METHOD FOR ANALYSIS OF AN IMAGE OF A TOOTHARCH

Description

technical field

[0001] The present invention relates to the field of image analysis of dental arches. State of the art

[0002] The latest orthodontic treatments use images to assess treatment situations. This assessment is typically carried out by an orthodontist, which requires the patient to send these images to the orthodontist, or even to make an appointment. A method for the automatic analysis of mouth images using neural networks is known from Kelwin Fernandez et al.: "Teeth / Palate and Interdental Segmentation Using Artificial Neural Networks", September 17, 2012 (2012-09-17), ARTIFICIAL NEURAL NETWORKS IN PATTERN RECOGNITION, SPRINGER BERLIN HEIDELBERG, BERLIN, HEIDELBERG, PAGE(S) 175 - 185, ISBN: 978-3-642-33211-1.

[0003] There is a continuing need for a process that facilitates the analysis of images of patients' dental arches.

[0004] One goal of the invention is to address this need. Summary of the invention

[0005] The invention proposes a method according to claim 1 for analyzing an image, called an "analysis image", of a patient's dental arch, a method in which the analysis image is subjected to a deep learning device, preferably a neural network, in order to determine at least one value of a tooth attribute relative to a tooth represented on the analysis image, and / or at least one value of an image attribute relative to the analysis image. Analysis by tooth

[0006] The invention proposes in particular a method for detailed analysis of an image called an "analysis image" of a patient's dental arch, said method comprising the following steps: 1) creation of a training set comprising more than 1,000 images of dental arches, or "historical images", each historical image comprising one or more areas each representing a tooth, or "historical tooth areas", to each of which, for at least one tooth attribute, a tooth attribute value is assigned; 2) training at least one deep learning device, preferably a neural network, using the training set; 3) submitting the analysis image to said at least one deep learning device so that it determines at least one probability relative to an attribute value of at least one tooth represented on an area representing, at least partially, said tooth in the analysis image, or "analysis tooth area";4) determination, based on said probability, of the presence of a tooth of said arch at a position represented by said analysis tooth zone, and of the attribute value of said tooth. ;

[0007] A first deep learning device, preferably a neural network, can in particular be implemented to evaluate a probability relative to the presence, at a location of said analysis image, of an analysis tooth zone.

[0008] A second deep learning device, preferably a neural network, can be implemented in particular to evaluate a probability relative to the type of tooth represented in a tooth analysis area.

[0009] As will be seen in more detail later in the description, a detailed analysis method according to the invention advantageously allows the content of the analysis image to be recognized immediately.

[0010] The analyzed image can be advantageously classified automatically. Furthermore, it is immediately usable by a computer program.

[0011] The invention relies on the use of a deep learning device, preferably a neural network, whose performance is directly linked to the richness of the training set. Therefore, there is also a need for a method to rapidly enrich the training set.

[0012] The invention therefore relates to a method for enriching a training database, particularly intended for the implementation of a detailed analysis method according to the invention, said enrichment method comprising the following steps: A) at an "updated" time, creation of a model of a patient's dental arch, or "updated reference model", and segmentation of the updated reference model so as to create, for each tooth, a "tooth model", and for at least one tooth attribute, assignment of a tooth attribute value to each tooth model; B) preferably less than 6 months, preferably less than 2 months, preferably less than 1 month, preferably less than 15 days, preferably less than 1 week, preferably less than 1 day before or after the updated time, preferably substantially at the updated time, acquisition of at least one, preferably at least three, preferably at least ten, preferably at least one hundred images of said arch, or "updated images", under respective actual acquisition conditions;C) For each updated image, search for suitable virtual acquisition conditions for acquiring an image of the updated reference model, called the "reference image," showing maximum concordance with the updated image under said virtual acquisition conditions, and acquisition of said reference image; D) identification, in the reference image, of at least one area representing a tooth model, or "reference tooth area," and, by comparison of the updated image and the reference image, determination, in the updated image, of an area representing said tooth model, or "updated tooth area"; E) assignment, to said updated tooth area, of the tooth attribute value(s) of said tooth model; F) addition of the updated image enriched with a description of said updated tooth area and its tooth attribute value(s), or "historical image," to the training database.

[0013] In particular, each execution of the process described in WO 2016 / 066651 preferably generates more than three, more than ten, preferably more than one hundred updated images which, by automated processing using the updated reference model, can produce as many historical images.

[0014] In one particular embodiment, the process of enriching a training database comprises, instead of steps A) to C), the following steps: A') at an initial time, creation of a model of a patient's dental arch, or "initial reference model", and segmentation of the initial reference model so as to create, for each tooth, a "tooth model", and for at least one tooth attribute, assignment of a tooth attribute value to each tooth model; B') at an updated time, for example separated by more than fifteen days, preferably by more than one month, or even more than two months from the initial time, acquisition of at least one, preferably at least three, preferably at least ten, preferably at least one hundred images of said arch, or "updated images", under respective real acquisition conditions;(c) for each updated image, search, by deformation of the initial reference model, for an updated reference model and suitable virtual acquisition conditions for acquiring an image of the updated reference model, called the "reference image", showing maximum concordance with the updated image under said virtual acquisition conditions.

[0015] This process advantageously allows, after the initial reference model has been generated, preferably using a scanner, for the training database to be enriched at various updated time points, without the need for a new scan, and therefore without the patient having to travel to the orthodontist. The patient can, in fact, acquire the updated images themselves, as described in WO 2016 / 066651.

[0016] A single orthodontic treatment can thus lead to the production of hundreds of historical images.

[0017] The invention further relates to a method for training a deep learning device, preferably a neural network, comprising enriching a training database according to the invention, and then using said training database to train the deep learning device. Overall analysis

[0018] The detailed analysis process described above advantageously allows for a fine analysis of the analysis image, with the situation of each being preferably evaluated.

[0019] Alternatively, the deep learning system can be used globally, with the training set containing historical images whose descriptions provide a global attribute value for the image. In other words, the image attribute value relates to the entire image, not just a part of it. The attribute is then not a "tooth" attribute, but an "image" attribute. For example, this image attribute can define whether, with regard to the image as a whole or a part of the image, the dental situation is "pathological" or "not pathological," without examining each individual tooth. The image attribute also allows for the detection of, for example, whether the mouth is open or closed, or, more generally, whether the image is suitable for further processing, such as checking occlusion.

[0020] The image attribute can be specifically related to a position and / or orientation and / or calibration of an acquisition device used to acquire said analysis image, and / or quality of the analysis image, and in particular relating to the brightness, contrast or sharpness of the analysis image, and / or the content of the analysis image, for example the representation of the arches, tongue, mouth, lips, jaws, gums, one or more teeth or a dental appliance, preferably orthodontic.

[0021] When the image attribute refers to the image content, the description of the historical images in the training database specifies a characteristic of that content. For example, it may specify the position of the tongue (e.g., "recessed") or the opening of the patient's mouth (e.g., open or closed mouth) or the presence of a representation of a dental appliance, preferably orthodontic, and / or its condition (e.g., intact, broken, or damaged appliance).

[0022] A tooth attribute value can be used to define a value for an image attribute. For example, if a tooth attribute value is "cavity tooth", the image attribute value could be "unsatisfactory dental condition".

[0023] The image attribute can be particularly related to a therapeutic situation.

[0024] The invention proposes a method for the global analysis of an image of a patient's dental arch, said method comprising the following steps: 1') creation of a training set comprising more than 1,000 images of dental arches, or "historical images", each historical image having an attribute value for at least one image attribute, or "image attribute value"; 2') training at least one deep learning device, preferably a neural network, using the training set; 3') submitting the analysis image to the deep learning device so that it determines, for said analysis image, at least one probability relative to said image attribute value, and determining, as a function of said probability, a value for said image attribute for the analysis image.

[0025] The image attribute can be specifically related to the orientation of the acquisition device during the acquisition of the analysis image. For example, it can take the values ​​"front view," "left view," and "right view."

[0026] The image attribute can also relate to image quality. For example, it can take the values ​​"insufficient contrast" and "acceptable contrast".

[0027] The image attribute can also be related to the patient's dental situation, for example, related to the presence of a cavity or the condition of a dental appliance, preferably orthodontic, worn by the patient ("deteriorated" or "in good condition" for example) or to the suitability of the dental appliance, preferably orthodontic, to the patient's treatment (for example "unsuitable" or "suitable").

[0028] The image attribute can also relate to the "presence" or "absence" of a dental appliance, preferably orthodontic, or to the state of mouth opening ("open mouth", "closed mouth" for example).

[0029] As will be seen in more detail later in the description, a global image analysis method according to the invention advantageously allows for an immediate and comprehensive evaluation of the content of the analyzed image. In particular, it is possible to assess a dental situation globally and, for example, to deduce the need to consult an orthodontist. Definitions

[0030] A "patient" is a person for whom a process according to the invention is implemented, regardless of whether that person is undergoing orthodontic treatment or not.

[0031] The term "orthodontist" refers to any person qualified to provide dental care, which also includes a dentist.

[0032] The term "dental component", in particular "orthodontic component", refers to all or part of a dental appliance, especially an orthodontic one.

[0033] An orthodontic appliance can be, in particular, an orthodontic splint. Such a splint extends to follow the successive teeth of the dental arch to which it is attached. It defines a generally U-shaped channel, the shape of which is determined not only to ensure the splint's attachment to the teeth, but also according to a desired target positioning for the teeth. More precisely, the shape is determined so that, when the splint is in its functional position, it exerts forces that tend to move the treated teeth toward their target position, or to maintain the teeth in that target position.

[0034] The "service position" is the position in which the dental or orthodontic piece is worn by the patient.

[0035] By "model," we mean a three-dimensional digital model. An arrangement of tooth models is therefore a model.

[0036] By "image," we mean a two-dimensional image, such as a photograph or an image extracted from a film. An image is made up of pixels.

[0037] A "reference image" is a view of a "reference" model.

[0038] The term "image of an arch" or "model of an arch" means a representation of all or part of the said arch. Preferably, such a representation is in color.

[0039] The "acquisition conditions" of an image specify the position and orientation in space of an image acquisition device relative to the patient's teeth (actual acquisition conditions) or to a model of the patient's teeth (virtual acquisition conditions), and preferably the calibration of this acquisition device. Acquisition conditions are considered "virtual" when they correspond to a simulation in which the acquisition device would be in the said acquisition conditions (theoretical positioning and, preferably, calibration of the acquisition device) relative to a model.

[0040] In virtual acquisition conditions of a reference image, the acquisition device can also be described as "virtual". The reference image is in fact acquired by a fictitious acquisition device, having the characteristics of the "real" acquisition device used to acquire the real images, and in particular the updated images.

[0041] The calibration of a scanning device consists of all the values ​​of its calibration parameters. A calibration parameter is an intrinsic parameter of the scanning device (unlike its position and orientation) whose value influences the acquired image. Preferably, calibration parameters are chosen from the group formed by aperture, exposure time, focal length, and ISO sensitivity.

[0042] "Discriminatory information" is characteristic information that can be extracted from an image ( "image feature" ), typically through computer processing of this image.

[0043] Discriminatory information can have a variable number of values. For example, edge information can be 1 or 0 depending on whether a pixel belongs to a boundary or not. Brightness information can take on a large number of values. Image processing allows the extraction and quantification of discriminatory information.

[0044] Discriminatory information can be represented in the form of a "map". A map is thus the result of processing an image to reveal discriminatory information, for example the outline of teeth and gums.

[0045] This is called "concordance" ( " match " Or "fit" in(English) between two objects a measure of the difference between these two objects. A concordance is maximal (" best fit") when it results from an optimization that minimizes said difference.

[0046] An object modified to achieve maximum concordance can be described as an "optimal" object.

[0047] Two images or "views" that exhibit maximum concordance represent essentially at least one of the same teeth, in the same way. In other words, the representations of the tooth in these two images are virtually superimposable.

[0048] The search for a reference image with maximum concordance with an updated image is carried out by searching for the virtual acquisition conditions of the reference image with maximum concordance with the real acquisition conditions of the updated image.

[0049] By extension, a model has a maximum match with an image when that model has been chosen from among several models because it allows a view that has a maximum match with said image and / or when that image has been chosen from among several images because it has a maximum match with a view of said model.

[0050] In particular, an updated image is in maximum concordance with a reference model when a view of that reference model provides a reference image in maximum concordance with the updated image.

[0051] The comparison between two images preferably results from the comparison of two corresponding maps. A measure of the difference between two maps or between two images is classically called "distance".

[0052] "Metaheuristic" methods are well-known optimization methods. They are preferably chosen from the group formed by evolutionary algorithms, preferably chosen from: evolutionary strategies, genetic algorithms, differential evolution algorithms, distribution estimation algorithms, artificial immune systems, Shuffled Complex Evolution path recomposition, simulated annealing, ant colony algorithms, particle swarm optimization algorithms, tabu search, and the GRASP method; the kangaroo algorithm, the Fletcher and Powell method, the noise method, stochastic tunneling, random restart hill climbing, the cross-entropy method, and hybrid methods between the metaheuristic methods mentioned above.

[0053] The "description" of an image refers to information relating to the definition of the image's tooth zones and their associated tooth attribute values, and / or to an image attribute value. There is no limit to the number of possible values ​​for a tooth attribute or an image attribute.

[0054] A "historical" image is an image of a dental arch enhanced with a description. The tooth areas in a historical image are referred to as "historical tooth areas".

[0055] "Comprising" or "comprising" or "presenting" should be interpreted in a non-restrictive manner, unless otherwise indicated. Brief description of the figures

[0056] Other features and advantages of the invention will become apparent upon reading the detailed description that follows and examining the attached drawing in which: there figure 1represents, schematically, the different stages of a detailed image analysis process, according to the invention; the figure 2 represents, schematically, the different stages of a process for enriching a learning database, according to the invention; the figure 3 represents, schematically, the different stages of a variant of a method for enriching a training database, according to the invention; the figure 4 represents, schematically, the different stages of a global image analysis process, according to the invention; the figure 5 represents, schematically, the different stages of a stage C) of a method for enriching a training database, according to the invention; the figures 6 And 18 schematically represent the different stages of a process for modeling a patient's dental arch, according to the invention; the figure 7represents, schematically, the different stages of a process for evaluating a patient's dental situation, according to the invention; the figure 8 schematically represents the different stages of a process for acquiring an image of a patient's dental arch, according to the invention; the figure 9 schematically represents the different stages of a process for evaluating the shape of a patient's orthodontic aligner, according to the invention; Figure 10 represents an example of a reference image of an initial reference model; the figure 11 (11a-11d) illustrates a treatment for determining tooth patterns in an initial reference model, as described in WO 2016 066651; the figure 12 (12a-12d) illustrates the acquisition of an image using a retractor, an operation to crop this image, and the processing of an updated image to determine the contour of the teeth, as described in WO 2016 066651; the figure 13schematically illustrates the relative position of reference marks 12 of a spacer 10 on updated images 141 and 142, according to the directions of observation represented by dashed lines; the Figures 14 and 15 represent an orthodontic splint, in perspective and top view, respectively; the figure 16 illustrates step e) described in WO 2016 066651; the figure 17 illustrates an enrichment process according to the invention. Detailed description

[0057] A detailed analysis method according to the invention requires the creation of a training dataset. This creation preferably implements a method comprising steps A) to F), or, in one embodiment, instead of steps A) to C), preferably steps A') to C'). First principal embodiment of the enrichment process

[0058] Step A)is intended for the creation of an updated reference model representing a patient's arch. It preferably includes one or more of the features of step a) of WO 2016 066651 to create an initial reference model.

[0059] The updated reference model is preferably created with a 3D scanner. Such a model, called a "3D" model, can be viewed from any angle. An observation of the model, from a specific angle and distance, is called a "view" or "reference image".

[0060] There figure 11a is an example of a reference image.

[0061] The updated reference model can be prepared from measurements taken on the patient's teeth or on a mold of their teeth, for example a plaster cast.

[0062] For each tooth, a model of said tooth, or "tooth model" ( figure 11d). This operation, known in itself, is called "segmentation" of the updated reference model.

[0063] In the updated reference model, a tooth model is preferably delimited by a gingival margin which can be broken down into an internal gingival margin (on the inside of the mouth relative to the tooth), an external gingival margin (oriented towards the outside of the mouth relative to the tooth) and two lateral gingival margins.

[0064] One or more tooth attributes are associated with tooth models depending on the teeth they model.

[0065] The tooth attribute is preferably chosen from a tooth number, a tooth type, a tooth shape parameter (e.g., tooth width, especially mesio-palatal width, thickness, crown height, mesial and distal incisal edge deflection index, or abrasion level), a tooth appearance parameter (especially translucency index or color parameter), a parameter relating to the tooth condition (e.g., "abraded," "broken," "carious," or "braced"—i.e., in contact with a dental appliance, preferably orthodontic), the patient's age, or a combination of these attributes. A tooth attribute is preferably one that pertains only to the tooth modeled by the tooth model.

[0066] A tooth attribute value can be assigned to each tooth attribute of a particular tooth model.

[0067] For example, the tooth attribute "tooth type" will have the value "incisor", "canine" or "molar" depending on whether the tooth model is that of an incisor, a canine or a molar, respectively.

[0068] Assigning tooth attribute values ​​to tooth models can be done manually or, at least partially, automatically. For example, if the value of a tooth attribute is the same for all tooth models, such as the "patient age" tooth attribute, assigning a value to one tooth model may be sufficient to determine the value of that attribute for all other tooth models.

[0069] Similarly, tooth numbers are typically assigned according to a standard rule. Therefore, knowing this rule and the number of a tooth modeled by a tooth model is sufficient to calculate the numbers of the other tooth models.

[0070] In a preferred embodiment, the shape of a particular tooth model is analyzed to define its tooth attribute value, for example, its number. This shape recognition is preferably performed using a deep learning device, preferably a neural network. Preferably, a library of historical tooth models is created, each historical tooth model having a value for the tooth attribute, as described below (step a)). The deep learning device is trained with views of the historical tooth models from this library, and then one or more views of the particular tooth model are analyzed with the trained deep learning device to determine the tooth attribute value of said particular tooth model.

[0071] The assignment of tooth attribute values ​​can then be carried out entirely without human intervention.

[0072] Step B)is intended for the acquisition of one or preferably several updated images. Step B) preferably includes one or more of the features of step b) of WO 2016 066651.

[0073] The acquisition of updated images is carried out using an image acquisition device, preferably chosen from among a mobile phone, a so-called "connected" camera, a so-called "smartwatch", a tablet or a personal computer, desktop or laptop, equipped with an image acquisition system, such as a webcam or a camera. Preferably the image acquisition device is a mobile phone.

[0074] Preferably, the image acquisition device should be positioned more than 5 cm, 8 cm, or even 10 cm away from the dental arch. This prevents condensation of water vapor on the image acquisition device's optics and facilitates focusing. Furthermore, the image acquisition device, particularly a mobile phone, should ideally not be equipped with any specific optics for acquiring updated images. This is made possible, in particular, by the distance from the dental arch during image acquisition.

[0075] Preferably, an updated image is a photograph or a still from a film. It is preferably in color, preferably in true color.

[0076] Preferably, the acquisition of the updated image(s) is carried out by the patient, preferably without the use of a support, resting on the ground and immobilizing the image acquisition device, and in particular without a tripod.

[0077] In one embodiment, the triggering of the acquisition is automatic, i.e. without action by an operator, as soon as the acquisition conditions are approved by the image acquisition device, in particular when the image acquisition device has determined that it is observing a dental arch and / or a retractor and that the observation conditions are satisfactory (sharpness, brightness, or even dimensions of the representation of the dental arch and / or the retractor).

[0078] The time interval between steps A) and B) is kept to a minimum so that the teeth do not move significantly between the creation of the updated model and the acquisition of the updated images. Reference images that are consistent with the updated images can then be acquired by observing the updated reference model.

[0079] Preferably, a 10mm dental retractor is used during step B), as shown in the figure 12a The retractor classically consists of a support with a rim extending around an opening and arranged so that the patient's lips can rest on it while allowing the patient's teeth to appear through said opening.

[0080] In step C), The updated reference model is explored to find, for each updated image, a reference image with maximum concordance to the updated image.

[0081] Step C) may include one or more of the features of steps c), d) and e) of WO 2016 066651, insofar as they relate to such exploration.

[0082] For each updated image, a set of virtual acquisition conditions is preferably determined, in a rough manner, to approximate the actual acquisition conditions during the acquisition of that updated image. In other words, the position of the image acquisition device relative to the teeth at the moment it acquired the updated image is estimated (the device's position in space and its orientation). This rough evaluation advantageously limits the number of tests on virtual acquisition conditions during subsequent operations, thus significantly accelerating these operations.

[0083] To perform this rough assessment, one or more heuristic rules are preferably used. For example, it is preferable to exclude from the virtual acquisition conditions that could be tested during subsequent operations those conditions which correspond to a position of the image acquisition device behind the teeth or at a distance from the teeth greater than 1 m.

[0084] In a preferred embodiment, as illustrated in the figure 13 , we use reference marks shown on the updated image, and in particular reference marks 12 of the spreader, to determine a substantially conical region of space delimiting virtual acquisition conditions that can be tested during the following operations, or "test cone".

[0085] More specifically, we preferably have at least three non-aligned reference marks 12 on the spacer 10, and we precisely measure their relative positions on the spacer.

[0086] The reference marks are then located on the updated image, as described previously. Simple trigonometric calculations allow us to determine approximately the direction in which the updated image was taken.

[0087] Next, for each updated image, a reference image is sought that exhibits the highest possible match with the updated image. This search is preferably performed using a metaheuristic method, preferably evolutionary, preferably by simulated annealing.

[0088] Preferably, at some point before step C4, the updated image is analyzed to produce an updated map representing, at least partially, discriminating information. The updated map thus represents the discriminating information within the reference frame of the updated image.

[0089] The discriminating information is preferably chosen from the group consisting of contour information, color information, density information, distance information, brightness information, saturation information, reflection information, and combinations of these.

[0090] The expert knows how to process an updated image to reveal the discriminating information.

[0091] For example, the figure 12d is an updated map relating to the contour of the teeth obtained from the updated image of the figure 12b .

[0092] The research in question comprises the following steps: C1) determination of virtual acquisition conditions "to be tested"; C2) creation of a reference image of the updated reference model under said virtual acquisition conditions to be tested; C3) processing of the reference image to create at least one reference map representing, at least partially, the discriminating information; C4) comparison of the updated and reference maps so as to determine a value for an evaluation function, said value for the evaluation function depending on the differences between said updated and reference maps and corresponding to a decision to continue or stop the search for virtual acquisition conditions approximating said real acquisition conditions more accurately than said virtual acquisition conditions to be tested determined at the last occurrence of step C1);C5) if said value for the evaluation function corresponds to a decision to continue said research, modification of the virtual acquisition conditions to be tested, then resumed at step C2). ;

[0093] At step C1), We begin by determining virtual acquisition conditions to be tested, that is to say a virtual position and orientation likely to correspond to the real position and orientation of the acquisition device when capturing the updated image, but also, preferably, a virtual calibration likely to correspond to the real calibration of the acquisition device when capturing the updated image.

[0094] The first virtual acquisition conditions to be tested are preferably roughly evaluated virtual acquisition conditions, as described previously.

[0095] At step C2),The image acquisition device is then virtually configured under the virtual acquisition conditions to be tested in order to acquire a reference image of the updated reference model under these virtual acquisition conditions. The reference image therefore corresponds to the image that the image acquisition device would have taken if it had been placed, relative to the updated reference model, and optionally calibrated, under the virtual acquisition conditions to be tested.

[0096] If the updated image was acquired at approximately the same time as the updated reference model was created by scanning the patient's teeth, the position of the teeth in the updated image is virtually identical to that in the updated reference model. If the virtual acquisition conditions being tested are exactly the same as the actual acquisition conditions, then the reference image is exactly superimposable on the updated image. Differences between the updated image and the reference image result from errors in evaluating the virtual acquisition conditions being tested, if they do not exactly match the actual acquisition conditions.

[0097] At step C3), we process the reference image, like the updated image, in such a way as to create, from the reference image, a reference map representing the discriminating information ( figures 11a and 11bThe expert knows how to process a reference image to bring out the discriminating information.

[0098] At step C4), The updated and reference maps, both depicting the same discriminating information, are compared, and the difference, or "distance," between the two maps is evaluated using a score. For example, if the discriminating information is the outline of teeth, the average distance between points on the tooth outline appearing in the reference image and points on the corresponding outline appearing in the updated image can be compared. The smaller this distance, the higher the score. The score could be, for example, a correlation coefficient.

[0099] Preferably, the virtual acquisition conditions include the calibration parameters of the acquisition device. The score is higher the closer the tested calibration parameter values ​​are to the calibration parameter values ​​of the acquisition device used to acquire the updated image. For example, if the tested aperture is far from that of the acquisition device used to acquire the updated image, the reference image will have blurred and sharp regions that do not correspond to the blurred and sharp regions of the updated image. If the discriminating information is the contour of the teeth, the updated and reference maps will therefore not represent the same contours, and the score will be low.

[0100] The score is then evaluated using an evaluation function. This function determines whether the cycling process on steps C1) to C5) should continue or stop. For example, the evaluation function could be equal to 0 if the cycling should stop, or equal to 1 if the cycling should continue.

[0101] The value of the evaluation function can depend on the score achieved. For example, it may be decided to continue the cycle if the score does not exceed a threshold. For instance, if an exact match between the updated and reference images results in a score of 100%, the threshold could be, for example, 95%. Naturally, the higher the threshold, the better the accuracy of the evaluation of the virtual acquisition conditions if the score manages to exceed that threshold.

[0102] The value of the evaluation function may also depend on scores obtained with previously tested virtual acquisition conditions.

[0103] The value of the evaluation function can also depend on random parameters and / or the number of cycles already performed.

[0104] In particular, it is possible that despite repeated cycles, virtual acquisition conditions sufficiently close to actual acquisition conditions may not be found for the score to reach the threshold. The evaluation function may then lead to the decision to exit the cycle even though the best score obtained did not reach the threshold. This decision may result, for example, from exceeding a predetermined maximum number of cycles.

[0105] A random parameter in the evaluation function can also allow further testing of new virtual acquisition conditions, even if the score appears satisfactory.

[0106] Evaluation functions classically used in metaheuristic optimization processes, preferably evolutionary, particularly in simulated annealing processes, can be used for the evaluation function.

[0107] At the stage C5), if the value of the evaluation function indicates that it is decided to continue the cycling, the virtual acquisition conditions to be tested are modified and the cycling is restarted on steps C1) to C5) consisting of producing a reference image and a reference map, comparing the reference map with the updated map to determine a score, and then making a decision based on this score.

[0108] The modification of the virtual acquisition conditions to be tested corresponds to a virtual movement in space and / or a change in orientation and / or, preferably, a change in the calibration of the acquisition device. This modification can be random, preferably so that the new virtual acquisition conditions to be tested always belong to the set determined during the rough evaluation. The modification is preferably guided by heuristic rules, for example, by favoring modifications that, according to an analysis of previous scores obtained, appear most likely to increase the score.

[0109] The cycling is continued until the value of the evaluation function indicates that it is decided to stop this cycling and continue to step D), for example if the score reaches or exceeds said threshold.

[0110] The optimization of virtual acquisition conditions is preferably performed using a metaheuristic method, preferably an evolutionary one, and preferably a simulated annealing algorithm. Such an algorithm is well known for nonlinear optimization.

[0111] If the cycling process is stopped without a satisfactory score being obtained, for example, without the score reaching the specified threshold, the process can be stopped (failure situation) or a new step (C) can be launched with new discriminating information and / or a new, updated image. The process can also be continued with the virtual acquisition conditions corresponding to the best score achieved. A warning can be issued to inform the user of the error in the result.

[0112] If we stopped the cycling process when a satisfactory score could be obtained, for example because the score reached or even exceeded the said threshold, the virtual acquisition conditions correspond substantially to the real acquisition conditions of the updated image.

[0113] Preferably, the virtual acquisition conditions include the calibration parameters of the acquisition device. This allows the values ​​of these parameters to be evaluated without needing to know the type of acquisition device or its settings. The acquisition of updated images can therefore be carried out without any special precautions, for example, by the patient themselves using their mobile phone.

[0114] Furthermore, the search for the actual calibration is carried out by comparing an updated image with views of a reference model under virtual acquisition conditions to be tested. Advantageously, it does not require the updated image to show a calibration gauge, that is, a gauge whose characteristics are precisely known to determine the calibration of the acquisition device.

[0115] Step C) therefore results in the determination of virtual acquisition conditions that exhibit maximum agreement with the actual acquisition conditions. The reference image is thus in maximum agreement with the updated image; that is, these two images are virtually superimposable.

[0116] In one embodiment, said search for virtual acquisition conditions in step C) is carried out using a metaheuristic method, preferably evolutionary, preferably a simulated annealing algorithm.

[0117] At step D), We identify the reference tooth areas on the reference image and transfer them to the updated image to define corresponding updated tooth areas.

[0118] In particular, the reference image is a view of the updated reference model segmented into tooth models. The boundaries of the representation of each tooth model on the reference image, or "reference tooth area", can therefore be identified.

[0119] Superimposing the updated and reference images then allows the boundaries of the reference tooth zones to be transferred onto the updated image, thus defining the updated tooth zones. Since the reference image is in maximum agreement with the updated image, the updated tooth zones essentially define the boundaries of the tooth models represented in the reference image.

[0120] At step E), For each updated tooth area, we assign the tooth attribute value(s) from the corresponding tooth model.

[0121] Specifically, the reference image is a view of the updated reference model in which the tooth models have been assigned respective tooth attribute values ​​for at least one tooth attribute, such as a tooth number. Each reference tooth area can therefore inherit the tooth attribute value of the tooth model it represents. Each updated tooth area can then inherit the tooth attribute value from the reference tooth area that defined it.

[0122] At the end of step E), we therefore obtain an updated image and a description of the updated image defining one or more updated tooth zones and, for each of these zones, a tooth attribute value for at least one tooth attribute, for example a tooth number.

[0123] The term "historical image" refers to an updated image enriched with its description.

[0124] There figure 17ashows an example of an updated image (acquired during step B)) being analyzed to determine the contours of the teeth. figure 17b shows the reference image with the greatest possible match to the updated image (resulting from step C)). The tooth numbers are displayed on the corresponding teeth. The figure 17c illustrates the transfer of tooth numbers onto the updated tooth areas (steps D) and E)).

[0125] At step F), The historical image is added to the learning database.

[0126] Steps A) to F) are preferably performed for more than 1,000, more than 5,000, or more than 10,000 different patients, or "historical patients".

[0127] As is now clear, the invention provides a particularly effective method for creating a learning base.

[0128] The invention also relates to a method for training a deep learning device, preferably a neural network, said method comprising enriching a training set by means of a process comprising steps A) to F) so as to acquire a plurality of historical images, and then using said training set to train said deep learning device. Second main embodiment of the enrichment process

[0129] The invention is not, however, limited to the embodiments described above.

[0130] In particular, the updated reference model is not necessarily the direct result of a scan of the patient's dental arch. The updated reference model may, in particular, be a model obtained by deforming an initial reference model that itself resulted directly from such a scan.

[0131] The process then preferably includes, instead of steps A) to C), steps A') to C').

[0132] The stage A') is identical to step A). ​​In step A'), however, the generated reference model is intended to be modified. It is therefore referred to as the "initial reference model," and not the "updated reference model," as in step A).

[0133] The initial reference model can be generated at a specific point in time prior to active orthodontic treatment, for example, less than 6 months, less than 3 months, or less than 1 month before the start of treatment. Steps B') to C') can then be implemented to track the progress of treatment between the initial point in time and the updated point in step B').

[0134] The initial time point can alternatively be a point at the end of active orthodontic treatment, for example, less than 6 months, less than 3 months, or less than 1 month after the end of treatment. Steps B') to C') can then be implemented to monitor for the appearance of any relapse.

[0135] Step B') is identical to step B). In step B'), however, the updated images are also intended to guide the modification of the initial reference model to define the updated reference model, in step C').

[0136] The time interval between steps A') and B') is not limited, since, as explained below, the initial reference model will be deformed to obtain an updated reference model with maximum agreement with the updated images. The time interval between steps A') and B') can be, for example, greater than 1 week, 2 weeks, 1 month, 2 months, or 6 months.

[0137] Step C') is more complex than step C) since the search for a reference image with maximum concordance to an updated image is not limited to finding the optimal virtual acquisition conditions. It also includes finding an updated reference model, that is, a reference model in which the teeth have essentially the same position as in the updated image.

[0138] Step C') preferably includes one or more of the features of steps c), d) and e) of WO 2016 066651, and in particular of step e) illustrated in the figure 16 .

[0139] The goal is to modify the initial reference model until an updated reference model is obtained that shows maximum concordance with the updated image. Ideally, the updated reference model is therefore a model of the arcade from which the updated image could have been taken if that model had been the arcade itself.

[0140] We therefore test a succession of reference models "to be tested", the choice of a reference model to be tested preferably depending on the level of correspondence of the reference models "to be tested" previously tested with the updated image.

[0141] Preferably, the search includes, for an up-to-date image, a first optimization operation allowing to search, in a reference model to be tested determined from the initial reference model, for virtual acquisition conditions that best correspond to the real acquisition conditions of the updated image, and a second optimization operation allowing to search, by testing a plurality of said reference models to be tested, for the reference model that best corresponds to the positioning of the patient's teeth during the acquisition of the updated image.

[0142] Preferably, a first optimization operation is performed for each test of a reference model to be tested during the second optimization operation.

[0143] Preferably, the first optimization operation and / or the second optimization operation, preferably the first optimization operation and the second optimization operation implement a metaheuristic method, preferably evolutionary, preferably a simulated annealing.

[0144] Step C') therefore leads to the determination of an updated reference model showing maximum concordance with the updated image, and of virtual acquisition conditions showing maximum concordance with the real acquisition conditions.

[0145] A process comprising steps A') to C') can be advantageously implemented in the context of active or passive orthodontic treatment, or, more generally, to monitor any development of the teeth.

[0146] In the various methods according to the invention, the enrichment of the learning base does not necessarily result from an enrichment method according to the invention.

[0147] In one embodiment, the training set is created by an operator. This operator analyzes thousands of analysis images. To enable the training set to be used for a detailed analysis process, the operator identifies tooth regions and assigns them tooth attribute values. To enable the training set to be used for a global analysis process, the operator assigns image attribute values ​​to each image. This allows the operator to create historical images. Detailed image analysis method

[0148] The detailed analysis method of an "analysis image" of a patient's dental arch according to the invention comprises steps 1) to 4).

[0149] Preferably, the analysis image, preferably a photograph or an image extracted from a film, preferably in color, preferably in true color, is acquired with an image acquisition device, preferably a mobile phone, held away from the dental arch by more than 5 cm, more than 8 cm, or even more than 10 cm, and which preferably is not equipped with any specific optics.

[0150] Preferably, the analysis image represents several teeth, preferably more than 2, more than 3, more than 4 or more than 5 of the patient's teeth.

[0151] There figure 12a or the figure 12b These could be examples of analysis images. The arrangement of the teeth is realistic, that is to say, it corresponds to that observed by the image acquisition device when it acquired the analysis image.

[0152] Preferably, the acquisition of the analysis image is carried out by the patient, preferably without the use of a support, resting on the ground and immobilizing the image acquisition device, and in particular without a tripod.

[0153] In one embodiment, the triggering of the acquisition of the analysis image is automatic, i.e. without action by an operator, as soon as the acquisition conditions are approved by the image acquisition device, in particular when the image acquisition device has determined that it is observing a dental arch and / or a retractor and that the observation conditions are satisfactory (sharpness, brightness, or even dimensions of the representation of the dental arch and / or the retractor).

[0154] In step 1),We create a training dataset containing more than 1,000, preferably more than 5,000, preferably more than 10,000, preferably more than 30,000, preferably more than 50,000, preferably more than 100,000 historical images. The larger the number of historical images, the better the analysis performed by the process.

[0155] Preferably, an enriched training database is used, following an enrichment process according to the invention.

[0156] The training set can, however, be created using other methods, for example, manually. To create a historical image of the training set, an operator, preferably an orthodontist, identifies one or more "historical" tooth areas on an image, then assigns, to each identified historical tooth area, a value for at least one tooth attribute.

[0157] In step 2),a deep learning device, preferably a neural network, is trained with the training set.

[0158] A "neural network" or "artificial neural network" is a set of algorithms well known to those skilled in the art.

[0159] The neural network can be specifically chosen from: networks specialized in image classification, called “CNN” (“Convolutional neural network”), for example AlexNet (2012) ZF Net (2013) VGG Net (2014) GoogleNet (2015) Microsoft ResNet (2015) Caffe: BAIR Reference CaffeNet, BAIR AlexNet Torch:VGG_CNN_S,VGG_CNN_ M,VGG_CNN_M2048,VGG_CNN_M_1 0 24,VGG_CNN_M_128,VGG_CNN_F,VGG ILSVRC-2014 16-layer,VGG ILSVRC-2014 19-layer,Network-in-Network (Imagenet & CIFAR-10) Google: Inception (V3, V4). Specialized networks for localization and detection of objects in an image, the Object Detection Network, for example: R-CNN (2013) SSD (Single Shot MultiBox Detector: Object Detection network), Faster R-CNN (Faster Region-based Convolutional Network method: Object Detection network) Faster R-CNN (2015) SSD (2015).

[0160] The above list is not exhaustive.

[0161] In step 2), the deep learning device is preferably trained by a learning process called "Deep learning". By presenting historical images (images + descriptions) as input to the deep learning system, the deep learning system gradually learns to recognize patterns in an image, in English. "patterns", and to associate them with tooth areas and tooth attribute values, for example tooth numbers.

[0162] In step 3), The image that we wish to analyze, or "analysis image", is submitted to the deep learning device.

[0163] Thanks to its training in step 2), the deep learning system is able to analyze the analysis image and recognize these patterns. In particular, it can determine a relative probability to: the presence, at a location in said analysis image, of an area representing, at least partially, a tooth, or "analysis tooth area", the attribute value of the tooth represented on said analysis tooth area.

[0164] For example, it is able to determine that there is a 95% chance that a shape in the analysis image represents an incisor.

[0165] Preferably, the deep learning device analyzes the entire scan image and determines probabilities for all scan tooth areas it has identified.

[0166] In step 4), We analyze the results of the previous step to determine the teeth represented in the analysis image.

[0167] When the training set contains more than 10,000 historical images, step 3) leads to particularly satisfactory results. In particular, such a training set makes it possible to establish a probability threshold such that if a probability associated with an analysis tooth area and a tooth attribute value for that analysis tooth area exceeds said threshold, the analysis tooth area effectively represents a tooth with said tooth attribute value.

[0168] Step 4) thus leads to the definition of an analysis image enriched with a description defining the analysis tooth zones and, for each analysis tooth zone, the values ​​of the attributes of the tooth represented by the analysis tooth zone. Global image analysis method

[0169] The method for global analysis of an updated image of a patient's dental arch according to the invention comprises steps 1') to 3').

[0170] The process is similar to the detailed analysis process described above, except that, according to the global analysis, it is not necessary to analyze the individual situation of each tooth. The analysis is global to the entire image. In other words, the deep learning system determines the value of an "image" attribute without having to first determine tooth attribute values.

[0171] For example, the deep learning device may conclude that, "overall", the dental situation is "satisfactory" or "unsatisfactory", without determining which tooth might be the cause of the dissatisfaction.

[0172] Step 1') is similar to step 1). However, historical images include a description specifying an image attribute value for each image.

[0173] Step 2') is similar to step 2).

[0174] At step 3'),the analysis image is submitted to the deep learning system.

[0175] Thanks to its training in step 2'), the deep learning system is able to analyze the image and recognize these patterns. Based on these patterns, it can, in particular, determine a probability relative to the value of the image attribute under consideration. Application to the modeling of a dental arch

[0176] A detailed analysis method according to the invention is particularly useful for modeling a dental arch, especially for establishing a remote diagnosis.

[0177] It is advisable for everyone to have their teeth checked regularly, particularly to ensure that their position is not changing unfavorably. During orthodontic treatment, this unfavorable change may lead to modifications of the treatment plan. After orthodontic treatment, an unfavorable change, known as "relapse," may necessitate further treatment. Finally, more generally and independently of any treatment, everyone may wish to monitor any potential shifting of their teeth.

[0178] Typically, checkups are performed by an orthodontist who uses specialized equipment. These checkups are therefore expensive. Furthermore, the appointments are inconvenient. Finally, some people are apprehensive about visiting an orthodontist and will forgo making an appointment for a simple checkup or to assess the feasibility of orthodontic treatment.

[0179] US 2009 / 0291417 describes a process for creating and modifying three-dimensional models, particularly for the manufacture of orthodontic appliances.

[0180] WO 2016 066651 describes a method for checking the positioning and / or shape and / or appearance of a patient's teeth. This method includes a step of creating an initial reference model of the teeth at a given time, preferably using a 3D scanner, and then, at a later time, or "update time," for example, six months after the initial time, creating an updated reference model by deforming the initial reference model. This deformation is carried out in such a way that the updated reference model allows observations substantially identical to images of the teeth acquired at the updated time, in particular to photographs or video footage taken by the patient themselves without special precautions, referred to as "update images."

[0181] The updated images are therefore used to modify the initial, highly accurate reference model. The updated reference model resulting from the deformation of the initial reference model, guided by the analysis of the updated images, is therefore also highly accurate.

[0182] The procedure described in WO 2016 / 66651, however, requires an appointment with an orthodontist to create the initial reference model. This appointment hinders preventative care. Indeed, a patient will not necessarily consult an orthodontist if they do not perceive it as necessary. In other words, the procedure is often only implemented when a malocclusion is observed and needs correction.

[0183] Therefore, there is a need for a process that addresses this problem by facilitating prevention.

[0184] One goal of the invention is to address this need.

[0185] To this end, the invention proposes a method for modeling a patient's dental arch, said method comprising the following steps: a) creation of a historical library containing more than 1,000 tooth models, referred to as "historical tooth models", and assignment to each historical tooth model of a value for at least one tooth attribute, or "tooth attribute value"; b) analysis of at least one "analysis image" of the dental arch according to a detailed analysis method according to the invention, so as to determine at least one analysis tooth area and at least one tooth attribute value associated with said analysis tooth area; c) for each analysis tooth area determined in the previous step, search, in the historical library, for a historical tooth model exhibiting maximum proximity to the analysis image or to the analysis tooth area, or "optimal tooth model"; d) arrangement of all the optimal tooth models so as to create a model that exhibits maximum concordance with the updated image, or "assembled model";e) Optionally, replace at least one optimal tooth model with another historical tooth model and repeat step d) in order to maximize the concordance between the assembled model and the analysis image; f) Optionally, repeat step b) with another analysis image and in step d) and / or e), search for maximum concordance with all the analysis images used.

[0186] The invention thus makes it possible, from a simple analysis image, for example a photograph taken using a mobile phone, to reconstruct, with good reliability, a dental arch in the form of an assembled model.

[0187] The analysis image can be acquired in particular as described in step 1) above.

[0188] Of course, analyzing a single scan image is not enough to generate an assembled model that precisely matches the patient's tooth arrangement. However, such precision is generally not essential for making an initial diagnosis of the patient's dental situation.

[0189] Furthermore, the accuracy of the assembled model can be increased if multiple analysis images are processed.

[0190] Steps b) to c) are preferably implemented for several analysis images and, in steps d) and e), optimal tooth models and an assembled model are sought to obtain maximum concordance with respect to the set of analysis images (step f)).

[0191] The invention also relates to a method for evaluating a patient's dental situation, comprising the following steps: i) creation of an assembled model according to a modeling process according to the invention; ii) transmission of the assembled model to a recipient, preferably an orthodontist and / or a computer; iii) analysis of the patient's orthodontic situation, by the recipient, from the assembled model; iv) preferably, informing the patient of the orthodontic situation, preferably via his mobile phone.

[0192] The patient can therefore very easily ask an orthodontist to check their dental situation, without even having to travel, simply by sending one or preferably several photos of their teeth.

[0193] A modeling process is now described in detail.

[0194] In step a), we create a historical library 20 ( figure 18) comprising more than 1,000, preferably more than 5,000, preferably more than 10,000 historical tooth models 22. The higher the number of historical tooth models, the more accurate the assembled model.

[0195] A historical tooth model can be obtained, in particular, from a model of a dental arch of a "historical" patient obtained with a scanner. This arch model can be segmented to isolate the representations of the teeth, as in the figure 11d Each of these representations, featuring a specific shade of grey on the figure 11d , can constitute a model of a historical tooth.

[0196] Preferably, the library is enriched with the tooth models resulting from the implementation of the process described in WO 2016 066651 or from step A) or A') described above.

[0197] One or more tooth attributes, specifically chosen from the list provided above, are associated with the tooth models. A tooth attribute value is assigned to each tooth attribute of a particular tooth model, as described previously (see the description of step A). ​​For example, a tooth model is that of an "incisor," "heavily worn," and whose color parameters are, in the L*a*b* color system according to standard NF ISO 7724, "a*=2," "b*=1," and "L*=58."

[0198] The historical library therefore contains historical tooth models and associated attribute values ​​that facilitate searching in step c). On the figure 18 , only 22 historical tooth models representing molars were represented in the historical library 20.

[0199] In step b),The analysis image is acquired, as described above for step B), before being analyzed. In particular, the analysis image is preferably a photograph or a still from a video, preferably taken with a mobile phone.

[0200] The analysis image can be acquired at any time after step a), for example more than 1 week, more than 1 month or more than 6 months after step a).

[0201] The analysis image is analyzed according to a detailed analysis method according to the invention. The optional features of this method are also optional in step b).

[0202] At the end of step b), we obtain an analysis image enriched with a description providing, for each analysis tooth area, a tooth attribute value for at least one tooth attribute, for example a tooth number.

[0203] In step c),In the historical library, for each analysis tooth zone determined in the previous step, we search for a historical tooth model that presents a maximum proximity to the analysis tooth zone. This tooth model is called the "optimal tooth model".

[0204] "Proximity" is a measure of one or more differences between the historical tooth model and the analysis tooth area. These differences can include a difference in shape, but also other differences such as a difference in translucency or color. Maximum proximity can be sought by successively minimizing several differences, or by minimizing a combination of these differences, for example, a weighted sum of these differences.

[0205] "Proximity" is therefore a broader concept than "concordance," concordance only measuring proximity relative to form.

[0206] Evaluating the similarity of a historical tooth model to an analysis tooth area preferably involves comparing at least one value of a tooth attribute in the analysis tooth area with the value of that attribute in the historical tooth model. Such an evaluation is advantageously very fast.

[0207] For example, if the description of the analysis tooth area provides a value for the tooth type or number, the thickness of the represented tooth and / or the height of its crown and / or its mesio-palatal width and / or the mesial and distal deflection index of its incisal edge, this value can be compared to the value of the corresponding attribute of each of the historical tooth models.

[0208] Preferably, we look for a historical tooth model that has, for at least one tooth attribute, the same value as the aforementioned analysis tooth area. The tooth attribute can, in particular, relate to the tooth type or the tooth number. In other words, we filter the historical tooth models to examine in more detail only those that relate to the same tooth type as the tooth represented in the analysis tooth area.

[0209] Alternatively, or preferably in addition to this comparison of attribute values, the shape of the tooth represented on the analysis tooth area can be compared to the shape of a historical tooth model to be evaluated, preferably using a metaheuristic method, preferably evolutionary, preferably by simulated annealing.

[0210] To this end, the historical tooth model to be evaluated is observed from different angles. Each view thus obtained is compared with the analysis image, preferably with the analysis tooth area, in order to establish a "distance" between this view and the analysis image or, preferably, the analysis tooth area. The distance thus measures the difference between the view and the analysis tooth area.

[0211] The distance can be determined after processing the view and the analysis image or, preferably, the analysis tooth area, so as to make the same discriminating information appear on corresponding maps, for example contour information, as described above in step C3) or in WO 2016 066651.

[0212] For each historical tooth model tested, a view is determined that provides a minimum distance from the analysis image or the analysis tooth area. Each historical tooth model examined is thus associated with a specific minimum distance, which measures its shape proximity to the analysis tooth area.

[0213] The optimal historical tooth model is the one which, with regard to the comparison(s) made, is considered to be closest to the analysis tooth area.

[0214] The minimum distances obtained for the different tested tooth models are then compared, and the model with the smallest minimum distance is selected to define the optimal tooth model. The optimal tooth model therefore exhibits maximum concordance with the analysis image.

[0215] The search for maximum concordance is preferably carried out using a metaheuristic method, preferably evolutionary, preferably by simulated annealing.

[0216] In a preferred embodiment, a first evaluation of historical tooth models is performed successively by comparing the values ​​of at least one tooth attribute, for example, the tooth number, with the corresponding values ​​in the analysis tooth area, followed by a second evaluation by shape comparison. The first, faster evaluation advantageously filters the historical tooth models so that only those models retained by the first evaluation are submitted to the second, slower evaluation.

[0217] For example, if a tooth analysis area represents tooth #15, the first evaluation allows us to retain only the tooth models modeling teeth #15. During the second evaluation, we search, among the set of historical tooth models modeling teeth #15, for the historical tooth model whose shape most closely resembles that of the represented tooth.

[0218] Preferably, several initial assessments are carried out before the second assessment. For example, the initial assessments allow for filtering historical tooth models to retain only those modeling teeth #15 with a crown height between 8 and 8.5 mm.

[0219] At the end of step c), an optimal tooth model was thus associated with each of the analysis tooth zones.

[0220] For example, on the figure 18The historical tooth model 22 1 can be observed to closely resemble an analysis area identified on the analysis image. It is considered optimal for this analysis area.

[0221] In step d), We create an assembled model by arranging the optimal tooth models.

[0222] According to one embodiment, at the beginning of step d), a first rough arrangement is created, that is to say, a rough model is made by assembling the optimal tooth models.

[0223] To establish the first rough arrangement, we can orient the optimal tooth models so that their optimal observation directions are all parallel, the optimal observation direction of a tooth model being the direction in which said tooth model has maximum concordance with the analysis image.

[0224] The first rough arrangement can also be established by considering the tooth attribute values ​​of the optimal tooth models. For example, if the tooth numbers of the optimal tooth models are those of the canines and incisors, these tooth models can be arranged along an arc 24 ( figure 18 ) classically corresponding to the region of the arch which bears these types of teeth.

[0225] The shape of this arc can be refined based on other tooth attribute values.

[0226] The order of the optimal tooth models is that of the corresponding analysis tooth zones.

[0227] Furthermore, the minimum distance associated with an optimal tooth model results from observing the tooth model along an "optimal" observation direction. In other words, the tooth that this model represents is most likely also observed in the analysis image along this direction. All optimal tooth models are therefore preferably oriented so that their respective optimal observation directions are all parallel.

[0228] It is therefore possible to define a first arrangement of optimal tooth models.

[0229] Preferably, the first arrangement of optimal tooth models is then iteratively modified so as to present maximum concordance with the analysis image.

[0230] To evaluate an arrangement, it is observed from different angles. Each view thus obtained is compared with the analysis image in order to establish a "distance" between this view and the analysis image. The distance thus measures the difference between the view and the analysis image.

[0231] The distance can be determined after processing the view and the analysis image so as to make discriminating information appear on one of the corresponding maps, for example contour information, as described above in step C3) or in WO 2016 066651.

[0232] For each arrangement examined, a view providing a minimum distance to the analysis image is thus determined. Each arrangement examined is therefore associated with a minimum distance.

[0233] The minimum distances obtained for the different tested arrangements are then compared, and the arrangement with the smallest minimum distance is selected to define the optimal arrangement. The optimal arrangement therefore exhibits the greatest possible match with the analysis image.

[0234] The search for maximum concordance is preferably carried out using a metaheuristic method, preferably evolutionary, preferably by simulated annealing.

[0235] At the end of step d), we obtain an optimal arrangement of the optimal tooth models, that is to say the assembled model 26.

[0236] At step e) Optionally, one or more optimal tooth models are replaced with other tooth models, and then we start again at step d) in order to maximize the concordance between the assembled model and the analysis image.

[0237] It is indeed possible that an optimal tooth model, in the "optimal" arrangement, no longer exhibits maximum concordance with the analysis image. Specifically, the tooth model might have been observed in an "optimal" direction that provided a view with minimal distance from the analysis image (which is why it was considered optimal). However, in the optimal arrangement, it is no longer oriented along the optimal direction.

[0238] A new search for an assembled model can therefore be carried out by modifying the tooth models, for example by replacing the optimal tooth models with close tooth models.

[0239] The search for tooth models to be tested is preferably carried out using a metaheuristic method, preferably evolutionary, preferably by simulated annealing.

[0240] In the preferred embodiment, the process therefore implements a double optimization, on the tooth models and on the arrangement of the tooth models, the assembled model being the arrangement of a set of tooth models that provides the minimum distance with the analysis image, considering all possible tooth models and all possible arrangements.

[0241] At step f), Optional and preferred, the method uses multiple analysis images of the patient's dental arch, preferably more than 3, 5, 10, 50, or even more than 100. The resulting assembled model is thus more complete. Even more preferably, the method implements optimization so that the resulting assembled model is optimal with respect to all the analysis images. In other words, the assembled model is preferably the one that maximizes agreement with all the analysis images.

[0242] In the preferred embodiment, the process therefore implements a double or, preferably a triple optimization, on the tooth models on the one hand, on the arrangement of the tooth models and / or on a plurality of analysis images on the other hand, the assembled model being the arrangement of a set of tooth models which provides the average minimum distance, on the set of analysis images, considering all possible tooth models and, preferably, all possible arrangements.

[0243] According to one embodiment, in step d) and / or e) and / or f), a metaheuristic method is used, preferably evolutionary, preferably by simulated annealing.

[0244] As is now clear, the invention makes it possible to construct an assembled model of a dental arch from simple analytical images, such as photographs taken with a mobile phone. Of course, the precision of the assembled model does not reach that of a scan. However, in certain applications, for example, for making an initial diagnosis of a patient's dental condition, such precision is not essential.

[0245] The assembled model can therefore be used to analyze the patient's orthodontic situation, following steps ii) to iv).

[0246] In step ii), The assembled model is sent to an orthodontist and / or to a computer equipped with diagnostic software.

[0247] In one embodiment, the assembled model is sent along with a questionnaire completed by the patient in order to improve the quality of the analysis in step iv).

[0248] In step iii),The orthodontist and / or the computer examines the assembled model. Unlike a digital image, the assembled model allows observation from any angle. The analysis is therefore significantly more precise.

[0249] At step iv), The orthodontist and / or the computer informs the patient, for example by sending a message to their phone. This message may inform the patient of an unfavorable situation and invite them to make an appointment with the orthodontist.

[0250] The orthodontist can also compare the assembled model with previously assembled models for the same patient. This analysis is particularly useful for assessing the patient's progress. The information provided can then inform the patient of any unfavorable changes in their condition, thus improving preventative care.

[0251] The assembled model can also be compared with one or more models obtained by scanning the patient's teeth or a mold of the patient's teeth, or with an updated reference model resulting from the implementation of a process described in WO 2016 066651. Application to embedded control

[0252] An image analysis according to the invention is also useful for guiding the acquisition of an image of a dental arch, particularly for establishing a remote diagnosis.

[0253] In particular, WO2016 / 066651 describes a method in which an initial reference model is deformed so as to obtain an updated reference model enabling the acquisition of reference images exhibiting maximum concordance with the "updated" images of the arcade acquired at the updated time.

[0254] Reference images are therefore views of the updated reference model, observed under virtual acquisition conditions that are as concordant as possible with the real acquisition conditions implemented to acquire the updated images of the patient's arcade.

[0255] The search for these virtual acquisition conditions is preferably carried out using metaheuristic methods.

[0256] To expedite this search, WO2016 / 066651 recommends conducting an initial, rough assessment of the actual acquisition conditions. For example, conditions corresponding to a position of the acquisition device greater than 1 meter from the teeth are excluded from the search.

[0257] However, there is a continuing need to accelerate the execution of the process described in WO2016 / 066651, and in particular, to search more quickly for virtual acquisition conditions that have maximum concordance with the actual acquisition conditions implemented to acquire an updated image of the patient's arch.

[0258] One aim of the invention is to address, at least partially, this problem.

[0259] The invention proposes a method for acquiring an image of a patient's dental arch, said method comprising the following steps: a') Activation of an image acquisition device so as to acquire an image, referred to as the "analysis image", of said arch; b') Analysis of the analysis image by means of a deep learning device, preferably a neural network, trained by means of a training set, preferably according to a detailed analysis method according to the invention, so as to identify at least one analysis tooth area representing a tooth on said analysis image, and to determine at least one tooth attribute value for said analysis tooth area, or according to a global analysis method according to the invention; c') Determination, for the analysis image, of a value for an image attribute, said value being a function of said tooth attribute value(s) if a detailed analysis method according to the invention was implemented in the preceding step; d') Optionally, comparison of said image attribute value with a setpoint;e') issuing an information message based on said comparison. ;

[0260] In one embodiment, in step b'), all said analysis tooth zones are identified, and at least one tooth attribute value is determined for each analysis tooth zone, and, in step c'), the value for the image attribute is determined as a function of said tooth attribute values.

[0261] In one embodiment, step b') comprises the following steps: 1) preferably before step a'), creation of a training set comprising more than 1,000 images of dental arches, or "historical images", each historical image comprising one or more areas each representing a tooth, or "historical tooth areas", to each of which, for said tooth attribute, a tooth attribute value is assigned; 2) training of at least one deep learning device, preferably a neural network, using the training set;3) submitting the analysis image to the deep learning device in such a way that it determines at least a probability relative to: the presence, at a location in said analysis image, of an area representing, at least partially, a tooth, or "analysis tooth area", the attribute value of the tooth represented on said analysis tooth area, 4) determining, as a function of said probability, the presence of a tooth at a position represented by said analysis tooth area, and the attribute value of said tooth.

[0262] In one embodiment, step b') comprises the following steps: 1') creation of a training set comprising more than 1,000 images of dental arches, or "historical images", each historical image having an attribute value for at least one image attribute, or "image attribute value"; 2') training at least one deep learning device, preferably a neural network, using the training set; 3') submitting the analysis image to the deep learning device in such a way that it determines, for said analysis image, at least one probability relative to said image attribute value.

[0263] In one embodiment, to create a historical image of the learning base, an operator, preferably an orthodontist, identifies one or more "historical" dent areas on an image, then assigns to each identified historical dent area a value for at least one dent attribute, and / or assigns to an image a value for at least one dent attribute.

[0264] In one embodiment, the information message is emitted by the acquisition device.

[0265] As will be seen in more detail later in the description, an acquisition method according to the invention makes it possible to verify whether an analysis image complies with a set of instructions and, if it does not, to guide the operator so that they acquire a new analysis image. The method thus enables "embedded control," preferably within the image acquisition device.

[0266] In particular, to implement the WO2016 / 066651 procedure, it may be desirable to acquire updated images from different acquisition directions, for example, a front image, a right-side image, and a left-side image. These updated images, acquired successively, can then be classified accordingly. This accelerates the search for virtual acquisition conditions that best match the actual acquisition conditions.

[0267] Indeed, the search can begin from virtual acquisition conditions in which the virtual acquisition device is in front of, to the left or to the right of the updated reference model, depending on whether the updated image under consideration is classified as a front, left or right image, respectively.

[0268] The operator, usually the patient, can, however, make a mistake when acquiring updated images. In particular, they may forget to take an updated image, for example, the frontal view, or swap two updated images. Typically, the operator might take an image on the right when they are expected to take an image on the left.

[0269] This reversal of the updated images can significantly slow down their processing. For example, if the updated image is supposed to be a left-facing image but was mistakenly taken from the right, the search for optimal virtual acquisition conditions—that is, those exhibiting maximum agreement with the actual acquisition conditions—will begin from a starting point offering a left-facing view of the reference model, whereas the optimal virtual acquisition conditions correspond to a right-facing view. The search will therefore be considerably slowed down.

[0270] Thanks to the invention, each updated image is an analysis image that can be analyzed and controlled, preferably in real time.

[0271] For example, the acquisition process can determine that the updated image was "taken from the right" and compare this image attribute value with the instruction given to the operator to take the updated image from the left. Since the attribute value of the updated image (taken from the right) does not correspond to the instruction (to acquire an updated image from the left), the acquisition device can immediately alert the operator so they can change the acquisition direction.

[0272] An acquisition method is now described in detail.

[0273] At step a'), The operator activates the image acquisition device in order to acquire an analysis image.

[0274] In one embodiment, the operator triggers the acquisition device so as to store the analysis image, preferably by taking a photo or video of his teeth, preferably using a mobile phone equipped with a camera.

[0275] Step a') can be performed as the acquisition of the updated images in step B) described above.

[0276] In another embodiment, the analysis image is not stored. In particular, the analysis image can be the image that appears in real time on the screen of the operator's mobile phone, usually the patient's.

[0277] In a first embodiment, at step b'), The analysis image is analyzed according to a detailed analysis method according to the invention. This analysis preferably leads to the assignment of a tooth attribute value to each identified analysis tooth zone, for example, to assigning a tooth number to each of the analysis tooth zones.

[0278] At step c'), An attribute value is determined for the analysis image based on the tooth attribute values. The analysis image attribute value can be relative to its overall orientation and can, for example, take one of the following three values: "right-hand view," "left-hand view," and "front view." The analysis image attribute value can also be a list of the numbers of the teeth shown, for example, "16, 17, and 18." The analysis image attribute value can also be, for example, the "presence" or "absence" of a dental appliance, preferably orthodontic, or the mouth opening status ("open mouth," "closed mouth").

[0279] In another embodiment, a global analysis method according to the invention is implemented in step b'). Advantageously, such a method allows a value for an image attribute to be obtained directly, without having to determine values ​​for a tooth attribute. It is therefore advantageously faster. However, the information resulting from a global analysis may be less precise than that resulting from a detailed analysis.

[0280] Steps a') to c') thus allow us to characterize the analysis image.

[0281] The characterization of the analysis image helps guide the operator if the analysis image does not correspond to the expected image, for example because its quality is insufficient or because it does not represent the desired teeth.

[0282] At step d'), we compare the image attribute value of the analysis image with a setpoint.

[0283] For example, if the instruction was to acquire a right-hand image and the image attribute value is "taken from the left", the comparison leads to the conclusion that the acquired image is "unsatisfactory".

[0284] At step e'), a message is sent to the operator, preferably via the acquisition device.

[0285] Preferably, the information message relates to the quality of the acquired image and / or the position of the acquisition device in relation to said arch and / or the setting of the acquisition device and / or the opening of the mouth and / or the wearing of a dental appliance, preferably orthodontic.

[0286] For example, if the acquired image is "unsatisfactory", the acquisition device may emit a light, for example red, and / or sound a beep, and / or generate a voice message, and / or vibrate, and / or display a message on its screen.

[0287] For example, if the image is to be acquired while the patient is wearing their braces and this is not the case, the acquisition device may emit the message "wear your braces for this image".

[0288] For example, if the image was acquired while the patient did not open their mouth wide enough or had their mouth closed, the acquisition device may emit the message "open your mouth more for this image".

[0289] In one embodiment, steps b') to c') are only performed if the operator records the scan image, i.e., presses the shutter button. The message then prompts the operator to acquire a new scan image. Optionally, the acquisition device deletes the unsatisfactory scan image.

[0290] In one embodiment, steps b') to c') are continuously performed while the acquisition device is operating, and the analysis image is an image displayed on a screen of the acquisition device. The acquisition device can thus, for example, emit red light as long as the analysis image is unsatisfactory, and emit green light when the analysis image is satisfactory. Advantageously, the acquisition device then stores only satisfactory analysis images.

[0291] As is now clear, the invention thus enables embedded control during the acquisition of analysis images. Applied to updated images of the WO2016 / 066651 process, steps a') to e') ensure that these images meet the requirements, and therefore significantly accelerate the execution of this process.

[0292] Steps d') and e') are optional. In one embodiment, the analysis image is associated only with its description, which specifies its image attribute value. This description also significantly speeds up the execution of the WO2016 / 066651 process because, when the analysis image is used as the updated image for this process, it allows for an approximate determination of the actual acquisition conditions of this image, eliminating the risk of a gross error, for example, due to a swap between two images.

[0293] Steps d') and e') are preferred, however. They allow, for example, to prevent the operator from forgetting a left-hand image or taking two redundant right-hand images. Application to the control of an orthodontic splint

[0294] Typically, at the beginning of orthodontic treatment, the orthodontist determines the desired tooth positioning at a specific point in the treatment, known as the "setup." The setup can be defined using an impression or a three-dimensional scan of the patient's teeth. The orthodontist then has, or fabricates, an orthodontic appliance tailored to this treatment.

[0295] The orthodontic appliance can be an orthodontic splint (“ align » (in English). A splint is classically presented as a removable one-piece device, classically made of a transparent polymer material, which has a channel shaped so that several teeth of an arch, generally all the teeth of an arch, can be housed in it.

[0296] The shape of the channel is adapted to hold the tray in position on the teeth, while also correcting the positioning of certain teeth ( Figures 14 and 15 ).

[0297] Typically, at the beginning of treatment, the shapes of the different aligners are determined for various stages of the treatment, and then all the corresponding aligners are manufactured. At predetermined times, the patient changes aligners.

[0298] Treatment with aligners offers the advantage of minimal discomfort for the patient. In particular, the number of orthodontist appointments is reduced. Furthermore, the pain is less than with a metal braces attached to the teeth.

[0299] The market for orthodontic aligners is therefore growing.

[0300] At regular intervals, the patient visits the orthodontist for a visual check, in particular to verify if the movement of the teeth is in accordance with expectations and if the aligner worn by the patient is still suitable for the treatment.

[0301] If the orthodontist diagnoses that the treatment is unsuitable, they will take a new impression of the teeth, or equivalently, a new three-dimensional scan, and then order a new set of aligners configured accordingly. On average, the number of aligners ultimately manufactured is estimated to be around 45, instead of the 20 typically planned at the start of treatment.

[0302] The need to travel to the orthodontist is a burden for the patient. The patient's trust in their orthodontist can also be affected. The misalignment can be unsightly. Finally, it results in an additional cost.

[0303] The number of check-up visits to the orthodontist should therefore be limited.

[0304] There is a need for solutions that address these problems.

[0305] One aim of the invention is to meet, at least partially, this need.

[0306] The invention provides a method for evaluating the shape of an orthodontic splint, said method comprising the following steps: a) acquisition of at least one image representing at least partially the splint in a service position in which it is worn by a patient, referred to as the "analysis image"; b) analysis of the analysis image using a deep learning device, preferably a neural network, trained using a training set, so as to determine a value for at least one tooth attribute of an "analysis tooth area" of the analysis image, the tooth attribute being relative to a spacing between the tooth represented by the analysis tooth area, and the splint represented on the analysis image, and / or for an image attribute of the analysis image, the image attribute being relative to a spacing between at least one tooth represented on the analysis image, and the splint represented on said analysis image; c") preferably, assessment of the suitability of the splint based on the value of said tooth or image attribute;d") preferably, issuing an informational message based on said assessment. ;

[0307] As will be seen in more detail later in the description, an evaluation method according to the invention greatly facilitates the assessment of the suitability of the aligner for the treatment, while also making this assessment particularly reliable. In particular, the method can be implemented using simple photographs or videos, taken without any special precautions, for example, by the patient. The number of appointments with the orthodontist can therefore be reduced.

[0308] Preferably, in step b''), all said analysis tooth zones are identified, and the value of said tooth attribute is determined for each analysis tooth zone, and, in step c"), the suitability of the gutter is determined according to said tooth attribute values.

[0309] Preferably, said tooth attribute is chosen from the group formed by a maximum spacing along the free edge of the tooth, an average spacing along the free edge of the tooth, and said image attribute is chosen from the group formed by a maximum spacing along the set of teeth represented, an average spacing along the free edges of the set of teeth represented, an overall acceptability of the spacing of the teeth represented.

[0310] The tooth attribute relating to a gap may be in particular the existence of a gap, this attribute being able to take the tooth attribute values ​​"yes" or "no"; or a value measuring the magnitude of the gap, for example a maximum gap observed or an assessment relative to a scale.

[0311] In step b''), a detailed analysis method according to the invention is preferably implemented, a tooth attribute of each historical tooth area of ​​each historical image of the training base being related to a gap between the tooth represented by the historical tooth area, and a groove carried by said tooth and represented on said historical image.

[0312] Preferably, step b'') includes the following steps: b"1) preferably before step a"), creation of a training set comprising more than 1,000, preferably more than 5,000, preferably more than 10,000 images of dental arches, or "historical images", each historical image representing a splint worn by a "historical" patient and comprising one or more areas each representing a tooth, or "historical tooth areas", to each of which, for at least one tooth attribute relating to a spacing between the tooth represented by the historical tooth area in question, and the splint represented, a tooth attribute value is assigned; b"2) training a deep learning device, preferably a neural network, using the training set;b) submitting the analysis image to the deep learning device in such a way that the deep learning device determines at least a probability relative to the presence, at a location of said analysis image, of an analysis tooth area, and the tooth attribute value of the tooth represented on said analysis tooth area; b) determining, as a function of said probability, the presence of a gap between the groove and the tooth represented by said analysis tooth area, and / or an amplitude of said gap.

[0313] Steps b''1) to b''4) may include one or more of the features, possibly optional, of steps 1) to 4) described above, respectively.

[0314] In one embodiment, in step b''), a global analysis method according to the invention is implemented, an image attribute of each historical image of the training base being related to a spacing between at least one tooth represented on the historical image, and a groove carried by said tooth and represented on said historical image.

[0315] Preferably, step b'') includes the following steps: b"1') creation of a training set comprising more than 1,000 images of dental arches, or "historical images", each historical image having an attribute value for at least one image attribute, or "image attribute value", relating to a gap between at least one tooth represented on the analysis image, and the splint represented on said analysis image; b"2') training at least one deep learning device, preferably a neural network, using the training set; b"3') submitting the analysis image to the deep learning device so that it determines, for said analysis image, at least one probability relating to said image attribute value, and determining, as a function of said probability, the presence of a gap between the splint and the tooth or teeth represented on the analysis image, and / or an amplitude of said gap.

[0316] Steps b''1') to b''3') may include one or more of the features, possibly optional, of steps 1') to 3') described above, respectively.

[0317] The process is now described when a detailed analysis is carried out in step b").

[0318] Prior to step a"), the training base must be enriched, preferably according to an enrichment process according to the invention, in order to contain historical images whose description specifies, for each of the historical tooth zones, a value for the tooth attribute relating to the spacing.

[0319] This information can be entered manually. For example, an operator, preferably an orthodontist, can be presented with an image representing one or more so-called "historical" tooth areas, and asked to identify these historical tooth areas and to indicate, for each historical tooth area, whether there is a gap or not and / or to assess the extent of this gap.

[0320] A historical image can be a photograph of a mouthguard worn by a historical patient. Alternatively, a historical image can be the result of processing an image of a bare dental arch (i.e., without a mouthguard) and an image of the same arch with the mouthguard. The image of the bare arch may, in particular, be a distorted model of the arch to achieve maximum concordance with the image of the arch with the mouthguard. Such processing can be especially useful for making the outline of the teeth and mouthguard more visible when the teeth are barely visible through the mouthguard.

[0321] At step a"), The acquisition of the analysis image can be carried out in the same way as the acquisition of the updated images in step B) described above.

[0322] Preferably, at least one reminder informing the patient of the need to create an analysis image is sent to the patient. This reminder can be in paper form or, preferably, electronically, for example, as an email, an automatic alert from a specialized mobile application, or an SMS. Such a reminder can be sent by the orthodontic practice or laboratory, by the dentist, or via the patient's specialized mobile application, for example.

[0323] Step a") is carried out when evaluation of the shape of a splint is desired, for example more than 1 week after the start of treatment with the splint.

[0324] The analysis image is an image representing the splint worn by the patient's teeth.

[0325] At step b"), The analysis image is analyzed according to a detailed analysis method according to the invention.

[0326] The deep learning device was trained using a training set containing historical images whose description specifies, for at least one, preferably each historical tooth area, a value for a tooth attribute relating to a gap between the tooth represented by the historical tooth area and the groove carried by said tooth and represented on said historical image.

[0327] The value for this tooth attribute therefore provides information about the shape of the splint relative to the shape of the patient's teeth.

[0328] The value for this tooth attribute can be a measure of the spacing, for example a measure of the maximum spacing, or of the average spacing for the tooth represented by the historical tooth area.

[0329] The deep learning device is therefore capable of analyzing the analysis image to determine, preferably for each of the "analysis tooth zones", the existence, or even the extent, of a gap in the groove of the tooth represented on the analysis tooth zone.

[0330] At step c"), Based on the results of the previous step, the suitability of the splint is assessed. For example, it is determined whether the gap between the splint and at least one tooth exceeds an acceptable threshold and, if so, a decision is made to replace the splint with a better-fitting one.

[0331] The suitability of a retainer can be assessed within the context of orthodontic treatment (whether the spacing is compatible with the orthodontic treatment), but also in the context of non-therapeutic treatment, particularly cosmetic treatment. Retainers can indeed be used to move teeth for purely aesthetic purposes, without this movement altering the patient's health. The suitability of the retainer can also be evaluated within the framework of a research program on the effectiveness of the retainer, for example, to evaluate a new retainer material on a human or another animal.

[0332] At step d"), Information relating to the assessment carried out in the previous step is issued, in particular to the patient and / or the orthodontist.

[0333] The orthodontist can then use this information, possibly in combination with additional information, for example the patient's age or the duration for which the aligner was worn, to establish a diagnosis and, if necessary, decide on an appropriate treatment.

[0334] In one embodiment, the process includes, in step b''), an overall analysis according to the invention. The other steps are not modified.

[0335] The analysis of the analysis image and historical images is then carried out globally, without identifying the individual situation of each of the teeth represented, and the image attribute is relative to the image as a whole.

[0336] For example, the image attribute related to spacing can relate to the acceptability of a dental situation due to one or more spacings, or to the overall extent of the spacing between the teeth. For example, the image attribute value could be "generally acceptable" or "generally unacceptable." The image attribute value could also be, for example, a measure of the spacing, such as the maximum spacing or the average spacing between the teeth shown in the analysis image and the splint.

[0337] As is now clear, a method according to the invention makes it possible, from simple photos or a simple film, to determine whether the splint is abnormally detached, or even, if a detailed analysis has been carried out in step b"), to determine the regions in which the splint has moved away from the teeth and to assess the extent of this separation.

[0338] The invention also relates to a method for adapting orthodontic treatment, a method in which a method for evaluating the shape of an orthodontic aligner according to the invention is implemented, and then, depending on the result of said evaluation, a new aligner is manufactured and / or the patient is advised, for example, to improve the conditions of use of his orthodontic aligner, in particular the positioning and / or the wearing times and / or the maintenance of his orthodontic aligner, in order to optimize the treatment.

[0339] The use of aligners is not limited to therapeutic treatments. In particular, an evaluation process can be implemented to assess an aligner used exclusively for aesthetic purposes.

[0340] The process can also be used to evaluate other dental parts or devices, including orthodontic ones. Computer program

[0341] The invention also relates to: a computer program, and in particular a specialized mobile phone application, comprising program code instructions for the execution of one or more steps of any process according to the invention, when said program is executed by a computer, a computer medium on which such a program is recorded, for example a memory or a CD-ROM.

[0342] Of course, the invention is not limited to the embodiments described and represented above.

[0343] In particular, the patient is not limited to a human being. A method according to the invention can be used for another animal.

[0344] The patient may be alive or dead. Preferably, he is alive.

[0345] The methods according to the invention can be implemented within the framework of orthodontic treatment, but also outside of any orthodontic treatment, and even outside of any therapeutic treatment.

Claims

1. Method for analysing an image of a dental arch of a patient, referred to as the "analysis image", comprising the patient acquiring the analysis image using a mobile telephone, the analysis image being a colour image and being chosen from a photograph and an image extracted from a film, the mobile telephone being more than 5 cm away from the dental arch at acquisition, said analysis method comprising the following steps: 1') creating a learning base containing more than 1000 images of dental arches, or "historical images", each historical image having an attribute value for an image attribute, or "image attribute value"; 2') training at least one neural network, by way of the learning base, by presenting the historical images at the input of the neural network so that the neural network gradually learns to recognize patterns in an image and to associate these patterns with image attribute values; 3') submitting the analysis image to the neural network so that it determines, for said analysis image, at least one probability relating to said image attribute value, and determining a value for said image attribute for said analysis image based on said probability; said image attribute relating to - the representation of the arches, the tongue, the mouth, the lips, the jaws, the gums, or a dental apparatus, the neural network being an object detection network.

2. Method according to the immediately preceding claim, wherein, in step 1'), in order to create a historical image for the learning base, an operator, preferably an orthodontist, assigns a value for at least one tooth attribute to an image.

3. Method according to either one of the two immediately preceding claims, wherein, in step 1), a learning base containing more than 10000 historical images is created.

4. Method according to any one of the preceding claims, comprising the following steps: a') activating an image acquisition apparatus so as to acquire the analysis image, b') analysing the analysis image by way of a neural network, trained by way of a learning base, in line with a method according to any one of the preceding claims.

5. Method according to Claim 4, comprising step d'): comparing said image attribute value with a setpoint.

6. Method according to the immediately preceding claim, comprising step e'): sending an information message based on said comparison, preferably via the mobile telephone.

7. Method according to any one of the preceding claims, the analysis image representing a plurality of teeth, preferably more than 2, more than 3, more than 4 or more than 5 teeth of the patient.

8. Method according to any one of the preceding claims, wherein the acquisition is triggered automatically, that is to say without any action from an operator, when the image acquisition apparatus determines that it is observing a dental arch and / or a retractor and that the observation conditions are satisfactory.

9. Method according to any one of the preceding claims, wherein the neural network is an R-CNN, a single shot multibox detector, or a faster region-based convolutional network.