A method for detection of deviations in skeletal structures of a fish spine, and a system for detection of deviations in skeletal structures of a fish spine

Abstract

The disclosure relates to a method for detection of deviations in skeletal structures of a fish spine (1), the method comprising the steps: capturing an x-ray image (4) of a fish (20) or a portion of a fish including a portion of or the complete spine (1), providing the x-ray image (4) of the spine (1) or the portion of the spine (1) of the fish, the x-ray image (4) spanning a range of two or more vertebrae (2), to a processing means (10), analyzing by the processing means (10) the x- ray image (4) (4), and determining from the x-ray image (4) a lesion state of one or more of: - any one or more vertebrae (2) of the range of vertebrae (2), and - any one or more, or the absence of, intervertebral spaces between two vertebrae (3) in the range of vertebrae (2). The disclosure further relates to a system for detection of deviations in skeletal structures of a fish spine (1).

Description

[0001] A method for detection of deviations in skeletal structures of a fish spine, and a system for detection of deviations in skeletal structures of a fish spine

[0002] Technical field

[0003] The present disclosure relates to a method for detection of deviations in skeletal structures of a fish spine, and a system for detection of deviations in skeletal structures of a fish spine. More specifically, the disclosure relates to a method for detection of deviations in skeletal structures of a fish spine, and a system for detection of deviations in skeletal structures of a fish spine as defined in the introductory parts of the independent claims.

[0004] Background art

[0005] Skeletal deformities are invariably present in fish aquaculture production, although in highly variable numbers. This is true for many aquaculture species, and specifically for Salmonidae such as for example Atlantic salmon.

[0006] Vertebral lesions of any size can be detected, from small deviations in one or two vertebrae to lesions involving most or all of the approximately 58 vertebrae in an individual salmon. Lesions need to be of a certain size in order to be detected on external examination. Only when the external shape of the fish is affected is the term "deformed fish" used. Whether or not a vertebral lesion leads to recognition of a deformed fish depends on size of lesion (number of affected vertebrae), the degree of deviation from normal shape of the affected vertebrae, the location of the lesion and to some extent the type and causality of the lesion.

[0007] During the period 1995-2010, deformed fish were recognized as a major problem in Norwegian salmon aquaculture, and a considerable research effort was done in order to identify causal factors and find ways of preventing the problem. During these years, slaughter groups in which 10-30% of fish were downgraded at slaughter were not uncommon. By 2010, the prevalence of deformed fish was reduced to a level where it no longer received much attention, typically 0-5% downgrading at slaughter. By this time, other health and welfare issues were considered more urgent, and this has been the case since. During 2016-2020, a new wave of deformed fish appeared, but less prominent than the previous. In the present situation (2023), the majority of slaughter groups still have a low prevalence of deformed fish (<5%), but reports indicate that the problem is now increasing again and is not under control.

[0008] The low level of focus on deformed fish in the industry from 2010 and onwards led to a gradual disintegration of competence and a near stand-still in development of knowledge. The producers tend to regard deformed fish as a quality issue, being secondary to the more urgent health problems which keep appearing. Producers are not required to record deformities under the fish health regulations, and consequently, no statistics on prevalence is publicly available. Thus it is assumed that the true prevalence of vertebral deformities is underestimated in all stages of salmon production, and that producers to a large extent are oblivious to the true situation. Fish health groups have raised deformities as an important and concerning issue in salt- and fresh-water stage production of rainbow trout and salmon.

[0009] In commercial production, fish which are identified as deformed during quality screening at slaughter should be sorted out. Fish with moderate deviation in body shape are downgraded from superior to ordinary, which sells at a lower price. Fish with severe deviations in body shape are discarded and the value of the product is lost. Quality screening of fish at slaughter is a matter of skill, and it is a problem in the industry that the resulting quality grading may be unreliable. In particular, in fish groups with high prevalence of deformities, it is a big problem that a drift in standards is observed, allowing for more fish with deformities to be sold as superior. Some markets may be tolerant towards deformities, whereas some markets require fish to be flawless. A relevant example is fish which are machine filleted at the customer's, where even a moderate vertebral deformity can cause technical problems and cause losses. In addition to any issues created between producer and consumer, it is a problem that there is increasing concern over the industry's reputation concerning fish welfare. A further problem is related to undetected deformities that represent a lost opportunity for learning and optimization of the production for the producer.

[0010] It is a problem that a large population of fish is bred to full production size before being sorted out or downgraded.

[0011] There is thus a need for improved processes for increasing the level of quality of produced fish, and to reduce downgraded or discarded portions of the production in fish farming industry.

[0012] Summary

[0013] It is an object of the present disclosure to mitigate, alleviate or eliminate one or more of the above-identified deficiencies and disadvantages in the prior art and solve at least the above mentioned problem. According to a first aspect there is provided a method for detection of deviations in skeletal structures of a fish spine, the method comprising the steps: capturing an x-ray image of a fish or a portion of a fish including a portion of or the complete spine, providing the x-ray image of the spine or the portion of the spine of the fish, the x-ray image spanning a range of two or more vertebrae, to a processing means, analyzing by the processing means the x-ray image, and determining from the x-ray image a lesion state of one or more of: any one or more vertebrae of the range of vertebrae, and any one or more, or the absence of, intervertebral spaces between two vertebrae in the range of vertebrae.

[0014] Vertebral deformities in fish can be correctly identified by use of X-ray, often at very early juvenile stage. Present disclosure defines how to determine the lesion state of the fish to enable early discovery of threats to raising a healthy specimen, and discharging of those which has a high probability of developing to non-healthy and / or deformed entities. According to some embodiments, the step capturing the x-ray image is capturing the x-ray image of a live fish.

[0015] Previous use of X-ray in the fish industry has been to analyze slaughtered fish. Present enclosure provides a novel use of X-ray also when the fish is alive, even in the beginning of the breeding process.

[0016] According to some embodiments, the step capturing the x-ray image is preceded by: euthanizing the fish.

[0017] The method and system may be used equally efficiently on both live and euthanized fish species.

[0018] Different lesion states need different measures to qualify whether a fish species is to be considered abnormal, and according to some embodiments further determining if the lesion state is abnormal or normal, wherein the lesion state may be derived for one or more of:

[0019] - individual vertebra,

[0020] - sub-groups of vertebrae, and

[0021] - the fish spine as a whole.

[0022] According to some embodiments, the analysis of the x-ray image the method comprises the steps, for each of the vertebrae :

[0023] - determining the four corners,

[0024] - deriving the midpoints along the four sides defined by the corners,

[0025] - determining the global centroid of the quadrilateral defined by the four corners.

[0026] Many of the abnormalities / lesion states of a fish may be predicted by analyzing these features. According to some embodiments, the method comprises the steps: determining the radiopacity of each vertebrae, finding the common vertex of two opposing internal cones of each of the vertebrae, determining the center of each of the vertebrae defined by the common vertex, and determining a center-line passing close to centers of the vertebrae by mapping a smooth model of the range of vertebrae.

[0027] The center-line may be used to find deviations in dorsal or ventral direction from an expected line / curvature.

[0028] According to some embodiments, the method further comprises the steps: determining a lesion state as a fusion when it is detected one or more of:

[0029] - the intervertebral space defined by the space between the right edge of a first vertebra and the left edge of a second vertebra to the right of the first vertebra, is less than a first relative threshold wherein the first relative threshold is a function of at least one of: the area defined by the four corners of the vertebra, the vertebral number, fish species, fish length, fish weight, and age,

[0030] - the first and second vertebra has no discernible intervertebral space, and - the number of neural spines emanating from a vertebra exceeds the expected number of emanating neural spines from a healthy vertebra.

[0031] According to some embodiments, the method further comprises the steps: determining a lesion state as an osteopenia when it is detected one or more of:

[0032] - the intervertebral space defined by the space between the right edge of a first vertebra and the left edge of a second vertebra to the right of the first vertebra, is above a second threshold, wherein the second threshold is a function of at least one of: an area defined by the four corners of the vertebra, the vertebral number, fish species, fish length, fish weight, age, in combination with one or more of:

[0033] - under-mineralised vertebrae, and

[0034] - intervertebral space, when relative large intervertebral space is detected in a series of vertebrae

[0035] According to some embodiments, the method further comprises the steps: determining a lesion state as a platyspondylia when it, for a series of vertebrae, is detected:

[0036] - a vertebra with a width-to-height ratio lower than a third threshold, and

[0037] - the intervertebral space defined by the space between the right edge of a first vertebra and the left edge of a second vertebra to the right of the first vertebra, is less than a fourth threshold (tN4), wherein

[0038] - the third and fourth thresholds are functions of at least one of: an area defined by the four corners of the vertebra, the vertebral number, fish species, fish length, fish weight, and age.

[0039] According to some embodiments, the method further comprises the steps: determining a lesion state as a hyper dense vertebra when it is detected:

[0040] - a single vertebra with increased radiopacity compared to the neighbouring vertebrae.

[0041] According to some embodiments, the method further comprises the steps: determining a lesion state as a cross stich vertebrae when it is detected:

[0042] - a series of vertebrae with decreased or missing intervertebral spaces and dorsoventral shifts.

[0043] According to some embodiments, the method further comprises the steps: determining a lesion state as a missing intervertebral space when it is detected:

[0044] - early cross stitch vertebrae visualizing as a small intervertebral space and opposing shifts.

[0045] According to some embodiments, the method further comprises the steps: determining a lesion state as an axial deviation of type scoliosis when it is detected:

[0046] - deviation from expected lateral curvature of the spinal axis.

[0047] According to some embodiments, the method further comprises the steps: determining a lesion state as an axial deviations of type lordosis or kyphosis when it is detected: - deviation locally from an expected line / curvature of the dorsal or ventral curvature respectively of the spinal axis.

[0048] According to some embodiments, the step: capturing an x-ray image of a fish or a portion of a fish including a portion of or the complete spine, is performed on a fish belonging to, but not limited to, any of the family: Salmonidae, Gadidae, and Scophthalmidae / Pleuronectidae.

[0049] Adapting the teachings of present disclosure in some machine learning (ML) settings provides for the following:

[0050] According to some embodiments, the method comprises the steps: providing a first storage means for storing x-ray images of a set of vertebrae of fish spine ; providing a first model generation means for generating a first determination model through machine learning, to which an x-ray image of a set of vertebrae is input and from which a set of keypoints of the set of vertebrae is output, wherein the keypoints are for each vertebra one or more of:

[0051] - the four corners, and

[0052] - the vertex of each of two opposing internal cones of the vertebra, the method further comprising the steps: using training data containing the x-ray images of the set of vertebrae stored in the first storage means, and the keypoints of the set of vertebrae ; providing a first reception means for receiving an input of an x-ray image of a set of vertebrae of a fish, and providing a first processing means for outputting, using the generated first determination model that has been generated by the first model generation means, a set of keypoints of the set of vertebrae based on the x-ray image of the set of vertebrae inputted to the first reception means.

[0053] According to some embodiments, the method comprises the steps: determining vertebra characteristics of each vertebra, being one or more, but not limited to, of:

[0054] - the midpoints along the four sides defined by the corners,

[0055] - the global centroid of the quadrilateral defined by the four corners,

[0056] - the center-line,

[0057] - the radiopacity, and further providing a second storage means for storing sets of vertebra characteristics and keypoints of set of vertebrae of a fish spine; providing a second model generation means for generating a second determination model through machine learning, to which a set of vertebra characteristics and keypoints of a set of vertebrae is input and from which a lesion state of the set of vertebrae is output, the method further using training data containing the vertebra characteristics and keypoints of the set of vertebrae stored in the second storage means, and the lesion state of the set of vertebrae; providing a second processing means for receiving an input of the set of vertebra characteristics and keypoints of the set of vertebrae, and outputting, by the second processing means using the generated second determination model that has been generated by the second model generation means, a lesion state of the set of vertebrae based on the set of vertebra characteristics and keypoints of the set of vertebrae inputted to the second processing means.

[0058] The ML concept may also be comprised in a single step such as:

[0059] According to some embodiments, the method comprises the steps: providing a third storage means for storing x-ray images of a set of vertebrae of fish spine ; providing a third model generation means for generating a third determination model through machine learning, to which an x-ray image of a set of vertebrae is input and from which a lesion state of the set of vertebrae is output, using training data containing the x-ray images of the set of vertebrae stored in the third storage means, and the lesion state of the set of vertebrae ; providing a third reception means for receiving an input of an x-ray image of a set of vertebrae of a fish, and providing a third processing means for outputting, using the generated third determination model that has been generated by the third model generation means, a lesion state of the set of vertebrae based on the x-ray image of the set of vertebrae inputted to the third reception means.

[0060] According to some embodiments, the method comprises the steps: providing a sorting device, and sorting the fish according to predefined sorting matrices based on the output lesion state.

[0061] According to a second aspect there is provided a system for detection of deviations in skeletal structures of a fish spine comprising: a first storage means for storing x-ray images of a set of vertebrae of fish spine ; a first model generation means for generating a first determination model through machine learning, to which an x-ray image of a set of vertebrae is input and from which a set of keypoints of the set of vertebrae is output, using training data containing the x-ray images of the set of vertebrae stored in the first storage means, and the keypoints of the set of vertebrae ; a first reception means for receiving an input of an x-ray image of a set of vertebrae of a fish, and a first processing means for outputting, using the generated first determination model that has been generated by the first model generation means, a set of keypoints of the set of vertebrae based on the x-ray image of the set of vertebrae inputted to the first reception means.

[0062] According to some embodiments the keypoints are for each vertebra one or more of:

[0063] - the four corners and

[0064] - the vertex of each of two opposing internal cones of the vertebra.

[0065] According to some embodiments, the system comprises: vertebra characteristics of each vertebra derived from the keypoints one or more, but not limited to, of:

[0066] - the midpoints mL,mB,mR,mTalong the four sides defined by the corners CNW,CNE,CSW,CSE,

[0067] - the global centroid me of the quadrilateral defined by the four corners CNW,CNE,CSW,CSE,

[0068] - the center-line / ,

[0069] - the radiopacity, a second storage means for storing sets of the vertebra characteristics and keypoints of set of vertebrae of a fish spine ; a second model generation means for generating a second determination model through machine learning, to which a set of the vertebra characteristics and keypoints of a set of vertebrae is input and from which a lesion state of the set of vertebrae is output, using training data containing the vertebra characteristics and keypoints of the set of vertebrae stored in the second storage means, and the lesion state of the set of vertebrae; a second processing means for receiving an input of the set of the vertebra characteristics and keypoints of the set of vertebrae, and outputting, by the second processing means using the generated second determination model that has been generated by the second model generation means, a lesion state of the set of vertebrae based on the set of the vertebra characteristics and keypoints of the set of vertebrae inputted to the second processing means.

[0070] According to a third aspect there is provided a system for detection of deviations in skeletal structures of a fish spine comprising: a third storage means for storing x-ray images of a set of vertebrae of fish spine ; a third model generation means for generating a third determination model through machine learning, to which an x-ray image of a set of vertebrae is input and from which a lesion state of the set of vertebrae is output, using training data containing the x-ray images of the set of vertebrae stored in the third storage means, and the lesion state of the set of vertebrae ; a third reception means for receiving an input of an x-ray image of a set of vertebrae of a fish, and a third processing means for outputting, using the generated third determination model that has been generated by the third model generation means, a lesion state of the set of vertebrae based on the x- ray image of the set of vertebrae inputted to the third reception means.

[0071] According to some embodiments: one or more of: any of the first, second, and third storage means are the same storage means, any of the first, second, and third model generation means are the same model generation means, any of the first, second, and third determination model are the same determination model, any of the first, second, and third reception means are the same reception means, and any of the first, second, and third processing means are the same processing means.

[0072] Effects and features of the third aspects are to a large extent analogous to those described above in connection with the first and second aspect. Embodiments mentioned in relation to the first and second aspects are largely compatible with the third aspect.

[0073] The present disclosure will become apparent from the detailed description given below. The detailed description and specific examples disclose preferred embodiments of the disclosure by way of illustration only. Those skilled in the art understand from guidance in the detailed description that changes and modifications may be made within the scope of the disclosure.

[0074] Hence, it is to be understood that the herein disclosed disclosure is not limited to the particular component parts of the device described or steps of the methods described since such device and method may vary. It is also to be understood that the terminology used herein is for purpose of describing particular embodiments only, and is not intended to be limiting. It should be noted that, as used in the specification and the appended claim, the articles "a", "an", "the", and "said" are intended to mean that there are one or more of the elements unless the context explicitly dictates otherwise. Thus, for example, reference to "a unit" or "the unit" may include several devices, and the like. Furthermore, the words "comprising", "including", "containing" and similar wordings does not exclude other elements or steps.

[0075] Terminology

[0076] The terms "vertebral deformities" and "vertebral lesions" are used to describe deviations in the skeletal structures of the spine, as observed on X-ray images.

[0077] The term "radiopacity" is used to describe any form of radiodensity, radiographic opacity, radiographic image density, commonly used to describe the opacity to the radio wave and X-ray portion of the electromagnetic spectrum. Opacity is a relative quantity used to describe any area that preferentially absorbs and therefore appears more opaque than the surrounding area on a radiograph, such as in present disclosure where the vertebrae is examined. The higher the radiopacity, the less dark the X-ray image: it reflects mineralization of the bone tissue.

[0078] The term "cross-stitch vertebrae" is a deformity associated with reduced fish welfare and significant pathological changes in the affected vertebrae. Axial, radially distributed lesions of the compact bone of vertebral endplates appear unique for cross-stitch vertebrae.

[0079] The term "set of vertebrae" is used to comprise any number of vertebrae, from one or more single vertebrae, one or more groups of vertebrae, and all vertebrae.

[0080] Brief descriptions of the drawings

[0081] The above objects, as well as additional objects, features and advantages of the present disclosure, will be more fully appreciated by reference to the following illustrative and non-limiting detailed description of example embodiments of the present disclosure, when taken in conjunction with the accompanying drawings.

[0082] Figure 1 shows an x-ray images of complete spine of a Salmon at 4 growth stages

[0083] Figure 2 shows an x-ray image of two smolt, one healthy (the upper) and one with a lesion state (the lower)

[0084] Figure 3 is a table showing various lesion states in set of vertebrae of fish species according to an embodiment of the present disclosure

[0085] Figure 4A shows an x-ray image of a healthy fish at production stage

[0086] Figure 4B shows an x-ray image of a fish with multiple lesion states at production stage Figure 4C shows an x-ray image of a fish with multiple lesion states at production stage

[0087] Figure 4D shows an x-ray image of a fish with multiple lesion states at production stage

[0088] Figure 5A show an example of key points derived from a group of 3 vertebrae

[0089] Figure 5B show a 3D drawing of a vertebra with two opposing internal cones of the vertebra and center point

[0090] Figure 6 shows an illustration of an example of a system according to an embodiment of the present disclosure.

[0091] Figure 7 shows 3 examples of scoliosis in X-ray image taken in a dorsoventral direction.

[0092] Detailed description

[0093] The present disclosure will now be described with reference to the accompanying drawings, in which preferred example embodiments of the disclosure are shown. The disclosure may, however, be embodied in other forms and should not be construed as limited to the herein disclosed embodiments. The disclosed embodiments are provided to fully convey the scope of the disclosure to the skilled person.

[0094] In present disclosure, when analyzing an x-ray image of a fish, it is in figure 1 - 5 presumed that the anterior / head end is pointing to the left, the posterior / tail end to the right, the dorsal / back of the fish is up, and the ventral / abdomen is down in a left right ventral view. In figure 7 it is shown X-ray image taken in a dorsoventral direction, to be able to show lateral deviations 17 in the vertebral column / spine. When analyzing a vertebra and the intervertebral spaces of one or more vertebrae, it is assumed that the vertebra is the left most entity, and the corresponding intervertebral space is to the right. The vertebral number when looking at the vertebral column / spine start at 1 at the head side and increase the further to the right it is shown on the X-ray image.

[0095] The first aspect of this disclosure shows a method for detection of deviations in skeletal structures of a fish spine 1, the method comprising the steps: capturing an x-ray image 4 of a fish 20 or a portion of a fish including a portion of or the complete spine 1, providing the x-ray image 4 of the spine 1 or the portion of the spine 1 of the fish, the x-ray image 4 spanning a range of two or more vertebrae 2, to a processing means 10, analyzing by the processing means 10 the x-ray image 4, and determining from the x-ray image 4 a lesion state of one or more of: - any one or more vertebrae 2 of the range of vertebrae 2, and - any one or more, or the absence of, intervertebral spaces between two vertebrae 3 in the range of vertebrae 2.

[0096] Digital radiography was introduced as an option for fish radiography in the 2000's. Digital imaging is, however, not much used due to the mobile units are very expensive, and the process of analyzing slaughtered fish species are a very manual process not feasible for a production optimization tool. It has now been discovered that vertebral deformities in fish can be correctly identified by use of X- ray, often at very early juvenile stage. Present disclosure defines how to determine the lesion state of the fish, in some embodiments enabling early discharging of deformed entities.

[0097] Present enclosure provides a novel use of X-ray also when the fish is alive, even in the beginning of the breeding process. Thus the step of capturing an x-ray image 4 may be executed on live fish 20. The advantage is that the process may be performed early, or during, the breeding process, not only at the end of life when the fish is slaughtered. An X-ray image may be taken of the specimen without any disturbance or danger to the specimen itself. When using the method for sorting out 22 defect fish, and sort in 21 healthy fish this is done at lowest possible risk to the healthy fish.

[0098] That said, there is no hindrance for the technique of present disclosure to euthanizing the fish 20 before capturing the x-ray image 4. Typically this may be done in a quality process handling and sorting the fish species before slaughtering.

[0099] It is a desire to determine as early as possible if a fish specimen is to be disposed of. Present disclosure provides a method and system for determining if a detected lesion state is abnormal or normal, wherein the lesion state may be derived for one or more of: - individual vertebra 3, - subgroups of vertebrae 2, and - the fish spine 1 as a whole.

[0100] Different lesion state requires different level of vertebral examination. For some states it is sufficient to detect one defect vertebra, for others a set of vertebrae is needed. Some other times a set of vertebrae is a need for constructing a comparator value for which the other examined vertebra / vertebrae is compared to.

[0101] Capturing an x-ray image 4 of a fish 20 or a portion of a fish including a portion 2 of or the complete spine 1, is encompassed by the present disclosure also when it is performed on a fish 20 belonging to any of the family: Salmonidae, Gadidae, and Scophthalmidae / Pleuronectidae.

[0102] Other examples of farmed fish species of interest for the present disclosure includes, but is not limited to: sea bass, sea bream, tuna, karp, pangasius, tilapia, breams, Anguillidae, koi, goldfish, perch, and catfish.

[0103] Most advantageous is the method and system for detection of deviations in skeletal structures of a fish spine according to present disclosure when applied to fish species wandering between fresh water and salt water before smoltification. Examples of such species are Salmonidae and Anguillidae. These transitions have great impact on the development of the spine, and most types of vertebral deformities may continue to develop as the fish grows, thus small lesions observed in juveniles and smolts can develop into more severe lesions by the time fish reach slaughter size.

[0104] Image reading has previously been a manual procedure. Each fish in the image is evaluated visually. A trained person may determine if an individual is normal. As soon as a deviation is noted, additional time is required in order to identify type of lesion and which of the vertebrae are affected. If considered relevant, additional measurements of morphological features may be done, e.g. width:height ratios of vertebrae or distance between vertebrae.

[0105] Studies have identified a number of different vertebral lesions in farmed salmon. A guide to classification was published in 2009 (Witten et al., 2009), but is of limited value in this context, as it describes the deviations of single vertebra and gives little guidance towards causal factors and interpretation of observations. Also, this guide does not take the size of fish or the progression of lesions with time into account.

[0106] In an example fish species, here Salmon, starts life in freshwater, and a number of studies indicate that conditions of egg incubation and juvenile rearing are key to development of normal bone structures. Vaccination is done in freshwater, at a size of approximately 30-60g. The standard size for transfer to seawater pens is 80-100g. Industry driven developments have led to a range of different rearing strategies around the smolt stage, and the trend is that smolt size is increasing and / or that smolts are kept in land-based units for a period of time beyond 100g, before transfer to open net pens. At the same time, the industry is working towards land based rearing also in the final on-growing stages towards slaughter. Fish in land-based units can be handled and monitored more closely than fish in open net pens.

[0107] Vertebral deformities can be induced at any life stage from fertilization of the egg and up to slaughter, but the juvenile stages in freshwater are the most critical. Most types of vertebral deformities may continue to develop as fish grows, thus small lesions observed in juveniles and smolts may develop into more severe lesions by the time fish reach slaughter size. The expected morphological expression of the various types of vertebral deformities varies over the life cycle.

[0108] There are typically several paths to a solution when adapting an image processing task, and with the development of machine learning approaches and the significant advances the field has shown in the last 10-15 years, the work-flow enables a significant shift with more of the engineering being handled by machine learning. A key method in the approach of the present disclosure is in one embodiment a form of transfer learning where a model developed for one task is adapted to a related task. This leads to good performance with minimal requirements on data and architecture development.

[0109] In one embodiment the design in present disclosure, as see in figure 5A, has framed the machine learning part as the task of detecting key-points in the image representing important information on each of the vertebrae. This may in one embodiment be the four corners of the vertebra. In a second embodiment the center of each vertebra might be targeted as an output with a similar methodology.

[0110] This low-level information may then be used to express diagnostics relative to fish at both the vertebral and spine levels. This approach hence combines the detection task of various diagnoses into a "black box" part and an "expert system" part, the latter incorporating the experience of experts in the domain. This way, outcomes may be better explained to users.

[0111] In a second embodiment the "expert system" might be substituted by a second machine learning task, wherein the key points are used as input and detection of diagnosis is the output.

[0112] The following outlines a first approach to the machine learning case providing for a "black box" Artificial intelligence step, and one "expert step" for determination of diagnosis. The example is developed using the fish species Salmon.

[0113] • Machine learning data set in present example: o X-ray images have been taken of a set of live / immobilised fish. The resolution and other parameters are known regarding the image. o A subset of these images has been annotated: the four corners of every vertebrae in every fish.

[0114] • Machine learning model development for: o The Detectron2 (PyTorch) framework is used to fine-tune a keypoint-detection model using a pre-trained model from the Detectron2 model zoo:

[0115] ■ COCO-Keypoints / keypoint_rcnn_R_50_FPN_3x.yaml: defines the model architecture.

[0116] ■ COCO-Keypoints / keypoint_rcnn_R_50_FPN_3x.yaml: defines the model weights. o The framework has been slightly modified so more data augmentations techniques is accessible via the configuration system (yaml files). This includes rotations and contrast. o Hyperparameter adjustments has been made to the learning rate and its reduction as a function of training iterations. o It was found that when the original images (ca. 20 Mpixels) were split into tiles and processing each tile independently, it was possible to run the system on a smaller processing machine (limited GPU / RAM resources). This enabled the method according to present enclosure to deviate from the original framework which normally re-sizes images to a fixed format and allowing for working with full resolution with no deterioration in quality. Also important is the choice of parameters for utilizing the image. o Post-processing of model output is performed in order to stitch the predictions from each tile to correspond to the original image and to remove overlapping detections from different tiles. o Further post-processing is necessary to remove outliers and to aggregate individual vertebral detections into spines / fish.

[0117] • With the set of key points extracted from each vertebra a range of metrics are calculated and described in the following sections. The key points and parameters of the metrics may be comprising one or more of:

[0118] Table 1: Metrics

[0119] When the "black box" module has provided the metrics and key points a diagnostic table may be constructed for diagnosis in the "expert step" module as follows: i

[0120] Table 2: Diagnosis table and symbol table

[0121] Present disclosure provides a method for the analysis of the x-ray image 4 comprising the steps, for each of the vertebrae 2:

[0122] - determining the four corners CNW,CNE,CSW,CSE,

[0123] - deriving the midpoints mL,mB,mR,mTalong the four sides defined by the corners CNW,CNE,CSW,CSE, and

[0124] - determining the global centroid me of the quadrilateral defined by the four corners CNW,CNE,CSW,CSE-

[0125] The method further comprises the steps: determining the radiopacity of each vertebra, finding the common vertex of two opposing internal cones 30 of each of the vertebrae 3, determining the center re of each of the vertebrae 2 defined by the common vertex, and

[0126] - determining a center-line I passing close to the vertebrae 2 centers by mapping a smooth model of the range of vertebrae 2.

[0127] When analyzing the X-ray images, and defining the key aspects of the vertebrae, the methods and embodiments may utilize a variation of apparatuses, diagnostic tools and image processing equipment. The picture / X-ray machine resolution, gain and quality varies. The thresholds and setup parameters are set accordingly. Thus, the thresholds are dynamically set in accordance with the equipment used. The thresholds will further depend on the objects / fish species to be examined, and the thresholds are set relative one or more of the selected number of vertebrae, fish species, fish length, fish weight, and fish age. Further aspects that may play a role may be variations in habitat, fodder composition, vaccination programs and others.

[0128] A lesion state is a fusion when it is detected one or more of:

[0129] - the intervertebral space / Rdefined by the space between the right edge of a first vertebra and the left edge of a second vertebra to the right of the first vertebra, is less than a first relative threshold tNiwherein the first relative threshold tNiis a function of at least one of: the area A defined by the four corners cNw, cNE, Csw, CSE of the vertebra, the vertebral number, fish species, fish length, fish weight, and age,

[0130] - the first and second vertebra has no discernible intervertebral space / R, and the number of neural spines emanating from a vertebra exceeds the expected number of emanating neural spines from a healthy vertebra.

[0131] The number of neural spines emanating from a vertebra can be used to count the number of original vertebrae, and hence grade the Fusion to actual number of vertebrae being fused.

[0132] A lesion state is an osteopenia when it is detected one or more of:

[0133] - the intervertebral space / Rdefined by the space between the right edge of a first vertebra and the left edge of a second vertebra to the right of the first vertebra, is above a second threshold tN2, wherein the second threshold tN2 is a function of at least one of: an area A defined by the four corners CNW, CNE, CSW, CSE of the vertebra, the vertebral number, fish species, fish length, fish weight, age, in combination with one or more of:

[0134] - undermineralized vertebrae 2, and

[0135] - intervertebral space when relative large intervertebral space / Ris detected in a series of vertebrae 2.

[0136] The orientation of the vertebrae, and the corresponding intervertebral space is in present disclosure chosen to be the intervertebral space to the right of the vertebra in question. This orientation is true when the fish lies on its right side, and the fish head is to the left. It should be understood that different orientation might be chosen within the inventive concept of present disclosure without deviating from the invention. Any other direction would cause he direction of mapping the vertebrae and their corresponding intervertebral spacing to change correspondingly.

[0137] A lesion state is a platyspondylia when it, for a series of vertebrae 2, is detected:

[0138] - a vertebra 2 with a width-to-height W / H ratio lower than a third threshold tN3, and - the intervertebral space / Rdefined by the space between the right edge of a first vertebra and the left edge of a second vertebra to the right of the first vertebra, is less than a fourth threshold (tN4), wherein

[0139] - the third and fourth thresholds tNs,tN4 are functions of at least one of: an area A defined by the four corners cNw, cNE, Csw, CSE of the vertebra, the vertebral number, fish species, fish length, fish weight, and age.

[0140] A lesion state is a hyper dense vertebra 2 when it is detected:

[0141] - a single vertebra 2 with increased radiopacity 4 compared to the neighbouring vertebrae 2.

[0142] A lesion state is a cross stich vertebrae 2 when it is detected:

[0143] - a series of vertebrae 2 with decreased or missing intervertebral spaces and dorsoventral shifts.

[0144] Cross stich vertebrae may comprise deformity associated with reduced fish welfare and significant pathological changes in the affected vertebrae. Axial, radially distributed lesions of the compact bone of vertebral endplates appear unique for cross-stitch vertebrae, (around vaccination time)

[0145] A lesion state is a missing intervertebral space when it is detected:

[0146] - early cross stitch vertebrae 2 visualizing as a small intervertebral space / Rand opposing shifts.

[0147] A lesion state is an axial deviation of type scoliosis when it is detected:

[0148] - deviation from an expected lateral curvature of the spinal axis.

[0149] Figure 7 shows an illustration in a dorsoventral projection of the spine of 3 fish examples suffering from deviation from the expected lateral curvature of the posterior / tail end of the spine, and shows the signs of scoliosis. The images have been captured by a further X-ray machine arranged to take dorsoventral images of the fish.

[0150] A lesion state is an axial deviations of type lordosis or kyphosis when it is detected:

[0151] - deviation locally from an expected line / curvature of the dorsal or ventral curvature respectively of the spinal axis.

[0152] Some of the diagnostics are based on well-defined rules involving certain metrics.

[0153] For example, a very relevant metric is the width-to-height ratio W / H. However, while both "tall" and "fat" vertebrae are symptomatic, the effective thresholds or scale are dependent on several factors including vertebra index, and the weight / length / age of fish.

[0154] The database used for training in the machine learning step contains a large number of fish for which diagnostic information is available, and this has been used to help define more precisely where the boundaries between "normal" and "abnormal" lies. The distinction between abnormal and normal is applied to individual vertebra, sub-groups of vertebrae (e.g. vertebral numbers 30-40) or to the fish as a whole.

[0155] Based on these discoveries, the idea of using image analysis and automated procedures in X-ray image evaluation was introduced. This increases the diagnostic capacity and gives more objective image reading, combined with the high capacity of digital imaging.

[0156] Present disclosure builds on the findings:

[0157] • Vertebral deformities are indicators of suboptimal rearing conditions

[0158] • Correcting such suboptimal conditions is beneficial for fish welfare

[0159] • The aquaculture industry must control fish deformities in order to gain respect in public opinion

[0160] • Controlling and reducing incidence of vertebral deformities will increase profit

[0161] • The key to controlling and reducing the incidence of vertebral deformities is access to tools that can give a correct assessment of the problem

[0162] • X-ray technology is the most relevant option for development of such a tool

[0163] According to a further embodiment of the present disclosure a method adapted to utilize machine learning in the diagnosis of fish based on x-ray image of the fish spine comprises the steps: providing a first storage 40 means for storing x-ray images 4 of a set of vertebrae 2 of fish spine 1; providing a first model generation means for generating a first determination model through machine learning, to which an x-ray image 4 of a set of vertebrae 2 is input and from which a set of keypoints of the set of vertebrae 2 is output, wherein the keypoints are for each vertebra 3 one or more of:

[0164] - the four corners CNW,CNE,CSW,CSE, and

[0165] - the vertex rc of each of two opposing internal cones 30 of the vertebra 3, and the method further the method comprises the steps: using training data containing the x-ray images 4 of the set of vertebrae 2 stored in the first storage means 40, and the keypoints of the set of vertebrae 2 ; providing a first reception means 50 for receiving an input of an x-ray image 4 of a set of vertebrae 2 of a fish 20, and providing a first processing means 10 for outputting, using the generated first determination model that has been generated by the first model generation means, a set of keypoints of the set of vertebrae 2 based on the x-ray image 4 of the set of vertebrae 2 inputted to the first reception means 50.

[0166] The method further comprises the steps: determining vertebra characteristics of each vertebra, being one or more of:

[0167] - the midpoints along the four sides defined by the corners CNW,CNE,CSW,CSE,

[0168] - the global centroid me of the quadrilateral defined by the four corners CNW,CNE,CSW,CSE,

[0169] - the radiopacity, and further providing a second storage 40' means for storing sets of the vertebra characteristics and keypoints of set of vertebrae 3 of a fish spine 1; providing a second model generation means for generating a second determination model through machine learning, to which a set of the vertebra characteristics and keypoints of a set of vertebrae 3 is input and from which a lesion state of the set of vertebrae 3is output, using training data containing the vertebra characteristics and keypoints of the set of vertebrae 3 stored in the second storage means 40', and the lesion state of the set of vertebrae; providing a second processing means 10' for receiving an input of the set of the vertebra characteristics and keypoints of the set of vertebrae 2, and outputting, by the second processing means 10' using the generated second determination model that has been generated by the second model generation means, a lesion state of the set of vertebrae 2 based on the set of the vertebra characteristics and keypoints of the set of vertebrae 2 inputted to the second processing means 10'.

[0170] According to a further embodiment of the present disclosure a method adapted to utilize machine learning in the diagnosis of fish based on x-ray image of the fish spine, encompassing both the "black box" module and "expert step" module in a single ML / AI approach comprises the steps: providing a third storage means 40" for storing x-ray images 4 of a set of vertebrae 2 of fish spine 1; providing a third model generation means for generating a third determination model through machine learning, to which an x-ray image 4 of a set of vertebrae 2 is input and from which a lesion state of the set of vertebrae 3 is output, using training data containing the x-ray images 4 of the set of vertebrae 2 stored in the third storage means 40", and the lesion state of the set of vertebrae 2; providing a third reception means 50" for receiving an input of an x-ray image 4 of a set of vertebrae 2 of a fish, and providing a third processing means 10" for outputting, using the generated third determination model that has been generated by the third model generation means, a lesion state of the set of vertebrae 2 based on the x-ray image 4 of the set of vertebrae 2 inputted to the third reception means 50", and providing a sorting device 25,26, and sorting the fish 20 according to a predefined sorting matrices based on the output lesion state.

[0171] The second aspect of this disclosure shows a system for detection of deviations in skeletal structures of a fish spine 1 comprising: a first storage means 40 for storing x-ray images 4 of a set of vertebrae 2 of fish spine 1; a first model generation means for generating a first determination model through machine learning, to which an x-ray image 4 of a set of vertebrae 2 is input and from which a set of keypoints of the set of vertebrae 2 is output, using training data containing the x-ray images 4 of the set of vertebrae 2 stored in the first storage means 40, and the keypoints of the set of vertebrae 2; a first reception means 50 for receiving an input of an x-ray image 4 of a set of vertebrae 2 of a fish 20, and a first processing means 10 for outputting, using the generated first determination model that has been generated by the first model generation means, a set of keypoints of the set of vertebrae 2 based on the x-ray image 4 of the set of vertebrae 2 inputted to the first reception means. The keypoints for each vertebra 3 may be, but not limited to, one or more of:

[0172] - the four corners CNW,CNE,CSW,CSE, and

[0173] - the vertex rc of each of two opposing internal cones 30 of the vertebra.

[0174] The system may further comprise: vertebra characteristics of each vertebra derived from the keypoints one or more, ut not limited to, of:

[0175] - the midpoints along the four sides defined by the corners CNW,CNE,CSW,CSE,

[0176] - the global centroid me of the quadrilateral defined by the four corners CNW,CNE,CSW,CSE,

[0177] - the center-line / ,

[0178] - the radiopacity, a second storage means 40' for storing sets of the vertebra characteristics and keypoints of set of vertebrae 3 of a fish spine 1; a second model generation means for generating a second determination model through machine learning, to which a set of the vertebra characteristics and keypoints of a set of vertebrae 3 is input and from which a lesion state of the set of vertebrae 3 is output, using training data containing the vertebra characteristics and keypoints of the set of vertebrae 3 stored in the second storage means 40', and the lesion state of the set of vertebrae; a second processing means 10' for receiving an input of the set of the vertebra characteristics and keypoints of the set of vertebrae 2, and outputting, by the second processing means 10' using the generated second determination model that has been generated by the second model generation means, a lesion state of the set of vertebrae 2 based on the set of the vertebra characteristics and keypoints of the set of vertebrae 2 inputted to the second processing means 10'.

[0179] The third aspect of this disclosure shows a system for detection of deviations in skeletal structures of a fish spine 1 comprising: a third storage means 40" for storing x-ray images 4 of a set of vertebrae 2 of fish spine 1; a third model generation means for generating a third determination model through machine learning, to which an x-ray image 4 of a set of vertebrae 2 is input and from which a lesion state of the set of vertebrae 3is output, using training data containing the x-ray images 4 of the set of vertebrae 2 stored in the third storage means, and the lesion state of the set of vertebrae 2; a third reception means 50" for receiving an input of an x-ray image 4 of a set of vertebrae 2 of a fish 20, and a third processing means 10" for outputting, using the generated third determination model that has been generated by the third model generation means, a lesion state of the set of vertebrae 2 based on the x-ray image 4 of the set of vertebrae 2 inputted to the third reception means 50".

[0180] It is also furnished a system wherein one or more of: any of the first, second, and third storage means 40, 40', 40" are the same storage means, any of the first, second, and third model generation means are the same model generation means, any of the first, second, and third determination model are the same determination model, any of the first, second, and third reception means 50, 50', 50" are the same reception means, and any of the first, second, and third processing means 10, 10', 10" are the same processing means.

[0181] The person skilled in the art realizes that the present disclosure is not limited to the preferred embodiments described above. The person skilled in the art further realizes that modifications and variations are possible within the scope of the appended claims.

[0182] Additionally, variations to the disclosed embodiments can be understood and effected by the skilled person in practicing the claimed disclosure, from a study of the drawings, the disclosure, and the appended claims.

Claims

Claims1. A method for detection of deviations in skeletal structures of a fish spine (1), the method comprising the steps: capturing an x-ray image (4) of a fish (20) or a portion of a fish including a portion of or the complete spine (1), providing the x-ray image (4) of the spine (1) or the portion of the spine (1) of the fish, the x-ray image (4) spanning a range of two or more vertebrae (2), to a processing means (10), analyzing by the processing means (10) the x-ray image (4), and determining from the x-ray image (4) a lesion state of one or more of:- any one or more vertebrae (2) of the range of vertebrae (2), and- any one or more, or the absence of, intervertebral spaces (lR) between two vertebrae (3) in the range of vertebrae (2).

2. The method according to claim 1, wherein the step capturing the x-ray image (4) is capturing the x-ray image (4) of an alive fish (20).

3. The method according to claim 1, wherein the step capturing the x-ray image (4) is preceded by: euthanizing the fish (20).

4. The method according to any one of claim 1 to 3, wherein: determining if the lesion state is abnormal or normal, wherein the lesion state may be derived for one or more of:- individual vertebra (3),- sub-groups of vertebrae (2), and- the fish spine (1) as a whole.

5. The method according to any one of the previous claims, wherein the analysis of the x-ray image (4) comprising the steps, for each of the vertebra (2):- determining the four corners (cNw, cNE, Csw, CSE),- deriving the midpoints (mt, mB,R, mT) along the four sides defined by the corners (cNw, cNE, Csw, CSE), deriving the global centroid (me ) of the quadrilateral defined by the four corners (cWw cNE,Csw, CSE)-6. The method according to claim 5, further comprising the steps: determining the radiopacity of each vertebra, finding the common vertex of two opposing internal cones (30) of each of the vertebrae (3), determining the center (rc ) of the vertebra (2) defined by the common vertex,- and determining a center-line ( / ) passing close to the centers of the vertebrae (2) center by mapping a smooth model of the range of vertebrae (2).

7. The method according to claim 5 and 6, further comprising the steps: determining a lesion state as a fusion when it is detected one or more of:- the intervertebral space ( / R) defined by the space between the right edge of a first vertebra and the left edge of a second vertebra to the right of the first vertebra, is less than a first relative threshold (tNi) wherein the first relative threshold (tNi) is a function of at least one of: the area (A) defined by the four corners (cNw, cNE, Csw, cEE) of the vertebra, the vertebral number, fish species, fish length, fish weight, and age, the first and second vertebra has no discernible intervertebral space ( / «), and the number of bone rays emanating from a vertebra exceeds the expected number of emanating bone rays from a healthy vertebra. center8. The method according to claim 5 and 6, further comprising the steps: determining a lesion state as an osteopenia when it is detected one or more of:- the intervertebral space ( / R) defined by the space between the right edge of a first vertebra and the left edge of a second vertebra to the right of the first vertebra, is above a second threshold (fe), wherein the second threshold (tW2) is a function of at least one of: an area (A) defined by the four corners (cNw, cNE, Csw, cEE) of the vertebra, the vertebral number, fish species, fish length, fish weight, age, in combination with one or more of:- undermineralized vertebrae (2), and when the relative large intervertebral space ( / R) is detected in a series of vertebrae (2).

9. The method according to claim 5 and 6, further comprising the steps: determining a lesion state as a platyspondylia when it, for a series of vertebrae (2), is detected:- a vertebra (2) with a width-to-height (W / H) ratio lower than a third threshold (tNs) and- the intervertebral space ( / R) defined by the space between the right edge of a first vertebra and the left edge of a second vertebra to the right of the first vertebra, is less than a fourth threshold (tN4), wherein the third and fourth thresholdsare functions of at least one of: an area (A) defined by the four corners (cNw, cNE, Csw, CSE) of the vertebra, the vertebral number, fish species, fish length, fish weight, and age, .10.The method according to claim 5 and 6, further comprising the steps: determining a lesion state as a hyper dense vertebrae (2) when it is detected:- a single vertebra (2) with increased radiopacity (4) compared to the neighbouring vertebrae (2).11.The method according to claim 5 and 6, further comprising the steps: determining a lesion state as cross stich vertebrae (2) when it is detected:- a series of vertebrae (2) with decreased or missing intervertebral spaces and dorsoventral shifts.12.The method according to claim 5 and 6, further comprising the steps: determining a lesion state as a missing intervertebral space when it is detected:- early cross stitch vertebrae (2) visualizing as a small intervertebral space ( / R) and opposing shifts.13.The method according to claim 5 and 6, further comprising the steps: determining a lesion state as an axial deviation of type scoliosis when it is detected:- deviation from expected lateral curvature of the spinal axis from a dorsoventral projection.14.The method according to claim 5 and 6, further comprising the steps: determining a lesion state as an axial deviations of type lordosis or kyphosis when it is detected:- deviation locally from an expected line / curvature of the dorsal or ventral curvature respectively of the spinal axis .15.The method according to any one of the previous claims, wherein the step: capturing an x- ray image (4) of a fish (20) or a portion of a fish including a portion (2) of or the complete spine (1), is performed on a fish (20) belonging to any of the family: Salmonid, Gadidae, and Scophthalmidae / Pleuronectidae.16.The method according to claim 1, further comprising the steps: providing a first storage (40) means for storing x-ray image (4)s of a set of vertebrae (2) of fish spine (1); providing a first model generation means for generating a first determination model through machine learning, to which an x-ray image (4) of a set of vertebrae (2) is input and from which the radiopacity and a set of keypoints of the set of vertebrae (2) is output, the keypoints are for each vertebra (3) one or more of: the four corners (cNw, cNE, Csw, CSE), and the vertex (rc ) of each of two opposing internal cones (30) of the vertebra (3), the method further comprising the steps: using training data containing the x-ray images (4) of the set of vertebrae (2) stored in the first storage means (40), and the keypoints of the set of vertebrae (2); providing a first reception means (50) for receiving an input of an x-ray image (4) of a set of vertebrae (2) of a fish (20), and providing a first processing means (10) for outputting, using the generated first determination model that has been generated by the first model generation means, a set of keypoints of the set of vertebrae (2) based on the x-ray image (4) of the set of vertebrae (2) inputted to the first reception means (50).17.The method according to claim 16, further comprising the steps: determining vertebra characteristics of each vertebra, being one or more of:- the midpoints mE,mB,mR,mTalong the four sides defined by the corners CNW, CNE, CSW, CSE,- the global centroid me of the quadrilateral defined by the four corners CNW, CNE, CSW, CSE,- the center-line / ,- the radiopacity, and further providing a second storage (40') means for storing sets of the vertebra characteristics and keypoints of sets of vertebrae (3) of fish spines (1); providing a second model generation means for generating a second determination model through machine learning, to which a set of the vertebra characteristics and keypoints of a set of vertebrae (3) is input and from which a lesion state of the set of vertebrae (3) is output, using training data containing the vertebra characteristics and keypoints of the group of vertebrae (3) stored in the second storage means (40'), and the lesion state of the group of vertebrae;providing a second processing means (10') for receiving an input of the set of the vertebra characteristics and keypoints of the set of vertebrae (2), and outputting, by the second processing means (10') using the generated second determination model that has been generated by the second model generation means, a lesion state of the set of vertebrae (2) based on the set of the vertebra characteristics and keypoints of the set of vertebrae (2) inputted to the second processing means (10').18.The method according to claim 17, further comprising the steps: providing a sorting device (25, 26), and sorting the fish (20) according to a predefined sorting matrices based on the output lesion state.

19. A system for detection of deviations in skeletal structures of a fish spine (1) comprising: a first storage means (40) for storing x-ray image (4)s of a set of vertebrae (2) of fish spine (1); a first model generation means for generating a first determination model through machine learning, to which an x-ray image (4) of a set of vertebrae (2) is input and from which a set of keypoints of the set of vertebrae (2) is output, using training data containing the x-ray image (4)s of the set of vertebrae (2) stored in the first storage means (40), and the keypoints of the set of vertebrae (2); a first reception means (50) for receiving an input of an x-ray image (4) of a set of vertebrae (2) of a fish (20), and a first processing means (10) for outputting, using the generated first determination model that has been generated by the first model generation means, a set of keypoints of the set of vertebrae (2) based on the x-ray image (4) of the set of vertebrae (2) inputted to the first reception means.

20. The system according to claim 19, wherein: the keypoints are for each vertebra (3) one or more of:- the four corners (cNw, cNE, Csw, CSE), and- the vertex (rc ) of each of two opposing internal cones (30) of the vertebra.21.The system according to claim 19 or 20, further comprising: vertebra characteristics of each vertebra derived from the keypoints one or more, but not limited to, of:- the midpoints mE,mB,mR,mTalong the four sides defined by the corners CNW,CNE,CSW,CSE,- the global centroid me of the quadrilateral defined by the four corners CNW,CNE,CSW,CSE,- the center-line / ,- the radiopacity, a second storage means (40') for storing sets of keypoints of set of vertebrae (3) of a fish spine(1); a second model generation means for generating a second determination model through machine learning, to which a set of the vertebra characteristics and keypoints of a set of vertebrae (3) is input and from which a lesion state of the set of vertebrae (3) is output, using training data containing the vertebra characteristics and keypoints of the set of vertebrae (3) stored in the second storage means (40'), and the lesion state of the set of vertebrae; a second processing means (10') for receiving an input of the set of the vertebra characteristics and keypoints of the set of vertebrae (2), and outputting, by the second processing means (10') using the generated second determination model that has been generated by the second model generation means, a lesion state of the set of vertebrae (2) based on the set of the vertebra characteristics and keypoints of the set of vertebrae(2) inputted to the second processing means (10').22.The system according to claim 21, wherein: one or more of: any of the first and second storage means (40, 40', 40") are the same storage means, any of the first and second model generation means are the same model generation means, any of the first and second determination model are the same determination model, any of the first, and second reception means (50, 50', 50") are the same reception means, and any of the first, and second processing means (10, 10', 10") are the same processing means.

Images