Direct inference based on undersampled MRI data from industrial samples

Abstract

The method for automatic non-invasive identification of predetermined features in a number of industrial samples (102) of a predefined sample type of the present invention comprises the steps of: a) transporting the industrial samples (102) of a predefined sample type to an MRI scanner (106); b) recording undersampled MRI data (300) for at least one slice or at least a partial volume of the industrial sample in an MRI measurement, the undersampled MRI data (300) comprising undersampled MRI raw data comprising a number of time-dependent signals for different phases and / or processed MRI data obtained from processing the undersampled MRI raw data; and c) analyzing the undersampled MRI data (300) in an inference module (200) for identifying predetermined features of the industrial samples (102) using a machine learning module (204) trained to identify the predetermined features in industrial samples (102) of the predefined sample type from the undersampled MRI data (300).

Description

[Technical field]

[0001] The present invention relates to a method for the automatic non-invasive identification of predetermined features in a large number of industrial samples of a given sample type, an inference module for analysing undersampled MRI data of the industrial samples, an MRI system including said inference module, and a corresponding computer program product. [Background technology]

[0002] In modern industries, advanced automation techniques have led to higher output and productivity, and industrial processes have changed significantly in the past decades. In recent years, various industries have significantly increased their overall production, i.e. the number of processed industrial products per hour, resulting in high throughput, i.e. high speed processing, and reduced processing time of individual industrial products. At the same time, the amount of customer expectations and / or official regulations on product quality has increased, and high product quality must be achieved in addition to high productivity. Furthermore, some industries, such as the fertilization status of eggs in hatcheries or the expected productivity of seeds in the agricultural industry, depend on the accurate evaluation of the characteristics and quality of the relevant basic materials. Therefore, allowing the acceleration of the analysis of industrial samples is highly relevant for both quality control and property prediction in various industrial sectors.

[0003] A promising approach to meet this demand and allow high-throughput analysis of samples is the use of techniques based on NMR (nuclear magnetic resonance), in particular magnetic resonance imaging (MRI). For example, the use of MRI techniques to provide information about the interior of industrial products has recently attracted attention. Several different concepts for analyzing the properties of industrial samples using imaging techniques have been described in the prior art, for example in EP 3483619 A1, US 6149956 A2, WO 02 / 059586 A1, US 10338015 A1 and US 2019 / 011383 A1. Due to the fundamental relevance of MRI for possible applications in industrial applications, much attention has been focused on optimizing sample throughput, which is heavily influenced by, among other things, the potential scanning speed and the overall MRI processing speed.

[0004] Apart from reducing the number of samples, an approach to reduce measurement time and increase processing speed is to reduce the digitization rate and / or reduce the number of experiments in the phase encoding dimension. As a result, undersampled MRI data are acquired. Furthermore, processing speed can be further increased by reducing the shot repetition time of each scan and / or reducing the number of scans in each sub-experiment. Using this approach, the total time required for each sample can be reduced, which can result in higher throughput, especially when combined with techniques such as parallel imaging.

[0005] However, acquiring undersampled MRI data in so-called k-space reduces the quality of the MRI image, especially the appearance of so-called aliasing artifacts, which, when combined with short shot repetition times and a small number of scans, can dramatically reduce the signal-to-noise ratio and overall quality of the obtainable MRI image. When using MRI as an imaging technique, the assessment of sample quality and / or specific sample characteristics in the prior art relies heavily on the analysis of the acquired MRI images, often performed by a process operator. As a result, the prior art has focused on improving the quality of MRI images obtained from undersampled MRI data using different approaches, including some concepts that rely on artificial intelligence, for example, utilizing artificial neural networks to remove artifacts from the images. Much of this development has been driven by improvements resulting from the application of MRI in the medical field and the diagnosis of medical conditions, where medical professionals rely on receiving high-quality, artifact-free MRI images. For related disclosures, see, for example, U.S. Patent Application Publication No. 2021 / 224634 (Patent Document 6), U.S. Patent No. 1,170,543 (Patent Document 7), and U.S. Patent Application Publication No. 2020 / 0305756 (Patent Document 8).

[0006] Although these techniques for improving image quality can be effective in obtaining high-quality MRI images from undersampled MRI data, the reconstruction of each aliased MRI image based on the undersampled MRI data typically requires a powerful processor and a large memory capacity, i.e., a high-performance computer. The amount of additional computing power required for the enhanced image reconstruction increases the initial resources required to perform the complex multidimensional Fourier transform required to obtain the basic MRI image from the undersampled MRI data in the first place. Depending on the available computing power, the overall image reconstruction itself can be a factor that reduces the overall throughput. [Prior art documents] [Patent documents]

[0007] [Patent Document 1] European Patent Application Publication No. 3483619 [Patent Document 2] U.S. Pat. No. 6,149,956 [Patent Document 3] WO 02 / 059586 [Patent Document 4] U.S. Patent No. 10,338,015 [Patent Document 5] US Patent Application Publication No. 2019 / 011383 [Patent Document 6] US Patent Application Publication No. 2021 / 224634 [Patent Document 7] U.S. Pat. No. 1,117,0543 [Patent Document 8] US Patent Application Publication No. 2020 / 0305756 Summary of the Invention [Problem to be solved by the invention]

[0008] The main object of the present invention was to overcome or at least reduce the deficiencies of the prior art.

[0009] In particular, it was an object of the present invention to provide a powerful method for analyzing industrial samples with high throughput allowing rapid sample processing, thus offering broad applicability in multiple industrial fields.

[0010] Here, it was an object of the present invention to make it possible for the respective method to reduce the computational power required for the analysis and to shorten the processing time.

[0011] It was hoped that each method would be applicable to a wide variety of industrial sample types, allowing efficient simultaneous analysis of several samples and / or subsequent analysis at high throughput rates.

[0012] Moreover, it was a further object of the present invention that the respective method should provide particularly reliable results, preferably with as little susceptibility to human error as possible, and most preferably with the elimination of the need for a skilled process operator.

[0013] Likewise, it was a further object of the present invention for each method to be operable on a wide variety of different MRI systems.

[0014] A second object of the invention was to provide a powerful MRI system and a corresponding inference module suitable for use in the respective methods. Likewise, a further object of the invention was to provide a computer program product enabling the respective methods.

[0015] The inventors of the present invention have surprisingly found that the above objectives can be achieved when a method for automatic non-invasive identification of a predetermined feature in a large number of industrial samples of a given sample type is used as defined in the claims. Here, the inventors have realized that the inherent uniformity and / or similarity between different industrial samples of the same type, e.g. between a large number of eggs or a large number of seeds, opens up a promising route to increase the processing speed when using MRI and machine learning techniques. In particular, the inventors have found that for industrial samples, unlike e.g. in the medical field, it is futile to provide MRI images as output and to analyze said MRI images. Instead, they have found that it may be sufficient to automatically identify the predetermined feature, i.e. its presence or absence or size, and to provide that information as output rather than processed MRI images. The inventors have developed a machine learning based approach that synergistically exploits this idea. For this, the inventors do not rely on techniques to enhance MRI image quality, but instead have found that machine learning techniques can be directly applied to undersampled MRI data, in particular undersampled MRI raw data and MRI images containing aliasing artifacts, in order to identify whether a predetermined feature is present in an industrial sample. In fact, it has been surprisingly found that aliasing artifacts arising from processing undersampled MRI raw data, which are known to sometimes result in the duplication of features (e.g., the indication of contaminants), can be exploited to synergistically increase the discriminatory power of said features by a machine learning module trained for this purpose. Thus, the inventors have confirmed that aliasing artifacts arising from processing undersampled MRI raw data can be designed to duplicate features of interest, and have found that instead of trying to mitigate aliasing artifacts, the surprising advantages of aliasing artifacts can be taken advantage of. During development, the inventors have found that this beneficial concept can be further exploited, since the machine learning module, unlike humans, does not rely on analyzing fully processed MRI images.Instead, by implementing said method, it has been surprisingly found that predetermined features can be directly identified from undersampled raw MRI data or processed MRI data that have not been fully transformed into MRI images, e.g. obtained after a 2D or 3D Fourier transformation, thereby beneficially reducing even the amount of computational resources required for general image reconstruction. In other words, the inventors have found that the high comparability of predefined types of industrial samples and the strategy of restricting the search to one or more predetermined features of interest in combination with powerful machine learning techniques and appropriate training allows direct inference based on undersampled imaging MRI data, thereby allowing a very reliable analysis of relevant important features with a very high throughput of industrial samples and a low risk of human error. Furthermore, the inventors have found that said method can be beneficially operable on a wide range of industrially relevant MRI systems if information about hardware and experimental parameters is included in the training set used to train the machine learning module. Overall, it has been found that the machine learning module does not significantly depend on a signal-to-noise ratio sufficient for reliable visual inspection, so that a very beneficial side effect of the respective method is that fewer scans and shorter shot repetition times can be employed in MRI measurements.

[0016] The above-mentioned object is achieved by the subject matter of the invention as defined in the claims. In the following, the subject matter of the invention is described in more detail and preferred embodiments of the invention are disclosed. It is particularly preferred to combine two or more preferred embodiments to obtain a particularly preferred embodiment. Correspondingly, particularly preferred are methods of the invention that define two or more features of preferred embodiments of the invention. Also preferred are embodiments in which features of one embodiment that is to some extent preferred are combined with one or more further features of another embodiment that is to some extent preferred. Features of the preferred MRI system, inference module and computer program product result from features of the preferred method. [Means for solving the problem]

[0017] The method of the present invention for automated, non-invasive identification of a predetermined feature in a large number of industrial samples of a given sample type comprises: a) transporting an industrial sample of a predefined sample type to an MRI scanner; b) recording undersampled MRI data for at least one slice or at least a partial volume of the industrial sample in an MRI measurement, said undersampled MRI data being Undersampled raw MRI data containing multiple time-dependent signals for different phases, and / or Processed MRI data obtained by processing the undersampled raw MRI data and c) analyzing the undersampled MRI data using a machine learning module trained to identify the predetermined features in industrial samples of the given sample type from the undersampled MRI data, with an inference module for identifying predetermined features of the industrial samples; The inference module includes a memory that stores the machine learning module and a processor that controls the inference module, the inference module being configured to provide the undersampled MRI data as input to the machine learning module and to analyze the undersampled MRI data using the machine learning module, the machine learning module being trained to identify the predetermined feature in industrial samples of the predetermined sample type using a training set including undersampled MRI data of different training samples of the predetermined sample type, some of the training samples including the predetermined feature and some of the training samples not including the predetermined feature. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0018] The method of the present invention is automated, i.e., operated primarily by automated equipment, allowing for high throughput and, in most cases, eliminating the need for a process operator. In other words, the method uses machines and electronics to perform most tasks, and as a result, can be performed by a machine or computer without the need for human control. In theory, the method can be only partially automated, for example, by including additional manual process steps or using a manual conveyor belt. However, mostly or completely automated operation is preferred.

[0019] The term "non-invasive" is well known to those skilled in the art and defines that the method is non-destructive and does not require incision of the sample or insertion of an instrument into the sample or extraction of contents from the sample.

[0020] The method of the present invention involves the identification of a predetermined feature in a large number of industrial samples of a given sample type.

[0021] The term "industrial sample" in this specification refers to a sample that comes from an industrial facility and / or is further processed in an industrial facility. In full agreement with the understanding of those skilled in the art, the term "industrial sample" does not include animals or humans. Therefore, the method of the present invention does not include the identification of predetermined characteristics in humans and animals. Based on the needs of each industry and the excellent compatibility of the method of the present invention, the method of the present invention is preferably specified for use in industries where the industrial sample is selected from the agricultural industry, particularly farms and hatcheries, and the food industry, particularly confectionery production or the use of seeds and nuts.

[0022] The industrial samples used in the method of the invention belong to a predefined sample type. The term "predefined sample type" in this specification relates to a sample class or sample category that is defined prior to the method of the invention and is most often industry-dependent. Thus, samples of a predefined type generally exhibit similar characteristics so that they can be compared with each other. In other words, a predefined sample type is a grouping criterion for grouping a number of industrial samples according to their type, nature, origin or properties, including shape, volume, composition and color. In particular, a number of samples of a predefined type may or may not exhibit a predefined characteristic. Correspondingly, the method of the invention is configured to identify whether a sample of a predefined type exhibits a predefined characteristic. Although this definition may at first glance seem cumbersome, there is no practical obstacle for a person skilled in the art to define a sample type and determine whether an industrial sample falls into a predefined sample type. For example, a predefined sample type may be "egg", or more specifically "chicken egg", or "praline", or more specifically "praline with cherries". In most cases, said predefined sample type may be any product produced in the respective facility. Those skilled in the art will easily understand that identifying a given feature in a more general predefined sample type, e.g. "seed" instead of "grape seed", requires more effort in training the machine learning module and typically requires a larger training set to obtain a suitable machine learning module, as disclosed below. This is not an issue for the method of the present invention, but defining a very general predefined sample type may reduce the overall value of the feature identification provided by the method, for example, because the features required for grape seeds in terms of expected yield may be completely uncorrelated with the yield of cereal seeds. Therefore, in most cases, it is preferable to define the predefined sample type as narrowly as reasonably possible and adapt the training set accordingly.

[0023] The industrial sample of the predefined sample type can in principle be any industrially usable product or item that exhibits substantially similar characteristics, i.e. that meets the preferred grouping criteria.However, the inventors have found that the method of the present invention can be used very efficiently for certain industrial samples, especially due to their chemical and structural components, and therefore their typical NMR properties, as well as their typical size, allowing efficient measurement in relatively small MRI devices that can be more easily arranged in a typical factory.In particular, the method of the present invention is preferably used in which the predefined sample type is selected from the group consisting of polymers, animal products, plants, and products derived from these raw materials, preferably plastics, organic tissues, meat, fish, eggs, fruits, seeds, and processed foods and beverages, more preferably eggs, seeds, nuts, and chocolate products.

[0024] The term "predetermined feature" herein refers to a feature that is identified, set, or established prior to the method of the present invention, and "predetermined" and "predefined" are used interchangeably herein to ascribe "feature" and "sample type," respectively, to allow a clearer distinction between the two aspects. The predetermined feature can be any feature, characteristic, or attribute of an industrial sample, such as a biomarker. Specifically, the predetermined feature can relate, for example, to the chemical composition, physical properties, and structural features of an industrial sample, but also, for example, to the size ratios of elements in the sample. Compared to most prior art techniques that visually analyze quality based on enhanced MRI images and derive conclusions from what is visible (either visually or with a conventional computer program), the method of the present invention searches for specific predetermined features in the undersampled MRI data by training a machine learning module for each of these features.

[0025] Those skilled in the art will appreciate that the method of the present invention can be used to simultaneously identify two or more predetermined features on the same sample using the same undersampled MRI data (e.g., the presence of cherry seeds and / or cherry stalks in a large number of cherries). Those skilled in the art will appreciate that the identification of more general predetermined features, such as, for example, the identification of more general contaminants in cherries (i.e., including cherry seeds, cherry stalks and insects), will require more effort in training the machine learning module and will typically require a larger training set to obtain a suitable machine learning module. Therefore, in most cases, it is preferable to define the predetermined feature as narrowly as reasonably possible. In comprehensive experiments, the inventors have identified several categories of predetermined features that can be identified very reliably with the method of the present invention. In particular, the method of the present invention preferably provides that the predetermined feature is selected from the group consisting of chemical composition, physical properties, especially magnetic properties, and structural features, preferably structural features, more preferably anatomical features, biological features, morphological dimensions, sample structure, spatial distribution and size ratios of elements in industrial samples, and the presence of impurities.

[0026] In a first step, the method of the invention comprises transporting the industrial samples to the MRI scanner. The skilled person will understand that the transport can be performed manually or, in a highly preferred embodiment, automatically. The transport can comprise a pre-sorting mechanism and / or a transport mechanism. In a preferred embodiment, the transport is controlled by a transport controller, in particular a transport controller in communication with a central controller controlling the MRI scanner. In particular, the transport can be performed for a number of industrial samples simultaneously, for example by using a suitable tray, and the method of the invention is preferably performed for all samples transported to the MRI scanner. In this respect, the method of the invention preferably comprises transporting a number of industrial samples to the MRI scanner in a sequential and / or parallel manner. Also, the method of the invention preferably comprises an arrangement in which a number of industrial samples are transported in a regular pattern, preferably in a matrix configuration, preferably on a holder or tray suitable for holding a number of industrial samples of a predefined sample type.

[0027] As mentioned above, there is a need for high throughput analysis to ensure the quality and / or characteristics of a large number of industrial samples per hour. Those skilled in the art will understand that although the method of the present invention is defined for an individual industrial sample for reasons of clarity, the method of the present invention is designed to analyze a large number of industrial samples, in particular a huge number of industrial samples per hour, rather than laboratory-scale experiments that are typically performed on only a single sample. Thus, the method of the present invention is preferably applied to 2 or more industrial samples, preferably 5 or more, more preferably 10 or more, most preferably 20 or more, particularly preferably 1000 or more, even more preferably 5000 or more industrial samples. Similarly, the term "industrial sample" in the above definition can be replaced with "one or more than two industrial samples", preferably "two or more industrial samples".

[0028] Particularly high throughput rates are typically obtained by exploiting the preferred possibilities of the method of the invention, due to its advantageously low demands on computing resources and the ability to use highly undersampled MRI data, to process several industrial samples simultaneously and / or in succession at high speed. Correspondingly, for essentially all embodiments, it is preferred to operate the method at high sample throughput rates and / or processing rates. Thus, the method of the invention is preferably applied to a large number of industrial samples in succession (meaning that a large number of samples are processed one after the other) and / or simultaneously to a large number of industrial samples, preferably simultaneously applied several times in succession to a large number of industrial samples. Correspondingly, the method of the invention preferably records undersampled MRI data for at least one slice or at least a partial volume of an industrial sample simultaneously for a large number of industrial samples, preferably 3 or more, more preferably 5 or more, most preferably 20 or more industrial samples. In particular, the method of the invention is preferably operated at a rate of 1000 or more industrial samples per hour, preferably 5000 or more industrial samples per hour. Additionally or alternatively, the method of the present invention preferably comprises a time for recording the undersampled MRI data for the industrial sample in the range of 1 to 30 seconds, preferably in the range of 1 to 10 seconds, more preferably in the range of 1 to 5 seconds.

[0029] The general structure and function of MRI scanners, as well as the concept of MRI measurements, including typical MRI experiments, pulse sequences, and techniques, are well known to those skilled in the art and need not be described herein. MRI scanners suitable for the method of the present invention are commercially available from several suppliers and in most cases include the software and pulse programs required to record MRI data or undersampled MRI data, respectively. Further disclosures on the general concept of MRI are readily available in the prior art, for example WO 2019 / 092265, but the inventors have identified features of MRI scanners and measurement schemes that have proven to be highly suitable for carrying out the method of the present invention in an industrial environment.

[0030] Regarding the hardware, the method of the invention preferably comprises that the MRI scanner comprises a magnet generating a static magnetic field, the magnetic field having a magnetic field strength preferably in the range of 0.05-9.4 Tesla, more preferably in the range of 0.05-7 Tesla, most preferably in the range of 0.05-1 Tesla, the MRI scanner typically comprises a measurement zone typically located in the center of the static magnetic field, the industrial sample being in the measurement zone during the MRI measurement. The respective magnetic field strength has proven to be a very efficient compromise between the price of the hardware, the respective maintenance costs and the safety of the workplace on the one hand, and the achievable signal-to-noise ratio on the other hand. Due to the surprisingly good performance of the method of the invention when applied to undersampled MRI data with poor signal-to-noise ratio, relatively low field strengths can be used, further contributing to the expansion of industrial applicability. In most cases, in the method of the invention, the MRI scanner comprises a static magnetic field, preferably a one-dimensional, two-dimensional or three-dimensional magnetic field gradient coil configured to induce a magnetic field gradient in the measurement zone, and / or the MRI scanner comprises one or more radio frequency coils preferably configured to apply a radio frequency magnetic field in the measurement zone.

[0031] With regard to MRI measurements, essentially all typical MRI experiments and MRI pulse sequences can be used to record undersampled MRI data, and said MRI experiments are adapted by the skilled person to the MRI scanner and industrial sample based on the general knowledge of the skilled person. For example, the skilled person does not consider performing 31P-MRI measurements on samples that do not contain phosphorus. In particular, the method of the present invention preferably provides that the MRI measurements are T1-weighted or T2-weighted or T2*-weighted MRI measurements, the respective sequences typically using spin echo sequences being known in the art. With regard to potential nuclei, the method of the present invention preferably provides that the MRI measurements are 1H-, 13C-, 23Na- or 31P-MRI measurements, preferably 1H- or 13C-MRI measurements, more preferably 1H-MRI measurements.

[0032] It is neither possible nor expedient to define experimental parameters for an MRI experiment in general, since the relevant parameters strongly depend on the MRI scanner, the MRI experiment, the nuclei under study, the type of sample, the industrial characteristics of interest, and several other parameters. However, adjusting the parameters of an MRI experiment is a routine task for a person skilled in the art who knows how to measure and / or calibrate the parameters required for the adjustment and / or calibration of an MRI scanner. In practice, different parameters are often optimized to obtain the best signal-to-noise ratio, for example by incrementally adjusting the individual parameters. Although it is possible to calibrate the experimental parameters for any reference sample, it is expedient to at least supplement this calibration with measurements on reference samples of a predefined sample type. Correspondingly, the method of the invention in particular preferably provides that the parameters of the MRI measurement, in particular the pulse length, the development time and the repetition time, are obtained on the basis of calibration experiments carried out on industrial samples of a predefined sample type.

[0033] In the method of the invention, undersampled MRI data for at least one slice or at least a partial volume of the industrial sample is recorded, said slice may have a certain thickness or the volume, sometimes called "excitation volume", may be the entire volume of the industrial sample. Said recording depends on the MRI measurement employed. In the case of slices, slice selectivity of the MRI measurement is achieved using known techniques, typically using gradient coils, providing a gradient of magnetic field strength resulting in a gradient of resonance frequencies, allowing spatial separation of nuclei in the sample. In the case of volumes, known techniques such as 3D encoding techniques can be employed.

[0034] In some applications, it may be very advantageous to record exactly one slice of an industrial sample of a given sample type, for example for small industrial samples or for MRI measurements that are generally time-consuming, such as for samples with very long T1 times. Such restrictions may be applied, for example, to the samples, allowing a reliable prediction to be made as to where each predetermined feature is expected (for example, in the case of cherry seeds, it should be at least near the center of the fruit, etc.), thereby removing the need to record undersampled MRI data for further slices. However, in other cases, for example for the identification of potentially randomly distributed predetermined features, such as impurities or defects in the industrial sample that have a significant depth perpendicular to the slice, it may be advantageous to record two or more slices of each industrial sample. Thus, the method of the invention preferably records undersampled MRI data for two or more, preferably three or more, more preferably five or more slices of the industrial sample, for each industrial sample, preferably determining the spacing between the samples so that the slices are evenly distributed throughout the industrial sample, and preferably the multiple slices are recorded simultaneously. More specifically, the method of the invention preferably comprises a method in which each slice has an average thickness in the range of 0.5-10 mm, preferably in the range of 0.75-5 mm, more preferably in the range of 1-2 mm. Alternatively, the method of the invention preferably comprises recording undersampled MRI data for only one slice of the sample. Furthermore, in some applications, the method of the invention preferably comprises recording undersampled MRI data for at least a partial volume, preferably essentially the entire volume, of the industrial sample.

[0035] For the purpose of understanding the method of the present invention, the distinction between "undersampled MRI data" and "undersampled MRI raw data" and "processed MRI data obtained from processing undersampled MRI raw data" is understood to be as relevant as the distinction from normal non-undersampled MRI data. In accordance with the understanding of the skilled artisan, MRI data includes MRI raw data obtained by sampling (or digitizing) continuous time-dependent NMR signals (e.g., FIDs or echoes) of a series of different experiments in a two-dimensional or three-dimensional experiment (also called k-space) in an MRI measurement, as well as any representative data of undersampled MRI raw data obtained by processing it, in particular by Fourier transforming it. To obtain "undersampled MRI data", sampling is performed at fewer points than are essentially required to analyze the highest frequency contributing to the continuous time-dependent NMR signal and / or to analyze the phase angle of magnetization. For example, in the time dimension, the sampling rate to obtain a discrete time-dependent signal is less than twice the highest frequency of the continuous time-dependent NMR signal, i.e., the Nyquist rate. A corresponding concept applies to a number of signals of different phases, which can also be described as time-dependent signals with different spatial frequency components, and undersampling occurs when the interval between phase increments is larger than the corresponding Nyquist interval. In agreement with the understanding of those skilled in the art, the term "undersampled MRI raw data" includes all minor modifications of the undersampled MRI raw data, which are merely arbitrary modifications, such as adding or subtracting a fixed value, or multiplying with a fixed value.

[0036] Apart from undersampled MRI raw data, the term "undersampled MRI data" also includes processed MRI data resulting from processing the undersampled MRI raw data, said processing can in principle involve any data processing operation, but most often involves a one-, two- or three-dimensional Fourier transformation that finally leads to an MRI image containing aliasing artifacts caused by undersampling. Both the Nyquist rate and the Nyquist interval are well known to those skilled in the art, and recording undersampled MRI data presents no problem to those skilled in the art.

[0037] It should be noted that while the undersampled MRI data detected directly by the detector is typically undersampled MRI raw data, the processing of the undersampled MRI raw data and acquisition of the processed MRI data is considered to be "recording", similar to the natural understanding of those skilled in the art who consider, for example, the acquisition of an MRI image to be the recording of an MRI image.

[0038] Undersampling allows for particularly rapid MRI measurements and generally allows for a high throughput of samples. This advantageous effect is more pronounced when MRI data is undersampled in two or three dimensions, in which case a three-dimensional image is constructed using two phase encoding dimensions rather than only one slice being recorded. In the prior art, undersampling in one dimension is typically considered "bad enough", while undersampling in two or more dimensions is often considered inconvenient in prior art methods. However, a major advantage of the present invention is that the method of the present invention can be readily employed for reliable feature identification using MRI data undersampled in two or more dimensions and / or MRI data where undersampling is particularly severe due to sampling with a large difference between the Nyquist rate and the Nyquist interval.

[0039] Considering the above description, the method of the present invention preferably provides that in the MRI measurement, the sampling rate of the time-dependent signals is less than the Nyquist rate and / or the interval between the phase increments is greater than the Nyquist interval, preferably the sampling rate of the time-dependent signals is set less than the Nyquist rate and the interval between the phase increments is set greater than the Nyquist interval, in order to obtain undersampled MRI data. The method of the present invention preferably provides that the sampling rate of the time-dependent signals is less than 75%, preferably less than 50%, more preferably less than 25% of the Nyquist rate.

[0040] Similarly, the method of the present invention preferably provides for a spacing between phase increments that is 200% or more, preferably 300% or more, more preferably 400% or more of the Nyquist spacing, which corresponds to skipping rows in k-space. Most preferably, each feature is implemented with a corresponding preference.

[0041] In general, undersampling or reducing the sampling rate, respectively, can be achieved along any coding dimension, i.e. along frequency, phase, or slice coding.

[0042] As mentioned above, undersampled MRI data can include one or both of "undersampled MRI raw data" and "processed MRI data resulting from processing undersampled MRI raw data". The latter refers to any type of processed MRI data obtained from undersampled MRI raw data by performing a Fourier transform, especially in the phase encoding dimension and / or the time encoding dimension, preferably in both dimensions. It also refers to other linear or non-linear encoding operations applied to the undersampled MRI raw data. Those skilled in the art will understand that advantageously, no artificial generation of additional sampling points or any other method of improving data quality is performed on the processed MRI data, and such steps are not required by the method of the present invention. Because it is based on undersampled MRI raw data, the processed MRI data obtained by Fourier transform along the undersampled dimension includes at least one aliasing artifact. As understood by those skilled in the art, aliasing artifacts are effects that make different signals of a continuous-time NMR signal indistinguishable in a discrete time-dependent signal. While these aliasing artifacts are often detrimental to prior art methods, they advantageously do not hinder the method of the present invention, or at least do not hinder it as much, and may even be beneficial for training the machine learning module, as the aliasing artifacts allow certain features to be replicated in the samples, providing the machine learning module with more features to recognize. Correspondingly, when undersampled raw MRI data is obtained, the method of the present invention may preferably comprise processing the undersampled raw MRI data by Fourier transforming along at least one undersampled dimension to obtain processed MRI data including at least one aliasing artifact.

[0043] If the undersampled MRI raw data is directly analyzed, typically the highest processing speed and highest sample throughput as well as the greatest reduction in required computing power are possible, and exclusive analysis of the undersampled MRI raw data is the preferred option in all embodiments. Overall, the ability of the machine learning module to identify predetermined features in industrial samples in the undersampled MRI raw data, especially in the undersampled k-space data, and the option to eliminate any Fourier transformation or image processing steps, is one of the most important advantages of the method of the present invention. Therefore, in particular, the method of the present invention preferably comprises the undersampled MRI raw data, and said undersampled MRI data is preferably composed of the undersampled MRI raw data. As mentioned above, the method of the present invention preferably also comprises the undersampled MRI raw data undersampled in both the time dimension and the phase dimension. Correspondingly, the method of the present invention preferably trains the machine learning module to identify predetermined features in industrial samples of a given sample type from the undersampled MRI raw data.

[0044] It has been found that training a machine learning module to identify a given feature in industrial samples of a given sample type from undersampled raw MRI data can be more difficult than providing more processed data, e.g. requiring a larger training set. Therefore, in some applications it is preferred to use processed MRI data. Therefore, the method of the present invention preferably includes the undersampled MRI data, and said undersampled MRI data preferably consists of processed MRI data.

[0045] Here, a highly preferred route uses not only highly processed MRI data but also MRI images with aliasing artifacts, in particular some MRI scanners are optimized to return MRI images as output, so that the respective data may be more readily available. In this respect, the method of the invention preferably is such that the processed MRI data is obtained from processing the undersampled raw MRI data using a one-, two- or three-dimensional, preferably two-dimensional, Fourier transform, in particular along the frequency- and / or phase-encoding dimension, and / or the processed MRI data is a one-, two- or three-dimensional, preferably two-dimensional, Fourier transform of the undersampled raw MRI data, in particular along the frequency- and / or phase-encoding dimension.

[0046] Similarly, the method of the present invention preferably comprises processing the undersampled raw MRI data by Fourier transforming along at least one undersampled dimension to obtain processed MRI data comprising at least one aliasing artifact, and the machine learning module is trained to identify a predetermined feature in an industrial sample of a predefined sample type from the processed MRI data comprising at least one aliasing artifact. In particular, the method of the present invention preferably comprises obtaining the processed MRI data as an MRI image by Fourier transforming or by linearly or nonlinearly encoding, preferably by Fourier transforming, the undersampled raw MRI data along an undersampled frequency encoding dimension and an undersampled phase encoding dimension, the MRI image comprising at least one aliasing artifact, and the machine learning module is trained to identify a predetermined feature in an industrial sample of a predefined sample type from the MRI image comprising at least one aliasing artifact.

[0047] However, when employing the method of the invention, a very efficient method of the invention is obtained, as a good compromise between efficiency of analysis and required computational power, where the processed MRI data is obtained from processing the undersampled MRI raw data by Fourier transforming along exactly one undersampled dimension, resulting in processed MRI data containing at least one aliasing artifact. The use of undersampled MRI raw data and / or less processed processed MRI data rather than the complete MRI image has been found to be a very advantageous embodiment.

[0048] In a third step, the method of the invention comprises analysing the undersampled MRI data with an inference module which identifies a predetermined feature of the industrial sample. The skilled person will appreciate that the result of the method of the invention is a feedback of the inference module and thus information of whether the predetermined feature is present in the industrial sample or is present at a particular magnitude, respectively. The method of the invention therefore contributes to an advantageously fast processing speed, since it returns a relatively simple answer, for example in the form of "yes", "no" or "partially". In view of the method of the invention, it is neither necessary nor advantageous for the inference module to also return MRI data, in particular processed MRI data, since this would in most cases only increase the amount of computing power required and would increase the need for additional hardware (e.g. displays). Thus, the method of the invention preferably does not comprise a means for said inference module to output a visualized output of the processed MRI data, in particular the MRI image. Similarly, the method of the invention preferably does not configure the inference module to improve the resolution of the MRI image and / or to reconstruct the MRI image and / or to remove artifacts from the MRI image and / or to otherwise improve the quality of the MRI image, and most preferably none of these processing steps are performed in the method of the invention. The method of the invention is particularly preferred where no processed MRI data is provided as output of the inference module.

[0049] In the majority of cases, the method of the invention is preferably configured such that the inference module provides the results of the analysis in the form of classification and / or evaluation parameters and / or in step c) the inference module provides the results of the analysis, preferably in the form of classification and / or evaluation parameters.

[0050] In view of the above, it is highly preferred to classify the industrial samples depending on the identification of the features. This means that two or more classes based on grouping criteria are provided and the industrial samples are assigned to one or more of these classes based on the identification of the predefined features. For example, if the predefined feature is the amount of water in a glass bottle, the classes can be for example i) less than 95%, ii) between 95% and 98% and iii) more than 98%, where for example only samples of class ii) are further processed by sealing the bottle, since they are neither too full nor too little. In a highly efficient embodiment, this classification can be performed automatically by an inference module. Therefore, the method of the invention preferably further comprises: d) classifying the industrial samples based on the results of the analysis, preferably the classification is performed by the inference module, preferably the inference module is configured to classify the predefined type of industrial sample based on the specific results of the predetermined features, most preferably the multiple industrial samples are removed from the MRI scanner and sorted based on the classification.

[0051] In view of the above disclosure, it is understood that a major benefit of the present invention consists in the high availability and reliability of the rapid identification of a given feature and / or the rapid classification of an industrial sample, whereby the method of the present invention is particularly suitable for controlling subsequent processing steps, such as, for example, the transfer of eggs to a hatchery or the packaging of a product free of impurities, based on the output. The method of the present invention therefore preferably identifies the presence or absence or magnitude of said given feature and / or controls subsequent processing steps of said industrial sample depending on said presence or absence or magnitude of said given feature. Furthermore, the method of the present invention preferably provides that said industrial sample is only provided for further processing if, as a result of said analysis, preferably the presence or absence or magnitude of said given feature, satisfies a given quality criterion, and preferably that said industrial sample is discarded and / or repurposed if said given quality criterion is not met.

[0052] The inference module used in the method of the invention is used to analyze undersampled MRI data to identify predetermined characteristics of the industrial sample. The inference module itself is a physical device and can be a typical data processing device, e.g. a computer. In line with this, the inference module comprises a memory, i.e. a computer readable storage device, for storing e.g. data or software, and a processor, i.e. a digital circuit, capable of performing operations on external data sources, for controlling the inference module. Since the hardware infrastructure of the inference module is not important for the method of the invention, it is convenient to define the presence of only the basic components of an electronic data processing system, i.e. a memory and a processor, for both of which typical commercially available components can be used, and the inference module may also comprise other components of a typical data processing system, e.g. a power supply, a mouse, a keyboard, a display, a network connector, etc. The inference module may for example be part of a central controller controlling a conveyor and / or an MRI scanner.

[0053] The processor of the inference module is used to control the inference module, and the inference module is configured to provide undersampled MRI data as input to a machine learning module and to analyze the undersampled MRI data using the machine learning module. Those skilled in the art will understand that this configuration is established in a typical manner, for example, by using software that can be programmed by the skilled person himself or provided by a typical specialized programming company. Thus, the inference module typically includes a computer executable code that can be stored in a memory, including machine executable instructions or programs that cause the processor and the inference module to perform the respective tasks in the method of the present invention. A suitable design for assembly from a processor and a memory is disclosed, for example, in EP 3704666.

[0054] A core component of the inference module is the machine learning module, which is stored in the memory of the inference module, and is therefore usable by the inference module to analyze undersampled MRI data by providing the undersampled MRI data to the machine learning module, which processes inputs and identifies the predetermined features, in particular the presence, magnitude or absence of a predetermined feature, and the output of the machine learning module is processed by the inference module and a processor, respectively, and the output of the machine learning module may, for example, be stored in memory, provided to a user via an interface, or provided as input to a subsequent processing device (e.g., a sorter).

[0055] In recent years, employing machine learning techniques or so-called artificial intelligence has become a promising approach to solving different technical problems in several industrial fields. Today, machine learning, its core concepts and its implementation are well-known concepts, and several companies provide commercial solutions for implementing machine learning. Some skilled persons focusing on the field of MRI and / or the field of quality control may have to study some literature before setting up some of the more complex machine learning solutions completely on their own, but this skill is not necessarily required to implement the method of the present invention. Suitable machine learning algorithms and codes are available from the prior art and from commercial service providers as needed. In practice, in most industries, the implementation of the machine learning aspects of the present invention will probably be provided by skilled persons in the field of machine learning who will have no problem providing a suitable machine learning module in view of the disclosure of the present invention, and these skilled persons may for example be part of a team of employees or individuals of an external IT service provider. The machine learning module is similar to an MRI scanner in that it is often provided by a specialized company.

[0056] The inference module uses the machine learning module to analyze the undersampled MRI data. In accordance with the understanding of those skilled in the art, a machine learning module is an entity (e.g., a program) that actually analyzes, for example, the undersampled MRI data, and can provide the desired identification of the predetermined features, typically without task-specific programming. To this end, the machine learning module typically comprises both data and procedures for using the data to make predictions, sometimes referred to as predictive algorithms. The machine learning module itself does not necessarily need to be capable of further "learning" or "training" itself, but can exist, for example, in the form of a program that comprises a machine learning module, and can continue to fulfill its task in the same way.

[0057] In accordance with the understanding of those skilled in the art, a machine learning module can be obtained by executing (or fitting) a machine learning algorithm, often called a learning rule or learning algorithm, on a data set, often called a training set. Thus, a machine learning module can be obtained by executing a machine learning algorithm on a training set of data, or an existing machine learning module can be modified. Thus, by using a machine learning algorithm on a training set, a machine learning module is trained for its specific task. In the framework of the present invention, in accordance with the established expression, a machine learning module is defined as being trained, and training is specified by defining a training set in particular. An exemplary disclosure of the use of machine learning in the field of MRI can be referred to EP 3704666, which provides further information for further understanding of the technical background.

[0058] In practice, the choice of machine learning algorithm used to obtain a machine learning module mostly depends on the industrial sample and often more importantly on the predetermined features that need to be identified. However, the inventors have been able to identify suitable machine learning algorithms that can be used to generate said machine learning module depending on the respective machine learning algorithm. In particular, the method of the present invention is preferably such that said machine learning module is based on a machine learning algorithm, said machine learning algorithm being selected from the group consisting of regression algorithms, linear classifiers, instance-based algorithms, regularization algorithms, decision tree algorithms, Bayesian algorithms, clustering algorithms, association rule learning algorithms, artificial neural network algorithms, deep learning algorithms, dimensionality reduction algorithms, and ensemble algorithms, preferably linear classifiers, artificial neural network algorithms, and deep learning algorithms. Correspondingly, the method of the present invention is preferably such that said machine learning module is obtained by applying a machine learning algorithm to a training set, said machine learning algorithm being selected from the group consisting of regression algorithms, linear classifiers, instance-based algorithms, regularization algorithms, decision tree algorithms, Bayesian algorithms, clustering algorithms, association rule learning algorithms, artificial neural network algorithms, deep learning algorithms, dimensionality reduction algorithms, and ensemble algorithms, preferably linear classifiers, artificial neural network algorithms, and deep learning algorithms. As used herein, said machine learning algorithms are themselves known in the art and can be adopted by the skilled artisan taking into account the industrial sample and the given characteristics.

[0059] Based on this, the inventors have identified certain types of machine learning modules that have proven to provide excellent results in identifying predetermined features in industrial samples and are therefore explicitly preferred for all embodiments.The method of the present invention is therefore particularly preferred in that the machine learning module is a deep learning network or an artificial neural network, preferably a deep learning network.In this specification, these types of machine learning modules are known to those skilled in the art and are described in more detail, for example, in EP 3704666.

[0060] Due to the large amount of undersampled MRI data typically recorded during industrial-scale use of the method of the invention, it is highly recommended to provide a mechanism for continuous training of the machine learning module to continuously improve the performance and accuracy of the identification. Therefore, the method of the invention is particularly preferred, in which the inference module comprises a machine learning algorithm for generating or training a machine learning module based on a set of undersampled MRI raw data collected using the method, the machine learning algorithm being selected from the group consisting of regression algorithms, linear classifiers, instance-based algorithms, regularization algorithms, decision tree algorithms, Bayesian algorithms, clustering algorithms, association rule learning algorithms, artificial neural network algorithms, deep learning algorithms, dimensionality reduction algorithms and ensemble algorithms, preferably linear classifiers, artificial neural network algorithms and deep learning algorithms.

[0061] While a person skilled in the art will select a machine learning algorithm based on his general knowledge and his specific needs, what connects the machine learning modules of the present invention to each other is the training of the machine learning modules, and therefore the training set. According to the present invention, the machine learning modules are trained to identify the predetermined feature in industrial samples of the predefined sample type using a training set comprising undersampled MRI data recorded for different training samples of the predefined sample type, some of the training samples containing the predetermined feature and some of the training samples not containing the predetermined feature. The training of the machine learning modules is therefore supervised training, i.e. training based on input data with known results. In this training process, predictions are made and, if they are wrong, they are corrected manually or automatically. The desired level of accuracy is achieved by repetition of the above process.

[0062] Thus, training samples for training of a machine learning module identifying said predetermined feature in industrial samples of said predetermined sample type are represented by egg training samples identifying the predetermined feature in industrial samples of the same predetermined sample type, i.e. eggs of known outcome, i.e. eggs exhibiting the predetermined feature or not. In other words, prior to training of the machine learning module, the presence or absence of the predetermined feature in the training samples is known. It is known that the typical accuracy of a prediction depends on the quantity and quality of the training samples.

[0063] For some machine learning based techniques, for example for improving image quality or for removing artifacts, methods are proposed to generate artificial training sets, as disclosed in, for example, EP 3704666. However, the inventors have found this to be almost impossible and, more importantly, not convenient for the method of the present invention. Due to the fact that the method of the present invention is used in high throughput industries and is directed to identifying features relevant to these industries, potential samples exhibiting (or not exhibiting) the given feature are typically available in sufficient numbers, at least after a few days of operation. In most cases, suitable samples can also be generated artificially, for example by deliberately not removing cherry seeds from a set of training cherries. The required training data is easily obtained by recording undersampled MRI data for at least one slice or at least a partial volume of the training sample in an MRI measurement, as described for the method of the present invention above. Since typically only a few slices or partial volumes of an industrial sample are analyzed, it is possible to easily multiply the amount of available training data by changing the position and / or orientation of the industrial sample in the MRI scanner and performing new MRI measurements. Overall, the use of training sets obtained on real samples with real predetermined characteristics is typically preferred to yield the most powerful machine learning modules. Here, the respective training sets of undersampled training MRI data can also be easily obtained by those skilled in the art by recording undersampled MRI data on exemplary samples from their own industrial processes, preferably using the same MRI scanner as that used in the method of the present invention.In view of the above, the method of the present invention preferably comprises undersampled MRI data of different training samples of a given sample type recorded on the same type of MRI scanner as used in the method, and preferably similar, more preferably essentially identical experimental parameters are employed for recording the undersampled MRI data of the training samples used in the method.

[0064] As an alternative to the above-mentioned generation of a training set, the skilled person can use an externally provided training set, e.g. recorded in a reference facility, preferably including information about the experimental parameters used in the respective MRI measurements and / or varying the experimental parameters of the MRI measurements during the acquisition of the training set, in order to increase the adaptability of the training set and / or the trained machine learning module to different MRI scanners and / or different experimental parameters.Accordingly, the method of the invention preferably comprises: said machine learning module being trained to identify predetermined features in industrial samples of said defined sample type using said training set, each undersampled MRI data of said different training samples being linked with information about the MRI scanner used to obtain said undersampled MRI data and / or with experimental parameters of said MRI measurements, said machine learning module being trained to identify said predetermined features in industrial samples of said defined sample type on a number of different MRI scanners and / or under different experimental conditions, said inference module being configured to provide information about said MRI scanner and / or experimental parameters of said MRI measurements of said method as input to said machine learning module, said experimental parameters being preferably selected from the group comprising pulse length, pulse sequence, evolution time, repetition time, sampling rate, phase increment, temperature and number of scans.

[0065] The inventors have found that best training results for machine learning modules are typically obtained when the amount of training samples containing the predefined features is relatively large. Specifically, the method of the present invention is preferably adapted for training samples containing the predefined features in the range of 20-80%, preferably in the range of 30-70%, more preferably in the range of 40-60%, and most preferably in the range of 45-55%. When two or more predefined features are identified, training samples that only exhibit feature A can be used as negative examples of feature B, so that the machine learning module can be trained with a training set that is advantageous in most cases in that the amount of training samples that do not exhibit any of the predefined features can be significantly reduced.

[0066] For the sake of completeness, it is noted that the skilled artisan will appreciate that a significant lack of consistency between the training samples and the industrial samples used in the method of the invention may reduce the efficiency of said method, especially when combined with a relatively general definition of the sample type. For example, if the skilled artisan intends to use the method of the invention on chicken eggs, he will be aware that the accuracy and efficiency of the method of the invention may be reduced if the training set used on eggs contains 90% undersampled MRI data on ostrich eggs.

[0067] Similarly, the skilled person is aware that the above definition of the method of the present invention and the training of the machine learning module imply that the training set includes undersampled MRI data of the same format as the undersampled MRI data analyzed. A skilled person intending to analyze undersampled k-space data on chicken eggs would not consider a machine learning module trained on undersampled 3D-MRI-images of chicken eggs.

[0068] The inventors of the present invention have applied the method of the present invention to three exemplary industrial applications. In all these cases, the method of the present invention has proven to be particularly effective in determining certain characteristics of industrial samples in a rapid and reliable manner. For example, the hatchability of incubated eggs was predicted based on the identification of the relative sizes of the egg compartments. Similarly, seeds were classified for expected productivity based on the identification of certain characteristics associated with increased productivity. Furthermore, food samples, in particular chocolate pralines, were analyzed for the presence or absence of contaminants, and in particular the presence or absence of cherry seeds.

[0069] In view of the above, the method of the invention preferably comprises a step of predicting the hatchability of an egg based on the presence or absence or magnitude of a predetermined feature, preferably the egg white volume, the egg white texture, the egg white chemical composition, the egg white relative position, the egg white absolute position, the egg white magnetic properties and features derived from said features, preferably the egg white volume, the egg white texture, the egg white texture, the egg white dimensional ratio, the egg white blastoderm position and the egg white relative angular position. Correspondingly, the method of the invention preferably comprises a step of predicting the hatchability of an egg based on the presence or absence or magnitude of a predetermined feature, preferably the egg white volume and / or the egg yolk volume and / or the egg white volume and / or the egg white relative volume ratio and / or the egg white blastoderm position.

[0070] Similarly, the method of the invention preferably wherein the predefined sample type and the industrial sample are seeds and the predefined characteristic is selected from the group consisting of seed filling, seed texture, seed volume, seed shape, seed moisture content and seed lipid content. The method of the invention preferably wherein seed quality and / or productivity is predicted based on the presence or absence or magnitude of the predefined characteristic.

[0071] Furthermore, the method of the invention preferably provides that the given sample type and industrial sample are food products and that the predetermined feature is selected from the group consisting of the internal structure of the food product, in particular the texture, the spatial distribution of food components and / or contaminants, and the location and size of contaminants in the food product. Also, the method of the invention preferably provides that compliance with the product standard is assessed based on the presence or absence or size of the predetermined feature.

[0072] It will be clear to those skilled in the art that the present invention also relates to an inference module for analyzing undersampled MRI data of industrial samples of a predefined sample type using a machine learning module, preferably in a method of the present invention, said inference module comprising a memory for storing the machine learning module and a processor for controlling said inference module, said inference module being configured to provide undersampled MRI data as input to the machine learning module and to analyze said undersampled MRI data using said machine learning module, said machine learning module being trained to identify a predefined feature in industrial samples of said predefined sample type using a training set comprising undersampled MRI data of different training samples of said predefined sample type, some of said training samples comprising said predefined feature and some of said training samples not comprising said predefined feature. The inference module can be connected to a suitable MRI, for example a conventional industrial MRI scanner, to carry out the method of the present invention.

[0073] In view of this, the present invention also relates to an MRI system for carrying out the method of the invention, said MRI system being a) an MRI scanner for obtaining undersampled MRI data of an industrial sample of a given sample type; b) a conveyor for transporting a number of industrial samples to an MRI scanner; and c) an inference module of the present invention connected to said MRI scanner.

[0074] The invention also relates to a computer program product comprising instructions which, when executed by a computer, preferably by a reasoning module of the invention, cause said computer to carry out step c) of the method of the invention.

[0075] Also disclosed herein is a computer readable data carrier having stored thereon a computer program product of the present invention and a data carrier signal carrying the computer program product of the present invention.

[0076] The present invention will now be described in more detail and preferred embodiments of the invention will be disclosed with reference to the drawings. [Brief description of the drawings]

[0077] [Figure 1] FIG. 1 is a schematic diagram of the process steps of the present invention. [Figure 1] FIG. 2 is a schematic diagram of the structure of the inference module of the present invention. [Diagram 3] FIG. 3 is a schematic diagram of the principle of the method of the invention. [Figure 4] FIG. 4 is a schematic diagram of an exemplary MRI system of the present invention.

[0078] To facilitate understanding of the present invention, a schematic diagram of an exemplary method for automatic non-invasive identification of a predetermined feature in a multitude of industrial samples 102 of a given sample type according to the present invention is shown in FIG. 1. In a first step 12 of the method, at least one industrial sample 102 is transported to an MRI scanner 106. In a second step 14, an MRI measurement is performed on the industrial sample 102, and undersampled MRI data 300 for at least one slice or at least a partial volume of the industrial sample 102 is recorded. In a third step 16, the undersampled MRI data 300 acquired in the second step 14 is analyzed using an inference module 200. The inference module 200 includes a memory 202 that stores a machine learning module 204 and a processor 206 configured to control the inference module 200. The analysis procedure in the inference module 200 in the third step 16 is performed using the machine learning module 204, and the undersampled MRI data 300 recorded in the second step 14 are input to the inference module 200 and the machine learning module 204, respectively.

[0079] A schematic diagram of the structure of an inference module 200 suitable for use in the methods of the invention, particularly the methods shown in Figures 1 and 3, and in the MRI system 100 of Figure 4, is shown in Figure 2. The machine learning module 204 is trained to identify a predetermined feature in an industrial sample 102 of a predefined sample type. The training is performed using a suitable machine learning algorithm, and the training set 302 comprises undersampled MRI data 300 of different training samples of the predefined sample type to allow supervised training. The training set 302 consists, for example, of undersampled MRI data 300 of training samples of the predefined sample type with the predetermined feature and training samples of the predefined sample type without the predetermined feature. In other words, some of the training samples include the predetermined feature and some of the training samples do not include the predetermined feature. For example, the percentage of training samples in the training set 302 that include the predetermined feature is in the range of 45-55%. Based on this training of the machine learning module 204, the method can identify the presence or absence, or the magnitude, of the predetermined feature.

[0080] 2, the inference module 200 does not comprise means for providing a visualized output of the processed MRI data, such as an MRI image, e.g., in the form of a display. Correspondingly, the inference module 200 is not configured to improve in any way the quality of the MRI image, but rather to provide the results of the identification, and potentially the classification, to, e.g., a subsequent processing device (e.g., a sorting device).

[0081] In a particularly preferred embodiment of the inference module 200 shown in Fig. 2, i.e. a preferred option for all embodiments of the method or the MRI system 100, the machine learning module 204 is based on an artificial neural network algorithm or a deep learning algorithm. The machine learning module is therefore an artificial neural network or a deep learning module, and preferably the inference module 200 also comprises a machine learning algorithm for further training the machine learning module on the basis of the undersampled MRI data 300 recorded in accordance with the method of the invention.

[0082] Fig. 1 shows a preferred embodiment of the method of the invention, which includes a fourth step 18, in which the industrial samples 102 are also classified by the inference module 200. Said classification is based on the results of the analysis performed in the third step 16, which includes information on the presence or absence and / or magnitude of a predetermined feature in the industrial samples 102 of a given sample type. In a preferred embodiment, the classification performed in the fourth step 18 is used to sort the industrial samples 102 based on the results, preferably immediately after the industrial samples 102 are removed from the MRI scanner 106. For example, the industrial samples 102 can be classified into three classes 304, 306 and 308, as in the exemplary method of the invention shown in Fig. 3. In this example, the first class 304, the second class 306, and the third class 308 correspond to classifications "suitable" for further processing, "not suitable" for further processing, and "ambiguous / repetitive", and only the industrial samples 102 classified in the first class 304 are further processed, while the industrial samples 102 determined to be "not suitable" in the second class 306 are removed from the MRI scanner 106 for disposal or recycling of the samples for another use. Finally, the industrial samples 102 classified as "ambiguous / repetitive" in the third class 308, for example due to insufficient acquisition of undersampled MRI data 300 in the second step 14, are reintroduced into the MRI scanner 106 and the method is repeated to be classified into the first class 304 or the second class 306. FIG. 3 is a highly schematic diagram of the use of the machine learning module 204 for subsequent classification by the inference module 200 described above. Undersampled MRI data 300 is provided to the inference module 200 by the MRI scanner 106 and analyzed using a machine learning module 204, as indicated by the dashed double arrow, which has been trained using a training set 302 that includes the undersampled MRI training data, as indicated by the dotted double arrow.

[0083] An exemplary MRI system 100 of the present invention, including an inference module 200 of the present invention and thus suitable for implementing the method of the present invention, is shown diagrammatically in FIG. 4. The MRI scanner 106 includes, for example, a magnet generating a static magnetic field, said magnetic field having a magnetic field strength, for example, in the range of 0.05 to 1 Tesla. The MRI scanner 106 includes a measurement zone located in the center of the static magnetic field, in which the industrial sample 102 is placed during the MRI measurement. The MRI scanner 106 typically includes a three-dimensional magnetic gradient coil configured to induce a magnetic gradient in the static magnetic field, and a radio frequency coil configured to apply a radio frequency pulse to the industrial sample 102, the latter typically being placed next to the measurement zone. For example, the gradient coil and the radio frequency coil can be controlled by a gradient controller and a radio frequency controller, respectively. A radio frequency detector can be used to detect the signal, which results in recording of undersampled MRI raw data that can be provided to the inference module 200.

[0084] 4, several industrial samples 102 are transported in the direction indicated by the arrow by a transport device 104 to an MRI scanner 106. The transport device 104 is controlled for example by a transport controller, which is managed for example by a central controller, which also supervises other controllers in the MRI scanner 106, for example via a data channel or wireless communication.

[0085] As shown in the example of Fig. 4, the industrial samples 102 are arranged in parallel in five rows, i.e. in a regular pattern defined by suitable holders (not shown). The method of the invention is therefore applied simultaneously to a number of industrial samples 102, i.e. 20 industrial samples 102. After the undersampled MRI data 300 have been recorded for a holder of industrial samples 102, said holder is removed from the measurement zone and the next holder is placed in the MRI scanner 106 for MRI measurement. The method is therefore applied simultaneously several times in succession to a number of industrial samples 102. For example, the undersampled MRI data 300 are recorded for only one slice of each industrial sample 102, the row arrangement of the industrial samples 102 allowing a very efficient measurement and an easy adjustment of the magnetic gradient coils 115. Alternatively, for example, the entire volume of each industrial sample 102 can also be recorded, for example by 3D acquisition techniques.

[0086] For example, the sampling rate of the time-dependent signals in the MRI measurement can be set to only 25% of the Nyquist rate. In addition, the spacing between phase increments is 400% of the Nyquist interval, thereby skipping rows in k-space. Other experimental parameters of the MRI measurement, such as pulse length or repetition time, can be obtained based on typical calibration experiments performed on industrial samples 102 of a given sample type (e.g., training samples included in the training set 302).

[0087] Preferably, the method is optimized so that the time to record undersampled MRI data 300 for one holder of industrial samples 102 is less than 5 seconds, and it is possible to operate the method at a rate of at least 5000 industrial samples 102 per hour.

[0088] In the above example method optimized for high throughput and reduced computational power, the undersampled MRI data 300 consists of undersampled raw MRI data (eg, undersampled k-space data).

[0089] In an alternative embodiment of the above method optimized to provide the highest verifiability via a process operator, the undersampled MRI data 300 comprises an undersampled MRI image that includes at least one aliasing artifact.

[0090] In a second alternative to the above method, which is believed to be an advantageous compromise between the other two methods, the undersampled MRI data 300 comprises processed MRI data obtained from a Fourier transform of undersampled MRI raw data along one undersampled dimension.

[0091] The method of the present invention has been used in three different applications in different industries, and it has been found that the method is highly suitable for identifying predetermined features in a wide variety of industrial samples 102 of a given sample type, and classifying the industrial samples 102 if necessary.

[0092] In a first demonstration, an egg 102 was tested using the method of the present invention, and both the egg yolk volume and the egg white volume were determined in a slice through the center of the egg. The egg hatchability was then predicted based on the magnitude of the egg white to egg yolk volume ratio. The inventors found that incubated chicken eggs showed an 83-88% increase in hatchability when the ratio of the estimated egg yolk size divided by the egg white size based on a slice through the center of the egg was in the range of 0.44-0.50. Correspondingly, the method of the present invention set the ratio in the range of 0.44-0.50 as a predetermined feature. Using this approach, a large number of chicken eggs could be analyzed, and those that exhibited the feature were classified as "high hatchability" and labeled accordingly. Each analysis used processed MRI data obtained by performing a 2-D Fourier transform of undersampled raw MRI data along both the frequency and phase encoding dimensions, where the undersampled raw MRI data contained aliasing artifacts. Alternatively, other types of undersampled MRI data 300 can be used. A training set 302 was easily obtained by using undersampled MRI data 300 recorded with the same MRI scanner 106 used previously for incubated eggs, allowing supervised learning by manually analyzing the sizes of compartments and their ratios in the training sample using typical software for image analysis.

[0093] In a second demonstration, chocolate pralines containing cherries were analyzed for possible incomplete removal of the cherry seeds in order to classify the pralines based on their compliance with product standards. Undersampled MRI raw data can be used in both time and phase dimensions to obtain a particularly fast measurement speed, resulting in a particularly high efficiency since the predefined features and their potential locations in the pralines are quite well defined, although other types of undersampled MRI data can be used instead. A training set 302 was easily obtained by using undersampled MRI data 300 recorded on chocolate pralines containing cherries whose seeds were not intentionally completely removed.

[0094] In a third demonstration, the method of the present invention was used to identify several features in apple seeds, such as volume and water content, features that are often associated with expected productivity. Here, undersampled MRI data 300 obtained by Fourier transforming the undersampled MRI raw data along one undersampled dimension was used, but other types of undersampled MRI data could be used instead. A training set 302 was easily obtained by using undersampled MRI data 300 recorded with the same MRI scanner used previously for apple seeds, and each feature in the training sample was manually analyzed using typical software for image analysis. Using the method of the present invention, it was found that a reliable and rapid identification of each feature was possible. [Explanation of symbols]

[0095] 12 Process a) 14 Process b) 16 Process c) 18 Process d) 100 MRI Systems 102 Industrial Samples 104 Conveyor 106 MRI Scanner 200 Inference Module 202 Memory 204 Machine Learning Module 206 Processors 300 Undersampled MRI data 302 Training Set 304 First Class 306 Second Class 308 Third Class

Claims

1. a) transporting an industrial sample (102) of a predetermined sample type to an MRI scanner (106); b) recording undersampled MRI data (300) for at least one slice or at least a partial volume of the industrial sample in an MRI measurement, wherein the undersampled MRI data (300) Undersampled raw MRI data containing multiple time-dependent signals for different phases, and / or Processed MRI data obtained from processing the undersampled raw MRI data and c) analyzing the undersampled MRI data (300) using a machine learning module (204) trained to identify the predetermined features in industrial samples (102) of the given sample type from the undersampled MRI data (300) with an inference module (200) that identifies predetermined features in the industrial samples (102); The inference module (200) includes a memory (202) that stores the machine learning module (204) and a processor (206) that controls the inference module (200), the inference module (200) being configured to provide the undersampled MRI data (300) as input to the machine learning module (200) and to analyze the undersampled MRI data (300) using the machine learning module (204), the machine learning module (204) being trained to identify the predetermined feature in industrial samples (102) of the predetermined sample type using a training set (302) that includes undersampled MRI data (300) of different training samples of the predetermined sample type, some of the training samples including the predetermined feature and some of the training samples not including the predetermined feature.

2. moreover d) classifying the industrial samples (102) based on the results of the analysis, wherein the classification is performed by the inference module (200), the inference module (200) being configured to classify the predetermined type of industrial samples (102) based on the results of the identification of the predetermined features, and the number of industrial samples (102) being removed from the MRI scanner (106) and sorted based on the classification.

3. 3. The method of claim 1, wherein the undersampled MRI data (300) comprises undersampled raw MRI data, the undersampled MRI data (300) consisting of undersampled raw MRI data.

4. 3. The method of claim 1, wherein the undersampled MRI data comprises processed MRI data, the undersampled MRI data consisting of processed MRI data.

5. 3. The method of claim 1, wherein the processed MRI data is obtained as an MRI image by Fourier transforming or linearly or nonlinearly encoding the undersampled raw MRI data along an undersampled frequency encoding dimension and an undersampled phase encoding dimension, the MRI image containing at least one aliasing artifact, and the machine learning module is trained to identify the predetermined feature in industrial samples of the predetermined sample type from the MRI image containing at least one aliasing artifact.

6. 3. The method of claim 1 or 2, operated at a rate of 1000 or more industrial samples (102) per hour.

7. 3. The method according to claim 1 or 2, wherein the method is applied to multiple industrial samples (102) sequentially and / or simultaneously to multiple industrial samples (102).

8. 3. The method of claim 1, further comprising identifying the presence or size of the predetermined feature and / or controlling subsequent processing of the industrial sample (102) depending on the presence or size of the predetermined feature.

9. 3. The method of claim 1, wherein the predetermined sample type is selected from animal products, plants, and products derived from these raw materials.

10. 3. The method according to claim 1, wherein the predetermined characteristics are selected from the group consisting of chemical composition, physical properties, in particular magnetic properties, structural features, anatomical features, biological features, morphological dimensions, sample structure, spatial distribution of elements in the industrial sample, and the presence of impurities.

11. the machine learning module (204) is trained to identify the predetermined feature in industrial samples (102) of the predetermined sample type using the training set (302), and each undersampled MRI data (300) of the different training samples is linked with information about the MRI scanner (106) used to obtain the undersampled MRI data (300) and / or experimental parameters of the MRI measurements, and the machine learning module (204) is trained to identify the predetermined feature in industrial samples (102) of the predetermined sample type on a number of different MRI scanners (106) and / or under different experimental conditions, and the inference module (200) is configured to provide information about the MRI scanner (106) and / or experimental parameters of the MRI measurements of the method as input to the machine learning module (204), the experimental parameters being selected from the group comprising pulse length, pulse sequence, evolution time, repetition time, sampling rate, phase increment, temperature, and number of scans; or 3. The method of claim 1, wherein the training set (302) comprises undersampled MRI data (300) of different training samples of the predetermined sample type recorded with the same type of MRI scanner as used in the method, and experimental parameters similar to those used in the method are used to record the undersampled MRI data (300) of the training samples.

12. The method of claim 1 or 2, wherein the machine learning module (204) is a deep learning network or an artificial neural network.

13. 3. The method of claim 1 or 2, wherein an inference module (200) analyzes undersampled MRI data (300) of industrial samples (102) of a predetermined sample type using a machine learning module (204), comprising: a memory (202) that stores the machine learning module (204) and a processor (206) that controls the inference module (200); configured to provide undersampled MRI data (300) as input to the machine learning module (204) and analyze the undersampled MRI data (300) using the machine learning module (204); The machine learning module (204) is trained to identify a predetermined feature in industrial samples (102) of the predetermined sample type using a training set (302) including undersampled MRI data (300) of different training samples of the predetermined sample type, some of the training samples including the predetermined feature and some of the training samples not including the predetermined feature, and an inference module (200).

14. a) an MRI scanner (106) for obtaining undersampled MRI data (300) of an industrial sample (102) of a given sample type; b) a conveyor (104) for transporting a number of industrial samples (102) to said MRI scanner (106); 14. An MRI system (100) for implementing the method of any one of claims 1 or 2, comprising: c) an inference module (200) according to claim 13 connected to said MRI scanner (106).

15. 14. A computer program product comprising instructions that, when executed by a computer by an inference module (200) according to claim 13, cause said computer to carry out step c) of the method according to any one of claims 1 or 2.

Images