, and edge 21 is a directed connection from node 16 to node 18. An edge 19 to 21 from a first node 6 to 18 to a second node 6 to 18 is called an incoming edge for the second node 6 to 18 and an outgoing edge for the first node 6 to 18.

[0045] In this embodiment, the nodes 6 to 18 of the artificial neural network 1 can be arranged in layers 2 to 5, wherein the layers can have an intrinsic order introduced by the edges 19 to 21 between the nodes 6 to 18. In particular, edges 19 to 21 can only be provided between adjacent layers of nodes 6 to 18. In the illustrated embodiment, there is an input layer 2 that contains only nodes 6, 7, and 8, each without an incoming edge. The output layer 5 comprises only nodes 17 and 18, each without outgoing edges, with hidden layers 3 and 4 further situated between the input layer 2 and the output layer 5. In the general case, the number of hidden layers 3 and 4 can be chosen arbitrarily.The number of nodes 6, 7, 8 in input layer 2 usually corresponds to the number of input values ​​into neural network 1, and the number of nodes 17, 18 in output layer 5 usually corresponds to the number of output values ​​of neural network 1.

[0046] In particular, a (real) number can be assigned to nodes 6 to 18 of neural network 1. Here, x(n) < i denotes the value of the i-th node (6 to 18) of the n-th layer (2 to 5). The values ​​of nodes 6, 7, 8 of input layer 2 are equivalent to the input values ​​of neural network 1, while the values ​​of nodes 17, 18 of output layer 5 are equivalent to the output values ​​of neural network 1. Furthermore, each edge 19, 20, 21 can be assigned a weight in the form of a real number. In particular, the weight is a real number in the interval [-1, 1] or in the interval [0, 1,]. Here, w (m,n)< i,j denotes the weight of the edge between the i-th nodes 6 to 18 of the m-th layer 2 to 5 and the j-th nodes 6 to 18 of the n-th layer 2 to 5. Furthermore, the abbreviation w i , j n for the weight w i , j n , n + 1 defined.

[0047] To calculate the output values ​​of neural network 1, the input values ​​are propagated through neural network 1. In particular, the values ​​of nodes 6 to 18 of the (n+1)th layer 2 to 5 can be calculated based on the values ​​of nodes 6 to 18 of the nth layer 2 to 5 by x j n + 1 = f ∑ i x i n ⋅ w i , j n .

[0048] Here, f is a transfer function, which can also be called an activation function. Well-known transfer functions include step functions, sigmoid functions (for example, the logistic function, the generalized logistic function, the hyperbolic tangent, the arctangent, the error function, the smoothstep function), and rectifier functions. The transfer function is primarily used for normalization purposes.

[0049] Specifically, the values ​​are propagated layer by layer through neural network 1, with values ​​of input layer 2 being given by the input data of neural network 1. Values ​​of the first hidden layer 3 can be calculated based on the values ​​of input layer 2 of neural network 1, values ​​of the second hidden layer 4 can be calculated based on the values ​​in the first hidden layer 3, and so on.

[0050] To determine the values w i , j n To determine the values ​​for edges 19 to 21, neural network 1 must be trained using training data. Specifically, training data includes training input data and training output data, referred to below as ti. For a training step, neural network 1 is applied to the training input data to determine computed output data. Specifically, the training output data and the computed output data comprise a number of values, where the number is determined by the number of nodes 17 and 18 in output layer 5.

[0051] In particular, a comparison between the calculated output data and the training output data is used to recursively adjust the weights within neural network 1 (backpropagation algorithm). Specifically, the weights can be adjusted accordingly. w ′ i , j n = w i , j n − γ ⋅ δ j n ⋅ x i n to be changed, where γ is a learning rate and the numbers δ j n can be calculated recursively as δ j n = ∑ k δ k n + 1 ⋅ w j , k n + 1 ⋅ f ′ ∑ i x i n ⋅ w i , j n based on δ j n + 1 , if the (n+1)th layer is not the starting layer 5, and δ j n = x k n + 1 − t j n − 1 ⋅ f ′ ∑ i x i n ⋅ w i , j n if the (n+1)th layer is the output layer 5, where f' is the first derivative of the activation function and y j n + 1 the comparison training value for the j-th node 17, 18 of the initial layer 5 is.

[0052] The following will be discussed with regard to Fig. 2 An example of a Convolutional Neural Network (CNN) is also given. It's important to note that the term "layer" is used slightly differently there than in classical neural networks. For a classical neural network, the term "layer" refers only to the set of nodes that form a layer, i.e., a specific generation of nodes. For a Convolutional Neural Network, the term "layer" is often used to describe an object that actively modifies data; in other words, a set of nodes of the same generation and either the set of incoming or outgoing edges.

[0053] Fig. 2 Figure 1 shows an embodiment of a convolutional neural network 22. In the illustrated embodiment, the convolutional neural network 22 comprises an input layer 23, a convolutional layer 24, a pooling layer 25, a fully connected layer 26, and an output layer 27. In alternative embodiments, the convolutional neural network 22 can contain multiple convolutional layers 24, multiple pooling layers 25, and multiple fully connected layers 26, just like other types of layers. The order of the layers can be chosen arbitrarily, with fully connected layers 26 typically forming the last layers before the output layer 27.

[0054] In particular, within a Convolutional Neural Network 22, the nodes 28 to 32 of one of the layers 23 to 27 can be understood as arranged in a d-dimensional matrix or as a d-dimensional image. Specifically, in the two-dimensional case, the value of a node 28 to 32 with indices i, j in the nth layer 23 to 27 can be denoted as x(n) < [i,j]. It should be noted that the arrangement of the nodes 28 to 31 of a layer 23 to 27 has no effect whatsoever on the computations within the Convolutional Neural Network 22 as such, since these effects are determined solely by the structure and the weights of the edges.

[0055] A convolution layer 24 is particularly distinguished by the fact that the structure and weights of the incoming edges form a convolution operation based on a specific number of kernels. In particular, the structure and weights of the incoming edges can be chosen such that the values x k n node 29 of the folding layer 24 as a folding x k n = K k * x n − 1 based on the values ​​x (n-1)< of the nodes 28 of the preceding layer 23, where the convolution * in the two-dimensional case can be defined as x k n i , j = K k * x n − 1 i , j = ∑ i ′ ∑ j ′ K k i ′ , j ′ ⋅ x n − 1 i − i ′ , j − j ′ .

[0056] In this example, the k-th kernel Kk is a d-dimensional matrix, in this embodiment a two-dimensional matrix that is typically small compared to the number of nodes 28 to 32, for example, a 3x3 matrix or a 5x5 matrix. In particular, this implies that the weights of the incoming edges are not independent but are chosen to generate the convolution equation above. In the example for a kernel forming a 3x3 matrix, there are only nine independent weights (where each entry of the kernel matrix corresponds to an independent weight), regardless of the number of nodes 28 to 32 in the corresponding layer 23 to 27. Specifically, for a convolution layer 24, the number of nodes 29 in the convolution layer 24 is equivalent to the number of nodes 28 in the preceding layer 23 multiplied by the number of convolution kernels.

[0057] If the nodes 28 of the preceding layer 23 are arranged as a d-dimensional matrix, the use of multiple kernels can be understood as adding another dimension, also called the depth dimension, so that the nodes 29 of the convolution layer 24 are arranged as a (d+1)-dimensional matrix. If the nodes 28 of the preceding layer 23 are already arranged as a (d+1)-dimensional matrix with one depth dimension, the use of multiple convolution kernels can be understood as an expansion along the depth dimension, so that the nodes 29 of the convolution layer 24 are likewise arranged as a (d+1)-dimensional matrix, the size of the (d+1)-dimensional matrix in the depth dimension being larger than in the preceding layer 23 by the factor formed by the number of kernels.

[0058] The advantage of using convolutional layers 24 is that the spatially local correlation of the input data can be exploited by creating a local connection pattern between nodes of neighboring layers, in particular by ensuring that each node has connections only to a small area of ​​the nodes of the preceding layer.

[0059] In the illustrated embodiment, the input layer 23 comprises thirty-six nodes 28 arranged as a two-dimensional 6x6 matrix. The convolution layer 24 comprises seventy-two nodes 29 arranged as two two-dimensional 6x6 matrices, each of which is the result of convolution of the values ​​of the input layer 23 with a convolution kernel. Similarly, the nodes 29 of the convolution layer 24 can be understood as arranged in a three-dimensional 6x6x2 matrix, the latter dimension being the depth dimension.

[0060] A pooling layer 25 is characterized by the fact that the structure and weights of the incoming edges, as well as the activation function of their nodes 30, define a pooling operation based on a nonlinear pooling function f. For example, in the two-dimensional case, the values ​​x(n)< of the nodes 30 of the pooling layer 25 can be determined based on the values ​​x(n+1)< of the nodes 29 of the preceding layer 24 as x n i , j = f x n − 1 id 1 , jd 2 , … , x n − 1 id 1 + d 1 − 1 , jd 2 + d 2 − 1 The number of nodes 29, 30 can be calculated by using a pooling layer 25. This is achieved by replacing a number of d1 x d2 neighboring nodes 29 in the preceding layer 24 with a single node 30, which is calculated as a function of the values ​​of the aforementioned number of neighboring nodes 29. Specifically, the pooling function f can be a maximum function, an averaging function, or the L2 norm. In particular, for a pooling layer 25, the weights of the incoming edges can be fixed and not modified by training.

[0061] The advantage of using a pooling layer 25 is that the number of nodes 29, 30 and the number of parameters are reduced. This leads to a reduction in the required computational effort within the Convolutional Neural Network 22 and thus to control overfitting.

[0062] In the illustrated embodiment, pooling layer 25 is a max-pooling layer in which four adjacent nodes are replaced by a single node whose value is the maximum of the values ​​of the four adjacent nodes. Max-pooling is applied to each d-dimensional matrix of the preceding layer; in this embodiment, max-pooling is applied to each of the two two-dimensional matrices, thus reducing the number of nodes from seventy-two to eighteen.

[0063] A fully connected layer 26 is characterized by the presence of a plurality, in particular all, edges between the nodes 30 of the preceding layer 25 and the nodes 31 of the fully connected layer 26, whereby the weight of each edge can be individually adjusted. In this embodiment, the nodes 30 of the preceding layer 25 and the fully connected layer 26 are shown both as two-dimensional matrices and as non-connected nodes (represented as a row of nodes, the number of which has been reduced for clarity). In this embodiment, the number of nodes 31 in the fully connected layer 26 is equal to the number of nodes 30 in the preceding layer 25. In alternative embodiments, the number of nodes 30 and 31 can be different.

[0064] Furthermore, in this embodiment, the values ​​of the nodes 32 of the output layer 27 are determined by applying the softmax function to the values ​​of the nodes 31 of the preceding layer 26. By applying the softmax function, the sum of the values ​​of all nodes 32 of the output layer 27 is one, and all values ​​of all nodes 32 of the output layer are real numbers between 0 and 1. 1. When the Convolutional Neural Network 22 is used to classify input data, the values ​​of the output layer 27 can in particular be interpreted as a probability that the input data falls into one of the different classes.

[0065] A Convolutional Neural Network 22 can also have a ReLU layer, where ReLU is an acronym for "rectified linear units". Specifically, the number of nodes and the structure of the nodes within a ReLU layer are equivalent to the number of nodes and the structures of the nodes in the preceding layer. The value of each node in the ReLU layer can be calculated, in particular, by applying a rectifier function to the value of the corresponding node in the preceding layer. Examples of rectifier functions are f(x) = max(0,x), the hyperbolic tangent, or the sigmoid function.

[0066] Convolutional neural networks 22 can be trained, in particular, based on the backpropagation algorithm. To avoid overfitting, regularization methods can be used, for example, dropout of individual nodes 28 to 32, stochastic pooling, use of artificial data, weight decay based on the L1 or L2 norm, or maximum norm restrictions. The present invention will now be explained with reference to several exemplary embodiments, in which, purely by way of example, the automatic determination of a category of the Risser sign as a measure of skeletal maturity is described. This means that the area of ​​interest is the pelvic region, in particular comprising the ilium as the main relevant feature defining the Risser sign, which indicates bone formation (ossification) there.This involves acquiring two- or three-dimensional X-ray data, usable as input for a trained evaluation function that determines the category of Risser's sign, as part of a scoliosis examination procedure that also aims to image the spine within this comprehensive imaging area. An X-ray system with a movable X-ray tube and a movable spectrally selective X-ray detector is used. The spectrally selective X-ray detector is exemplified as a two-layer detector.

[0067] In general, for all embodiments described below, the process begins with step S1, at least one two-dimensional X-ray image of the entire imaging area, here in a frontal image plane. Spectral X-ray images, which thus include X-ray data for different energy spectra, can already be acquired at this stage. The X-ray images are acquired using slot scanning to reduce scatter radiation effects and avoid geometric distortions as much as possible. In the specific embodiment, for example, three such X-ray images, which are then combined to form a single image, can be acquired: one of the upper back and neck region, one of the mid-torso region, and one of the pelvic region.

[0068] In the first variant shown here according to Fig. 3 In step S1, the entire image is first acquired. Then, in step S2, the area of ​​interest, i.e., the pelvic region, is automatically localized using image-based methods. In step S3, the input data is compiled by selecting the spectral X-ray data from the entire image that lie within the area of ​​interest, as determined by the localization information from step S2. This includes several two-dimensional X-ray image datasets for the different energy spectra. In step S3, a combined dataset is also created as an additional two-dimensional X-ray image dataset. This combined dataset, as known from XDA, relates to the bone mineral content in the area of ​​interest, thus highlighting bone minerals.In this case, the input data therefore includes a two-dimensional high-energy X-ray image dataset of the area of ​​interest, a two-dimensional low-energy X-ray image dataset of the area of ​​interest, and the two-dimensional combined image dataset obtained by weighted linear combination of the high-energy X-ray image dataset and the low-energy X-ray image dataset.

[0069] A second variant presents the exemplary embodiment according to Fig. 4 This involves, in which the 2D scan of the recording area was first carried out completely in step S1, so that the complete image is available. Just as in the first embodiment according to Fig. 3 In step S2, the area of ​​interest is localized, preferably image-based on the overall image. After this has been confirmed, for example by a user, the acquisition mode of the X-ray system is changed, and in step S4, a 3D scan of the area of ​​interest, in this case the pelvic region, is performed. Tomosynthesis is then carried out to acquire three-dimensional spectral X-ray data, which thus comprise a three-dimensional high-energy X-ray image dataset and a three-dimensional low-energy X-ray image dataset. It should be noted that the area of ​​interest can, of course, be located off-center relative to the acquisition area, depending on the area of ​​interest and the clinical question.

[0070] In step S3, the input data are compiled. In this second variant, a combined dataset highlighting the bone mineral content is generated from both the two-dimensional spectral X-ray data of the overall image in the area of ​​interest and the three-dimensional X-ray data of the 3D scan. The input data consists of the two-dimensional high-energy X-ray image dataset, the two-dimensional low-energy X-ray image dataset, the two-dimensional combined image dataset, the three-dimensional high-energy X-ray image dataset, the three-dimensional low-energy X-ray image dataset, and the three-dimensional combined image dataset. It is also conceivable to work solely with the three-dimensional spectral X-ray data.

[0071] In a particularly preferred third embodiment according to Fig. 5 The process begins with a 2D scan of the imaging area. However, during the scan, step S5 continuously monitors whether the area of ​​interest has been reached. This is done by evaluating the X-ray images acquired during the 2D scan. As long as the area of ​​interest has not been reached, the process continues with step S1.

[0072] However, if the area of ​​interest is reached in step S5, in this case the pelvic region, the 2D scan of the acquisition area is aborted, and a message can optionally be sent to a user who can then confirm that the area of ​​interest has been reached. The acquisition mode is then changed again, just like the collimation, to acquire a 3D scan of the area of ​​interest in step S4, as in the second variant, a tomosynthesis scan. Once the tomosynthesis scan is complete, and the three-dimensional spectral X-ray data is available, this data is used in step S6 to generate a synthetic radiography of the area of ​​interest as a two-dimensional synthetic image. This synthetic image is then fused with the two-dimensional X-ray images acquired in step S1, so that a new acquisition of the area of ​​interest does not have to be performed, thus saving X-ray dose.

[0073] In step S7, it is checked whether the entire scan area is captured in the existing 2D scan and by the synthetic image. This should be the case for a scan area involving the spine with the pelvic region as the area of ​​interest, provided the scan started from the top. If this is not the case, the 2D scan from step S1 is continued before the input data is compiled in step S3. As explained regarding the second variant, the input data preferably includes the three-dimensional spectral X-ray data of the entire image, the two-dimensional combined dataset, and the three-dimensional X-ray data from the tomosynthesis scan together with the three-dimensional combined dataset.

[0074] Fig. 6 The subsequent evaluation, common to all three variants, then begins with step S3 – the compilation and provision of the input data. It should be noted at this point, however, that although the Risser sign refers to the ilium, particularly the iliac crest, as the relevant anatomical feature for its definition, the area of ​​interest is broader and also includes other anatomical features. These additional anatomical features can provide supplementary information that is equally useful in determining the category of the Risser sign, i.e., the bone information.

[0075] This was demonstrated using the following: Fig. 7 The diagram, which schematically shows the bones of a pelvic region as the area of ​​interest (33), is explained in more detail below. The lower end of the spine (34) is visible, with intervertebral discs (35), lumbar vertebrae (36), sacrum (37), and coccyx (38). The ilium (39) is essential for Risser's sign, as Risser's sign relates to the formation of the iliac crest (40) (often simply called the iliac crest) as a bone, as shown in the diagram. Fig. 8 The different categories are explained in more detail below to better distinguish them from other reference marks. Fig. 8 marked with a preceding "K". The following categories from 0 to 5 (here K0 to K5) are typically used: Category K0: No ossification of the iliac apophysis, Category K1: Less than 25% ossification, Category K2: 25-50% ossification, Category K3: 50-75% ossification, Category K4: 75-100% ossification, Category K5: Complete ossification and fusion of the apophysis.

[0076] Returning to Fig. 7 Further anatomical features, specifically bones, are shown, specifically the pubic bone 41, the pubic symphysis 42, and the femoral heads 43, which form the joint with the thigh bone 44 (femur). The femur has proximal epiphyses 45. Furthermore, in Fig. 7 The sacroiliac joint is marked 46.

[0077] The area of ​​interest 33 evidently encompasses more than just the iliac crest 40 as the defining anatomical feature 47, but also additional, adjacent anatomical features 48, such as the proximal femoral epiphyses 45, which radiologists also consider when assessing Risser's sign. Thus, the trained evaluation algorithm, which according to Fig. 6 When applied to the input data in step S8, such features and, if applicable, further relationships in the vicinity of the defining anatomical feature 47 are taken into account.

[0078] In the present case, the trained evaluation function, which is applied to the input data in step S6, comprises a Convolutional Neural Network 22, which can preferably be implemented as a DenseNet. The trained evaluation function determines bone information from the input data provided in step S3, specifically the category of the Risser sign as a measure of skeletal maturity. DenseNets are particularly suitable for this purpose, as it involves a classification into one of the categories relating to Fig. 8 The categories shown are 0 to 5. The trained evaluation function can also be configured to determine intermediate values, for example, category 1.8 or similar. The evaluation function was previously trained using spectral X-ray data to which Risser symbol categories were assigned by experts.

[0079] After the Risser sign is classified based on spectral X-ray data, bone density, particularly in the form of bone mineral content (BMC), is considered alongside bone morphology. This is motivated by the fact that skeletal maturation is influenced by the bone growth process, which in turn is based on the local bone mineral content. Therefore, in the preferred embodiments presented here, a combination image specifically tailored to highlight bone mineral content is used in the input data. This image, as is known from DXA (dual-energy X-ray absorptiometry), can be determined by combining a high-energy X-ray image dataset with a low-energy X-ray image dataset, with appropriate weighting to emphasize, in particular, the signal components of hydroxyapatite. This occurs, as already explained, in step S3.

[0080] Naturally, these statements also apply to other bone density information, especially skeletal maturity measurements, so that the addition of information on energy dependence also provides a benefit for the accurate, robust and reliable determination of bone information.

[0081] In step S9, the bone information, in this case the classification, i.e., category, of the Risser sign is then output, for example on a monitor together with the overall image or X-ray data. Saving it to a storage medium for later use, transferring it, especially to a diagnostic workstation or an archiving system, and the like are of course also conceivable.

[0082] Fig. 9 Figure 1 schematically shows an embodiment of an X-ray device 49 according to the invention, which is configured here as a ceiling-mounted, robotic imaging system. For example, robotic arms 50, here telescopically designed, can be guided in a rail system 51 and, as parts of the imaging arrangement, carry an X-ray source 52 and a spectrally selective X-ray detector 53, here a two-layer detector. A patient 54, standing for the assessment of scoliosis, can be positioned between the X-ray source 52 and the X-ray detector 53. Due to the mobility of both the X-ray source 52 and the X-ray detector 53 in various, in particular all, degrees of freedom, both slit scanning using a corresponding collimator (not shown in detail here) and tomosynthesis are possible.

[0083] The operation of the X-ray device 49 is controlled by a control device 55, which in this case also includes an investigative device 56 according to the invention. Fig. 10 Figure 55 shows the functional structure of the control unit 55 in this embodiment in more detail. The control unit 55 initially comprises a recording unit 57, which controls the recording operation of the X-ray device 49. This means that it also controls, in particular, the recording processes as described in the Figuren 3, 4 and 5 as explained, and thus can control steps S1 to S7.

[0084] As described in step S3, the compiled input data is then provided to the investigation unit 56 via an interface 58 (in this case, an internal interface) and used by an evaluation unit 59 to apply the trained evaluation algorithm, as described in step S8. The determined bone information, as output data, is provided to the investigation unit 56 via a second interface 60. It can, for example, be displayed on a monitor 61 of the X-ray unit 49 or stored in a storage device 62 of the control unit 55, which may also be at least partially part of the investigation unit 56.

[0085] The storage medium can contain the variant used to record the X-ray data, for example, one of the variants of Figuren 3 - 5The examination program, in this example for scoliosis, is stored as a diagnostic program, whereby the examination program provides both the overall image of the scan area and the bone information. It should be noted that, of course, further automatic evaluation algorithms can also be used by the control unit 55 within the examination program, for example, to assess the current severity of the scoliosis, in particular to calculate the Cobb angle, but also to determine further information for assessing the risk of scoliosis progression, whereby a risk assessment using bone information is also conceivable.

[0086] Although the invention has been illustrated and described in detail by the preferred embodiment, the invention is not limited by the disclosed examples and other variations can be derived by the person skilled in the art without leaving the scope of protection of the invention.

Claims

1. Computer-implemented method for determining an item of bone information describing the skeletal maturity of a patient (54), comprising the steps: - providing input data, comprising spectrally resolved X-ray data of a region of interest (33) of the patient (54) to be used for assessing the skeletal maturity, with at least two X-ray image datasets relating to X-ray spectra, - applying a trained evaluation function to the input data to obtain bone information as output data, and - outputting the bone information, wherein the X-ray data is acquired at least to some extent as part of or in the context of an examination procedure relating to an acquisition region that is larger than the region of interest (33), at least an X-ray image of the examination procedure which is acquired in two dimensions and / or an item of user information received for localising the region of interest (33) is evaluated, wherein the evaluation result is used for subsequent or intermittent acquisition of three-dimensional X-ray data relating only to the region of interest (33) and / or for selecting two-dimensional X-ray data of the X-ray images to be used as input data.

2. Method according to claim 1, characterised in that a combination image dataset determined from the X-ray image datasets which highlights bone minerals is also used as input data.

3. Method according to claim 1 or 2, characterised in that the X-ray data is acquired by means of a spectrally selective X-ray detector (53).

4. Method according to one of the preceding claims, characterised in that at least two two-dimensional and / or at least two three-dimensional X-ray image datasets are provided in each case for the different spectra.

5. Method according to claim 4, characterised in that the two-dimensional X-ray image datasets are acquired by means of slit scanning and / or the three-dimensional X-ray image datasets are acquired by way of tomosynthesis.

6. Method according to one of the preceding claims, characterised in that a two-dimensional synthetic image of the region of interest (33) is determined from the three-dimensional X-ray data, which image is fused with the at least one X-ray image.

7. Method according to one of the preceding claims, characterised in that the region of interest (33), comprising an additional adjacent anatomical feature (48) that provides at least one item of additional information, is selected in the case of an item of bone information relating to a specific anatomical feature (47).

8. Method according to one of the preceding claims, characterised in that the trained evaluation function comprises a convolutional neural network, in particular a DenseNet.

9. Method according to one of the preceding claims, characterised in that a degree of skeletal maturity is determined as bone information.

10. X-ray facility (49), having an X-ray emitter (52) and an X-ray detector (53) which is in particular spectrally selective, and a control facility (55) designed to carry out a method according to one of the preceding claims.

11. Determination facility (46) for determining an item of bone information describing the skeletal maturity of a patient (54), having: - a first interface (58) for providing input data, comprising spectrally resolved X-ray data of a region of interest (33) of the patient (54) to be used for assessing the skeletal maturity with at least two X-ray image datasets relating to different X-ray spectra, - an evaluation unit (59) for applying a trained evaluation function to the input data to obtain the bone information as output data, and - a second interface (60) for outputting the bone information, wherein the X-ray data is acquired at least partially as part of or in the context of an examination procedure relating to an acquisition region that is larger than the region of interest (33), at least an X-ray image of the examination procedure that is acquired in two dimensions and / or an item of user information received for localising the region of interest (33) is evaluated, wherein the evaluation result is used for subsequent or intermittent acquisition of three-dimensional X-ray data relating only to the region of interest (33) and / or for selecting two-dimensional X-ray data of the X-ray images to be used as input data.

12. Computer program, which carries out the steps of a method according to one of claims 1 to 9 when it is executed on a computer facility.

13. Electronically readable data carrier, on which a computer program according to claim 12 is stored.