Computer-implemented method for processing a dentate region of interest of a patient by means of at least one prediction model that has undergone machine learning for predicting residual bone volume (RBV) after healing of a single or multiple extraction site

Abstract

The invention relates to a computer-implemented method for processing an image of a dentate region of interest of a patient, the method comprising: a) obtaining an image of the jaws of a patient on which a region of interest is defined comprising a dentate sector with one or more teeth to be extracted; b) processing the region of interest in order to predict at least one edentulous sagittal 2D section, referred to as edentulous 2D section, the edentulous 2D section being a sagittal section perpendicular to the occlusion curve representative of a residual bone volume of a bony plate at an extraction site after extraction and healing, the processing implementing at least one model for predicting an edentulous section trained on the basis of pairs of dentate and edentulous 2D sections that are symmetrical with respect to the axis of symmetry passing through the middle of the mandibular symphysis of a reference patient for the lower jaw and by the median axis passing through the nasal spine for the upper jaw.

Description

[0001] DESCRIPTION

[0002] TITLE OF THE INVENTION: Computer-implemented method for processing a patient's tooth region of interest using at least one machine-learned prediction model enabling prediction of the Residual Bone Volume (RBV) after healing of a single or multiple extraction site.

[0003] TECHNICAL FIELD

[0004] This presentation concerns the field of dental medical imaging. This presentation concerns in particular, obtaining the prediction of sections representing a residual bone volume (designated by VOR) either from an examination using an imaging technique based on the digital analysis of the absorption of a conical beam called CBCT or from a 2D radiographic panoramic.

[0005] STATE OF THE ART

[0006] During a clinical examination at the dentist, a patient may present one or more irretrievable teeth which will have to be extracted and replaced by a mobile prosthesis or by one or more implants.

[0007] However, following tooth extraction and the healing process, it may turn out that the residual alveolar bone volume after healing is insufficient for the placement of an implant.

[0008] Indeed, following the extraction of a tooth or several teeth, a physiological healing process takes place which can lead to more or less significant bone loss which varies depending on the person, resulting radiologically in a reduction in the alveolar bone volume of the bone table.

[0009] One of the problems for the dentist is that it is not possible to predict this bone loss in advance before or during the tooth extraction. Consequently, it will only be a posteriori after the healing process that the practitioner will be able to assess this bone loss, allowing him to determine the residual "implantable" alveolar bone volume.In addition, the practitioner will then systematically use an examination using a 3D imaging technique based on the digital analysis of the absorption of a cone beam (CBCT examination, according to the English acronym Cone Beam Computed Tomography) and / or using a scanner to establish the anatomical morphology of the bone table in order to make measurements of the height and width of this residual bone volume (designated by VOR) allowing him to decide on the nature of his surgical procedure (implantation with or without filling biomaterial or pre-implant surgery as a bone graft) and the type of implant to be placed if necessary. The systematic use of a CBCT examination is restrictive.

[0010] This impossibility of predicting this residual or implantable bone volume (designated by VOI) can lead, in the event of significant bone loss, to a modification of the treatment plan previously presented to the patient and therefore to additionally plan (1) surgery as a bone graft or other(s) in order to restore a bone volume allowing the placement of one or more implants (2) a temporary waiting prosthesis restoring the chewing function during the healing period of the pre-implant intervention (3) a differentiated follow-up over time of the implant procedure. In addition, it should be noted that the placement of an implant as part of a single-stage surgery (immediate extraction-implantation) does not allow the maintenance of the bone in a patient with a physiology at risk of bone loss as defined later, which leads to clinical complications of the inflammatory type described in the literature as peri-implantitis.All of these interventions and setbacks are elements that modify the cost and duration of the treatment which can result in disputes between the practitioner and his patient.

[0011] Furthermore, there is no clinical and / or technical and / or any other means of determining the extent of bone loss following an extraction as well as its consequences on the implant and / or prosthetic procedure envisaged by the practitioner during one or more dental extractions scheduled in his treatment plan.

[0012] GENERAL STATEMENT

[0013] The invention is based on new findings both at the biological and technical levels, making it possible to overcome several of the shortcomings and drawbacks mentioned above.

[0014] The invention uses several prediction models developed and trained to allow the use of slices coming from a CBCT examination but also generated from a panoramic Rx. The tool used for this prediction is machine learning. The invention relates, according to a first aspect, to a computer-implemented method for processing an image of a toothed region of interest of a patient, the method comprising: a) obtaining an image of the jaws of a patient on which is defined a region of interest comprising a toothed sector with one or more teeth to be extracted;b) processing the region of interest to predict at least one 2D sagittal edentulous slice, called 2D edentulous slice, the 2D edentulous slice being a sagittal slice perpendicular to the occlusion curve representative of a residual bone volume of a bone table at an extraction site after extraction and healing, the processing implementing at least one model for predicting an edentulous slice trained from pairs of 2D toothed and edentulous slices symmetrical with respect to the axis of symmetry passing through the middle of the mental symphysis of a reference patient for the lower jaw and by the median axis passing through the nasal spine for the upper jaw.;

[0015] According to a first embodiment of the method according to the first aspect, in step a) the image is a 2D sagittal section of the region of interest of the patient's jaws, called a 2D dentate section, the 2D dentate section is a sagittal section perpendicular to an occlusion curve defined on the region of interest (ROI) and in step b) the 2D dentate sagittal section is processed to predict the 2D edentate section, the 2D dentate section (Cd) being representative of a dentate bone volume (VOD).

[0016] According to a second embodiment complementary to the first embodiment, in step a) a step of obtaining a 3D tomographic examination of a patient, obtained by an imaging technique based on the digital analysis of the absorption of a conical beam called 3D CBCT or a scanner, the 3D CBCT showing the patient's jaws with symmetrical toothed and edentulous sectors, in step a) the 2D toothed section is extracted from the 3D CBCT.

[0017] According to a third embodiment complementary to the first embodiment: before step a) a step of obtaining a 2D radiographic dental panoramic of a patient on which the region of interest of a toothed sector comprising a tooth to be extracted and an occlusion curve are defined, in step a) the 2D toothed section is obtained by processing the dental panoramic using a general prediction model (M1) trained to predict from the region of interest of the 2D radiographic dental panoramic of the toothed sector, a 2D section representative of a toothed volume (VOD) called a toothed section, the general prediction model (M1) having been trained from reference 2D radiographic dental panoramics with which 2D sections of the reference CBCT of the same reference patient are associated.

[0018] According to a fourth embodiment complementary to one of the first, second and third embodiments, in step b) the prediction model (M2a) of the edentulous cut is trained to predict the edentulous cut from the toothed cut obtained in step a).

[0019] According to a fifth embodiment of the invention according to the first aspect and possibly supplemented by one of the first, second, third and fourth embodiments, the prediction model of an edentulous section (M2a, M2b) is trained using reference images from the following two patient populations: patients presenting unilateral and / or embedded edentulism of one or more teeth with an edentulous sector symmetrical to this toothed sector

[0020] According to a sixth embodiment, complementary to the fifth embodiment, a classification (M3a) using a classification model of the 2D toothed section in order to define a risk of bone loss characterizing an insufficiency of residual bone volume to place an implant according to a percentage of bone loss obtained is implemented.

[0021] According to a seventh embodiment, complementary to one of the third, fourth, fifth and sixth embodiments, the general prediction model (M1) was trained from the following reference images for each reference patient: a reference 2D radiographic dental panoramic; and a reference 3D tomographic examination associated with the same patient, obtained by an imaging technique based on the digital analysis of the absorption of a conical beam called 3D CBCT or a scanner, the reference 3D CBCT showing the individual's jaws with toothed and edentulous sectors according to the examinations; and / or a virtual panoramic from the 3D CBCT; and a set of 2D sagittal sections taken perpendicular to the plane passing through the occlusion curve, modified or not, so that it is aligned with that of the reference 2D radiographic dental panoramic of this patient.

[0022] According to an eighth embodiment, complementary to the seventh embodiment, the learning comprises a step of re-aligning for the same reference patient, the virtual panoramic with the reference 2D radiographic panoramic so as to reposition the reference 2D radiographic panoramic within the 3D volume of the 3D CBCT of this patient allowing the obtaining of sagittal sections correctly located and oriented in relation to the defined region of interest of the Rx panoramic of this patient.

[0023] According to a ninth embodiment, complementary to the seventh embodiment, the learning comprises a step of re-aligning the reference 2D radiographic panoramic in the volume of the corresponding reference 3D CBCT.

[0024] According to a tenth embodiment of the invention according to the first aspect and possibly supplemented by one of the first to ninth embodiments, the prediction model (M2a) of the edentulous section was trained during a learning step from the following reference images for each reference patient: an associated reference 3D tomographic examination, obtained by an imaging technique based on the digital analysis of the absorption of a conical beam called 3D CBCT or a scanner, the reference 3D CBCT showing the jaws of the individual with symmetrical toothed and edentulous sectors whose extent of the sectors is variable; a set of 2D sagittal sections taken perpendicular to the plane passing through the occlusion curve, modified or not, so that it is recalibrated on that of the reference 2D radiographic dental panoramic of this patient when an Rx panoramic is used as an input image

[0025] According to an eleventh embodiment of the invention and possibly supplemented by one of the first to tenth embodiments, the model for predicting an edentulous section (M2a, M2b) is trained using reference images from the following two patient populations: patients presenting unilateral and / or embedded edentulism of one or more teeth with an edentulous sector symmetrical to this toothed sector.

[0026] According to a twelfth embodiment, the invention according to the first aspect comprises before step a) a step of obtaining a 2D radiographic dental panoramic of a patient on which the region of interest of a toothed sector comprising a tooth to be extracted and an occlusion curve are defined, in step a) the image is the region of interest of the 2D radiographic dental panoramic of the patient, in step b) the prediction model (M2b) of the edentulous section is, said model (M2b) being trained to predict the edentulous section from the region of interest of the 2D radiographic dental panoramic of the toothed sector comprising the tooth to be extracted

[0027] According to a second aspect, the invention relates to a computer-implemented method of determining a percentage of bone loss (PL), comprising the following steps:

[0028] - determination of a 2D toothed cut and an edentulous cut by means of the method according to the first aspect and optionally supplemented by one of the first to twelfth embodiments;

[0029] - processing (S4) of the 2D toothed section and the edentulous section to calculate a percentage of bone loss (PO).

[0030] In an embodiment of the method according to the second aspect, bone loss (PO) is defined by PO = (MC dentate side - MC edentate side) / MC dentate side when point M locating the alveolar nerve can be determined, or by PO = (BC dentate side - BC edentate side) / BC dentate side, when point M is not identified, for regions of interest symmetrical with respect to the axis of symmetry passing through the middle of the mental symphysis at the level of a lower maxilla, one being a dentate region and the other being an edentate region, with MC the total height of an alveolar table and BC a total height of a maxillary bone table where BC is replaced by the measurement of SC at the level of the upper maxilla, where point S is the point vertical to point C at the level of the sinus floor at the level of the tooth to be extracted and at the symmetrical edentate site.

[0031] The invention is based on acquired biological knowledge. Indeed, the invention addresses a lack of understanding of the individual variability of bone physiology in our patients following one or more dental extractions. Preliminary morphometric work was undertaken to study and clinically address this gap.Following morphometric studies on sagittal sections perpendicular to the occlusal plane and symmetrical with respect to the median axis passing through the middle of the mental symphysis from mandibular CBCT examinations involving three groups of subjects comprising 100 dentate subjects without dental malposition, 100 subjects with bilateral edentulism in the premolar-molar sectors and 100 unilaterally edentulous subjects still in the premolar-molar sectors, the applicant observed that: (1) the dentate subjects present at the level of the mandible a great volumetric bone variability with however a dimensional symmetry of the right and left bone tables in the premolar-molar sectors.

[0032] (2) bilateral mandibular edentulous subjects present a right and left dimensional symmetry of the bone tables and their VOR in the edentulous sectors in terms of height, width and surface. Given that these extractions were delayed in time, carried out by different practitioners and for different etiologies (infection, dental fracture, periodontal disease, etc.), the applicant considered that this bilateral symmetry of the VOR of the bone tables and volumetrically different from one individual to another corresponded to a biological and / or genetic predetermination of the healing phenomena or other mechanisms resulting in a determined amplitude of bone loss.

[0033] (3) as, in addition, in the same examination for unilaterally edentulous subjects, the bone table can be observed before and after extraction, it is therefore possible to measure bone loss after extraction individually. This bone table after extraction and healing represents the residual bone volume (RBV). The morphological dimension of this residual bone volume will determine the possibilities or not of implant placement.

[0034] The applicant therefore concluded that from the sagittal sections perpendicular to the occlusal plane and taken symmetrically with respect to the median axis passing through the mandibular symphysis in the group of subjects who were unilaterally edentulous or had embedded edentulism, it is possible to predict on the one hand the VOR (anatomy of the bone table after extraction and healing) and on the other hand to determine the individual bone loss linked to dental extractions by the difference in surface area between the surface of the bone table on the toothed side and the symmetrical surface measured on the edentulous side.

[0035] The invention also lies in the location of the cuts on the edentulous side. These must be cuts symmetrical to those on the toothed side defined by the axis of symmetry passing through the middle of the mental symphysis for the lower maxilla, also noted as the mandible.

[0036] The same principle of symmetrical cuts with respect to a median axis is applied to the upper jaw. This axis is the median axis passing through the nasal spine for the teeth of the upper jaw. This result of prediction of the morphology of the bone table and the bone loss thus predicted even before the surgical act of dental extraction are at the origin of certain elements of the first aspect of the presentation.

[0037] The invention has several advantages:

[0038] - Include Oral Medicine in a 5P Medicine (personalized, preventive, predictive, participatory and relevant) by allowing for a personalized assessment of individual bone loss following dental extraction. Adapted and differentiated monitoring according to the patient's risk of bone loss thus predicted could be put in place before any fixed or mobile prosthetic therapy.

[0039] - Replace the standard surgical implant procedure usually performed with a personalized implant surgery procedure based on the patient's bone physiology thanks to the prediction of the VOR even before tooth extraction. Similarly, adapted and differentiated monitoring according to the patient's risk of bone loss thus predicted can be put in place even before implant surgery.

[0040] According to a third aspect, the invention relates to a computer-implemented method for evaluating a bone region of a toothed or edentulous sector of a patient, the method comprising a) Obtaining a 2D radiographic dental panoramic of the maxillae of a patient on which a region of interest of a toothed sector or an edentulous sector and an occlusion curve are defined; b) Processing the region of interest by means of a general prediction model (M1) to predict a 2D sagittal slice perpendicular to the occlusion curve such that:

[0041] - if the region of interest is a dentate sector, a 2D slice representative of a dentate bone volume (VOD).

[0042] - if the region of interest is an edentulous sector, a 2D section representative of an implantable bone volume (OIV);

[0043] According to this third aspect, it is therefore possible to equip oneself with a tool allowing the exploration of morphological and volumetric anatomy in the form of 2D sagittal sections from a simple 2D panoramic radiographic examination. A CBCT is then no longer necessary.

[0044] Also, according to this third aspect for a toothed sector, the bone volume is used in particular to analyze the bone morphology and to be able to give an initial opinion on the possibilities of implant surgery. For an edentulous sector, the bone volume is used to analyze the bone morphology, take measurements and evaluate the implantable bone volume for the placement of an implant.

[0045] The method according to the third aspect, general and not linked to the prediction of the VOR, is advantageously supplemented by the following characteristics, taken alone or in any of their technically possible combinations:

[0046] - the general prediction model named M1 is trained using reference images from the following two patient populations: fully dentate patients, patients with unilateral and / or embedded edentulism of one or more teeth symmetrical to the dentate sector.

[0047] - the general prediction model was trained during a learning step from the following reference images for each reference patient: a reference 2D radiographic dental panoramic; a reference 3D tomographic examination associated with the same patient, obtained by an imaging technique based on the digital analysis of the absorption of a conical beam called 3D CBCT or a scanner, the reference 3D CBCT showing the individual's maxillae with toothed and edentulous sectors according to the examinations; a virtual panoramic from the 3D CBCT; a set of 2D sagittal slices taken perpendicular to the plane passing through the occlusion curve, modified or not, so that it is located on the region of interest of the reference 2D radiographic dental panoramic of this patient.

[0048] - the learning includes a step of re-aligning for the same reference patient, the virtual panoramic with the reference 2D radiographic panoramic in order to re-align the virtual panoramic from the 3D CBCT and the reference 2D radiographic panoramic.

[0049] The invention also relates to a computer program product comprising code instructions for executing the methods according to one of the embodiments mentioned above, when said program is executed on a computer.

[0050] The invention therefore proposes a method for generating sagittal sections from a 2D Rx dental panoramic using a prediction model that has been the subject of machine learning. To this end, this method is advantageous in that it allows a practitioner to dispense with a CBCT type examination to obtain a 2D sagittal section of the dentate and / or edentulous bone tables. According to the embodiment, it is thus possible to obtain information on a visualization of the Denture Bone Volume (DVB) of a tooth, the implantable bone volume (IBV) of an edentulous sector. In combination with at least one other prediction model, it is possible to obtain a prediction of the Residual Bone Volume (RBV) with a calculation of the bone loss (BL) even before the extraction of one or more teeth. A risk index on the possibilities of implant placement based on this VOR and bone loss (PO) is then determined.

[0051] It is also possible, from the generated and predicted sections, to perform morphological measurements using image processing to determine the percentage of bone loss. A classification of the toothed sagittal sections, adding a notion of risk, is carried out using a classification model that has been subject to machine learning.

[0052] Also in the invention a classification of the annotated Rx panoramic of this bone loss from which the sagittal section of the tooth to be extracted and that of its site after extraction are derived by adding a notion of risk is carried out by means of a classification model which has been the subject of machine learning.

[0053] Thus, thanks to these elements, it is possible to replace the standardized implant surgery with personalized implant surgery based on the patient's bone physiology and to allow, among other things, appropriate monitoring of patients while avoiding certain disputes.

[0054] PRESENTATION OF FIGURES

[0055] Other characteristics, aims and advantages of the invention will emerge from the following description, which is purely illustrative and non-limiting, and which must be read in conjunction with the attached drawings.

[0056] - Figure 1 illustrates methods according to the present disclosure for obtaining sagittal sections of toothed sectors (VOD), sagittal sections of edentulous sectors (VOI), sagittal sections of a tooth to be extracted (VOD) and the prediction of the residual bone volume (VOR) after extraction and healing of the extraction site as well as the calculation of the bone loss (PO) likely to occur during healing.

[0057] - Figure 2 illustrates data selection steps for training the prediction models implemented in this presentation. - Figure 3 illustrates the characteristics of obtaining and orienting the sagittal slices used in this presentation. On the left, an Rx panoramic and a corresponding virtual panoramic, on the right, a 3D image of a lower jaw on which are visible: the occlusion curve (Co), the plane passing through the occlusion curve (plco), and the sagittal slice perpendicular to this occlusion plane (Cs) as well as sagittal slices.

[0058] - Figure 4 illustrates the different steps used for learning models implemented in this presentation.

[0059] - Figure 5 illustrates a first procedure for recalibrating an occlusion curve of the virtual panoramic on that of the corresponding Rx panoramic.

[0060] - Figure 6 illustrates the preparation of sections which are segmented, masked, centralized, verticalized, and normalized.

[0061] - Figure 7 illustrates a second registration procedure used for the registration of a panoramic Rx in the volume of the CBCT associated with it.

[0062] - Figure 8 illustrates the different steps used for training the VOR prediction model when only the CBCT scan is present.

[0063] - Figure 9 illustrates images for learning a model for predicting an edentulous cut from a toothed cut;

[0064] - Figure 10 illustrates images for training a sagittal slice prediction model, this figure shows: a dental radiographic panoramic of a patient with embedded edentulism of a tooth and two regions of interest (ROI) one centered on a tooth representing the input image and the output image which is the predicted 2D slice showing a dentate bone volume (VOD), the other centered on the edentulous sector and the predicted 2D slice showing an implantable bone volume (VOI).

[0065] - Figure 11 illustrates a dental radiographic panoramic of a patient with unilateral edentulism of several teeth and showing different sections that can be obtained according to the present presentation.

[0066] - Figure 12 illustrates the measurement of bone loss implemented in the invention.

[0067] - Figure 13 illustrates a possible architecture in which the invention is implemented.

[0068] In all figures, similar elements have identical references. The figures illustrate, in a non-limiting manner, the prediction techniques applied to the lower jaw. Similar techniques are carried out in the upper jaw but using the median axis passing through the nasal spine of the maxilla to determine the symmetrical toothed and edentulous sections or sectors.

[0069] DETAILED DESCRIPTION

[0070] This disclosure relates to several methods described in relation to Figure 1.

[0071] Obtaining a 2D sagittal slice from a dental panoramic radiograph

[0072] One aspect of the disclosure relates to obtaining a 2D sagittal slice of a region of interest comprising a dentate or edentulous sector of a 2D radiographic dental panoramic of a patient (step S21).

[0073] A 2D dental radiographic panoramic (or Rx panoramic) is acquired by a practitioner in a known manner (step S01) and on which a region of interest is defined by an ad hoc means (step S11). Such an Rx panoramic and in particular the region of interest constitutes an input image to the method for obtaining a 2D section. An occlusion curve is defined on the Rx panoramic.

[0074] The region of interest includes, for example, a tooth (or toothed sector) or an edentulous sector or even a tooth to be extracted.

[0075] Several prediction models are then implemented to exploit this input image including the region of interest of a toothed or edentulous sector.

[0076] A prediction model M1 (called general model) makes it possible to predict (step S21) from the input image at least one 2D sagittal section taken perpendicular to the plane passing through the occlusion curve of the region of interest (hereinafter 2D section).

[0077] In this presentation the term cut must be understood as one or more cut(s).

[0078] The 2D cut predicted by the M1 prediction model is:

[0079] - a 2D toothed slice, noted CdM1, representing the toothed bone volume (designated by VOD) if the region of interest includes a tooth; or

[0080] - a 2D edentulous slice, denoted CedM1, representing the implantable bone volume (designated by VOI) available to place an implant if the region of interest is edentulous. Indeed, the 2D CdM1, CedM1 slice thus predicted makes it possible to obtain information on the bone volume. These predicted 2D slices can also be displayed to directly show the bone volume of the region of interest.

[0081] Obtaining a 2D slice from a panoramic Rx has the advantage of making the practitioner's work easier, as he can in fact do without a CBCT examination.

[0082] Obtaining VOR from a panoramic Rx or CBCT

[0083] One aspect of the presentation concerns obtaining the VOR from an image of a patient's jawbones on which a region of interest is defined, including a toothed sector with a tooth to be extracted. This image comes from a panoramic Rx or a CBCT examination or even a scanner. We understand that this residual bone volume, which is a prediction, actually makes it possible to evaluate an implantable bone volume.

[0084] An image of a patient's jaws comprising a toothed sector with a tooth to be extracted is acquired (step SO). In particular, a panoramic Rx (step S01) or a CBCT examination or a scanner (step S02) is acquired. A region of interest of a toothed sector comprising a tooth to be extracted is then defined (step S1). Such a region of interest constitutes an input image to the method for obtaining the VOR from a panoramic Rx or a CBCT examination or a scanner.

[0085] This region of interest is then processed (steps S2, S3) using one or more prediction models to obtain a 2D sagittal edentulous slice (CedM2), called a 2D edentulous slice, the 2D edentulous slice being a sagittal slice perpendicular to the occlusion curve representative of a residual bone volume (ROV) of a bone table at an extraction site after extraction and healing.

[0086] As will be detailed, such processing implements at least one model (M2a, M2b) for predicting an edentulous cut trained from pairs of 2D toothed and edentulous cuts symmetrical with respect to the axis of symmetry passing through the middle of the mental symphysis of a reference patient.

[0087] According to one embodiment, the VOR is obtained from a region of interest of a toothed sector defined on an Rx panoramic (step S11). In this case, one possibility is to then obtain a toothed slice CdM1 from the prediction model M1 (step S21) and then from the toothed slice CdM1 thus predicted by the model M1, to obtain an edentulous slice CedM2 representative of the residual bone volume (VOR) predicted by means of a prediction model M2a (step S3a). Another possibility is to obtain an edentulous slice CedM2 representative of the VOR directly from the Rx panoramic by means of a trained model M2b (step S3b).

[0088] This residual bone volume (RBV) predicts the bone volume available after healing for implant placement.

[0089] According to an embodiment, independent of that using the panoramic Rx, the VOR is predicted from a region of interest defined on a CBCT examination or a scanner (step S12). In this case, a 2D slice of the region of interest of the toothed sector to be extracted is extracted from the CBCT (step S22), also called a toothed slice Cd. Then, as previously, from this toothed slice Cd thus extracted, an edentulous slice CedM2 representative of the residual bone volume (VOR) is predicted using a prediction model M2a (step S3a).

[0090] Residual bone volume is the volume of healed bone remaining after extraction. This includes the bone volume defined by the height and width of the alveolar bone at the bone table at the end of the healing process.

[0091] Reference data

[0092] The methods of the present disclosure of the invention exploit several reference data for different learnings.

[0093] Figure 2 illustrates a selection of different examinations, these examinations come from a database comprising on the one hand reference 3D CBCTs and on the other hand reference 3D CBCTs associated with reference Rx panoramics of the same patient (step 10).

[0094] The various patient examinations used come from patients who are either fully dentate or have unilateral and / or embedded multiple or single edentulism (step 11). Alternatively, the methods can similarly use images from a CBCT or a scanner alone. These imaging techniques are well known to those skilled in the art and will not be discussed further.

[0095] Rx panoramics and CBCT reference images are anonymized and come from patients who have already been treated. CBCT images are in Dicom format.

[0096] In the population of patients with unilateral and / or embedded edentulism of one or more teeth, the mandibular and / or maxillary edentulism comes from patients whose extractions were carried out at least 2 to 3 years before the learning of the prediction model(s) and where, on the analysis of the sagittal sections of the CBCT, the edentulous crestal cortex is regular and where there is homogeneity of the underlying trabecular bone reflecting a phenomenon of completed healing.

[0097] To ensure data quality, the following exclusion criteria were used (this list is non-limiting and non-exhaustive): presence of dental agenesis or malposition, presence of implants in the studied sectors, presence of an extraction socket in the process of healing (recent extraction), tooth included in the studied sectors, mandible with maxillofacial surgery, bone preservation surgery or presenting osteosynthesis plates; presence of bone pathology (cementomas, ameloblastomas, etc.); CBCT resolution not allowing measurements (artifact, etc.), examination with a truncated basal edge in the studied sectors.

[0098] Figure 3 illustrates images used in this presentation. On the left are superimposed an Rx panoramic and a virtual panoramic. On the right, a 3D image of a lower jaw on which is shown, an occlusion curve Co, a plane passing through the occlusion curve plco, and the sagittal section perpendicular to this occlusion plane noted Cs. Below several sagittal sections Cs are shown. This figure illustrates the characteristics of obtaining and orienting the sagittal sections used in the learning processes of the different prediction models. The 3D image is obtained by means of a CBCT examination or a scanner.

[0099] Dental regions are identified using an MO model for detecting and identifying teeth and edentulous areas (step 12). Such a model is a convolutional neural network (CNN) taking as input a panoramic Rx and giving as output a JSON file containing the identifiers (numbering of areas containing or having contained a tooth) and their position (geometric coordinates) with rectangles surrounding each of the areas including the numbers of the present and absent teeth as well as the type of edentulism (unilateral and / or embedded). Bridges as well as implants are also identified and annotated. Such a model uses a convolutional neural network of the "Faster-RCNN" type which is widely used for object detection (see Ren, Shaoqing, et al. "Faster r-cnn: Towards real-time object detection with region proposal networks." Advances in neural information processing systems. 2015.The annotated Rx panoramics and the corresponding reference CBCT images are preferably stored in an SQLite database (step 13).

[0100] A similar procedure is performed on virtual panoramic images from CBCT and CT scans to identify teeth and edentulous areas.

[0101] This database is organized to allow for the collection of examinations of dentate patients, examinations of partially edentulous patients, and examinations of edentulous patients who must have a dentate sector symmetrical to the edentulous sector. Different types of data are used depending on the learning.

[0102] Data preparation

[0103] The training of the different prediction models uses data prepared from reference exams selected on the SQLite database previously described.

[0104] Data for training models M1, M2b and M3b

[0105] According to one embodiment, illustrated in figure 4, from the 3D CBCT, a virtual panorama is obtained (step E11).

[0106] Then, on the virtual panoramic, the dental regions are identified by means of an MO model for detecting and identifying teeth and edentulous sectors (step E12), the Rx panoramic being already annotated at this stage.

[0107] According to this embodiment, the virtual panoramic is then recalibrated on the Rx panoramic from this same patient, among other things, by aligning the occlusion curve of the virtual panoramic with that of the Rx panoramic (step E13). The aim is to obtain a virtual panoramic having the same characteristics in terms of size and arch curvature as those observed on the Rx panoramic associated with it, based among other things on the occlusion curve of the Rx panoramic and the dental geometric coordinates.

[0108] To proceed with the registration of the Rx and virtual panoramic views (step E13), as visible in Figure 5, a neck occlusion curve is determined (step E131) on the Rx panoramic view either from the coordinates of the top and center of the identification rectangles as defined previously or from the tip of the dental cusps easily identifiable by white pixels given the highly mineralized structure of dental enamel. A co2 occlusion curve is also determined in the same way on the corresponding virtual panoramic view (step E132). Then, a geometric transformation to align and size the co2 curve to the neck curve is performed (step E133). This transformation, added to the registration of the geometric coordinates of the teeth in the 3D volume, is then applied to create and extract a virtual panoramic view from the 3D (step E134) registered on the Rx panoramic view.

[0109] From this virtual panoramic image, reformatted and resized if necessary to be re-aligned with the corresponding Rx panoramic image where teeth and edentulous sectors have been identified, 2D sections perpendicular to the plane passing through the previously modified occlusion curve are obtained corresponding to the section of the Rx panoramic image entered as input (step E14).

[0110] These sagittal sections from this realigned virtual panoramic thus present a correct orientation and alignment in relation to the corresponding section of the teeth of the reference Rx pano.

[0111] The 2D slices are then segmented, masked, centered, straightened and normalized with or without identification of the different cortical, trabecular and alveolar nerve structures for the bone part alone and / or dental, thus differentiating the part of the image which will be trained in machine learning from the rest of the image (step E15).Figure 6 illustrates these different stages of processing the sagittal section (rectified (top left); segmented (top center); centered (top center); normalized (top right) then a segmentation detail: a 2D toothed section (image a), the same 2D segmented section to differentiate the global learning zone (bone and tooth OD) from the background B of the image (image b) and finally the same 2D segmented section to distinguish the different structures within this learning zone that are the dental morphology and at the level of the bone table, the cortical bone, the trabecular bone and the location of the alveolar nerve (image c).

[0112] Thus, each section of an Rx panoramic is associated with the corresponding 2D slice(s) with the exact orientation and location and prepared for machine learning (step E16).

[0113] Alternatively, it is possible to do without a virtual panoramic. According to this embodiment, a registration of the Rx panoramic in the CBCT volume is implemented and then on this Rx panoramic the dental regions are identified by means of an MO model for detecting and identifying teeth and edentulous sectors (step E12). According to this embodiment the previous step E11 is deleted. According to this embodiment, a registration procedure is implemented in relation to figure 7. To carry out the registration of the Rx panoramic in the CBCT volume it is a question of

[0114] (I) resize the panoramic Rx, without changing its aspect ratio, to the dimensions of the volume of the associated CBCT using as reference the total length of the occlusion curve traced on an axial slice of the associated CBCT and modifying the resolutions of the examinations

[0115] (II) reposition by projection all the pixel coordinates as well as the gray levels of this Rx panoramic in the CBCT volume by means of the percentage method among others and the occlusion curve traced on the axial section

[0116] (III) Then define a point on this panoramic Rx thus recalibrated in the CBCT where we want to have a sagittal section.

[0117] (IV) At the level of the projection of this point chosen and recalibrated in the CBCT, at this defined point, we can determine on an axial section the normal vector to the occlusion curve defined previously. This normal vector makes it possible to give the direction of a section segment.

[0118] (V) Pixels can then be sampled along a slice segment and by repeating this sampling process using the same slice segment on all axial slices of the CBCT, the sagittal slice CS perpendicular to the occlusal plane passing through this point can be reconstructed.

[0119] Data for training the M2a model (E10: figure 8 and figure 9)

[0120] The M2a prediction model uses 3D CBCT scans of patients with unilateral and / or symmetrical embedded edentulism of a toothed sector.

[0121] Dental regions are identified using a model for detecting and identifying teeth and edentulous areas (step E101).

[0122] 2D sagittal sections perpendicular to the occlusal plane passing through the occlusal curve are obtained for each tooth and edentulous sector (step E102).

[0123] The symmetrical 2D sagittal slices of the dentate and edentate with respect to the median axis passing through the middle of the mental symphysis are segmented and the bone tables of the sagittal slices are centered on the image then vertically straightened and finally normalized (step E103). The aim of this operation is to obtain sagittal slices whose dentate and edentate bone tables are positioned and located similarly in the image subjected to machine learning. The segmentation makes it possible to obtain a dentate and edentate bone table on a black background allowing machine learning of only the dentate and edentate bone table. The slices thus prepared are then subjected to prediction models such as neural networks or generative adversarial networks (GAN).

[0124] Figure 9 schematically illustrates these data for this learning from left to right:

[0125] - a CBCT of a patient with unilateral symmetrical edentulism of the toothed sector 45, 46 and 47;

[0126] - a toothed cut Cd and its edentulous symmetrical Ced_s and learning the edentulous cut from the toothed cut below;

[0127] - the verification of the model, on the first line from left to right the input toothed section Cd, in the middle the “Real edentulous section coming from the CBCT” Ced_s and on the right the predicted edentulous section CedM2; on the second line the real edentulous section Ced_s the predicted edentulous section CedM2, the superposition of the two sections allowing a visual comparison between the real edentulous section and the predicted edentulous section.

[0128] Learning prediction models

[0129] The prediction models used here are, for example, neural networks or Generative Adversarial Networks (GANs). After a machine learning phase (supervised or unsupervised), a prediction model "learns" and becomes able to apply, in an identical way to unknown data, what it has learned using reference data.

[0130] Learning the M1 model for predicting sagittal slices (step AM1)

[0131] The M1 learning prediction model uses either a convolutional neural network (CNN) or a generative adversarial network (GAN). The main difference between these two types of architectures comes from the optimized cost function. Classical convolutional networks optimize a cost function based on the pixel-wise error between the generated image and the reference image. Adversarial networks use the pixel-wise error as well as the adversarial error. The adversarial error comes from a part of the architecture called the discriminator whose function is to determine whether the image generated by the network is true (real) or false (generated).

[0132] Convolutional network is for example: SegNet or others. GAN network is for example Pix2Pix or Pix2PixHD or SPADE or others.

[0133] The result of this learning in the M1 model is the prediction of sagittal sections of the dentate and edentulous sectors.

[0134] The M1 model training uses panoramic Rx and CBCT scans of the same patient. These scans come from dentate and partially edentulous patients.

[0135] The learning (step E1, figure 4) of the M1 model is done using on the one hand the previously prepared data, namely Rx panoramics with teeth and edentulous sectors identified as described previously and on the other hand sagittal slices identified and correctly aligned and oriented to the corresponding teeth of the Rx panoramic. These identified sagittal slices from the virtual pano coming from the 3D CBCT were prepared as mentioned previously. The learning is done in the same way for the toothed and edentulous sectors of the Rx panoramic.

[0136] The M1 model is then trained to predict 2D slices from an Rx panoramic.

[0137] The learning (step E2) of the prediction model M1 is therefore carried out from pairs formed by a panoramic section centered on the tooth of which we want a representation in sagittal view and the 2D slices, corresponding to this section of tooth, extracted from the virtual pano of the CBCT after registration of the virtual panoramic on the panoramic Rx or by registration of the panoramic Rx in the CBCT volume as described previously. We use as training data the examinations of totally dentate people and the examinations of unilaterally or partially edentulous people for the sagittal representation of the edentulous section.

[0138] The result of this M1 model therefore allows, from a region of interest of the Rx panoramic input, to obtain as output the 2D slices corresponding to this tooth (2D toothed slices (CdM1) corresponding to the VOD). Similarly, from an edentulous sector of the 2D Rx dental panoramic, the M1 model allows to obtain as output the 2D slices corresponding to this edentulous sector (2D edentulous slices (CedM1) corresponds to the VOI).

[0139] In addition to training, the training data is enhanced by one or more augmentation treatments: random transformation such as rotation, enlargement, modification of contrast, brightness, etc. Such treatments not only increase the quantity of training data but also allow for the proper processing of images obtained under different conditions (brightness, framing, etc.). These treatments are applied to all or some of the data.

[0140] Once the model is correctly trained, it is advantageously subjected to new learning with different examinations which had been excluded for the basic learning: presence of implants in the sectors studied; presence of an extraction socket in the process of healing (recent extraction).

[0141] This additional learning makes it possible to increase prediction capabilities while testing the implemented model.

[0142] All the steps described are identical for the prediction of 2D sections of the dentate (VOD) and / or edentate (VOI) bone tables of teeth of the upper jaw.

[0143] In summary, this first M1 model makes it possible to dispense with a CBCT type examination to obtain 2D sections of the toothed and / or edentulous bone tables.

[0144] Figure 10 illustrates a panoramic Rx (Pano Rx) with two regions of interest designated by ROI1 and ROI2 which correspond to symmetrical regions. The dentate (CdM1) and edentulous (CedM1) slices are shown in this figure.

[0145] Learning models M2a, M2b for predicting residual bone volume (RBV) (step AM2a, AM2b)

[0146] Like the M1 prediction model, these M2a and M2b prediction models each use a convolutional neural network or a generative adversarial network (GAN). The convolutional network is for example: SegNet. The GAN network is for example Pix2Pix or Pix2PixHD or SPADE or others.

[0147] The M2a prediction model of the edentulous slice uses CBCT scans exclusively of unilateral and / or embedded edentulous patients with edentulousness of one or more symmetrical teeth in the dentate sector. The M2a prediction model predicts an edentulous slice CedM2 representative of the VOR of a dentate region. This CedM2 slice is to be distinguished from the edentulous slice CedM1 which shows a VOI of an edentulous region designated on the panoramic Rx (see model M1).

[0148] The learning of the M2a model associates / learns with the section of the tooth to be extracted, the 2D slices of this tooth and the symmetrical 2D edentulous slices. The M2a model is trained to predict the edentulous slice CedM2 from a toothed slice Cd and uses toothed slice / symmetrical edentulous slice pairs.

[0149] The M2b prediction model uses panoramic Rx and CBCT examinations and in particular examinations of unilateral and / or embedded edentulous patients presenting edentulism of one or more symmetrical teeth of the dentate sector. The M2b prediction model makes it possible to predict the edentulous CedM2 section representative of the VOR of a dentate region.

[0150] The learning of the M2b model associates / learns with the section of the tooth to be extracted from the Rx panoramic, the 2D slices not of this tooth but the symmetrical edentulous 2D slices of this examination. This learning is done as for that of the M1 model described above from the 2D slices previously prepared and coming from the CBCT corresponding to the Rx panoramic of the same patient whose occlusion curves and teeth have been previously registered. The M2b model is trained to predict the edentulous slice(s) CedM2 which corresponds to the dentate region of interest of an Rx panoramic and uses pairs of dentate region of the Rx pano / corresponding symmetrical edentulous slices of this dentate region.

[0151] The possibility of prediction either from a panoramic Rx or directly from a CBCT is allowed thanks to the previously described morphometry results where symmetry of the right and left residual bone volumes was demonstrated on bilateral edentulous patients.

[0152] The unilateral edentulous examinations used in the learning of the M2a model where there is a single or embedded edentulism of one or more teeth therefore have the characteristic of presenting on the same examination, the VOD of the tooth to be extracted and the VOR of the tooth after extraction and healing, thus allowing personalized machine learning of the residual bone volume from the dentate bone volume. The learning of the M2a model (step AM2a) therefore takes into account pairs of predicted 2D slices as for the learning of the M1 model but whose major difference with the M1 model is that we make the symmetrical edentulous sagittal slices representing the residual bone volume learn / correspond by means of machine learning to the dentate sector put as input.

[0153] The M2a model therefore makes it possible to predict, from the dentate sagittal sections Cd, edentulous sagittal sections corresponding to the extraction site after healing even before the extraction of one or more teeth by learning the association of the dentate sagittal sections Cd with the symmetrical edentulous sagittal sections CedM2 on the examinations of edentulous patients presenting unilateral or embedded edentulation of one or more teeth symmetrical to a dentate sector. This M2a model uses as an original examination either a panoramic Rx associated with a CBCT examination of the same patient or only a CBCT (figures 4 and 8).

[0154] The M2b model allows for the direct prediction of symmetrical 2D edentulous CedM2 sections of the toothed sector of the panoramic Rx of the tooth to be extracted, based on examinations of edentulous people presenting a toothed sector symmetrical to unilateral or embedded edentulism of one or more teeth.

[0155] Figure 11 shows a panoramic Rx on which an X axis of symmetry passing through the middle of the mental symphysis has been defined. The examination illustrates a panoramic showing the lower and upper jaw of which only the lower jaw is exploited. It presents a unilateral edentulism whose X axis passing through the middle of the mental symphysis makes it possible to define the teeth symmetrical to the edentulous sector. On the left dentate area are designated areas for which dentate slices can be obtained: cda, cdb. These slices are dentate slices and show a dentate volume VODa, VODb. On the right of the axis of the slices Céda, Cedb are shown. The Céda and Cedb slices are slices which make it possible to obtain an implantable bone volume VOIa and VOlb.However, as detailed previously, the Cda cut and the Céda cut form a pair for learning the M2a model, tooth 47 of the corresponding panoramic forming with the Céda cut a pair for learning the M2b model.

[0156] Obtaining bone loss (designated PO)

[0157] According to one aspect of the present disclosure, when the toothed section Cd and the edentulous section CedM2 are available, the latter make it possible to calculate a percentage of bone loss (denoted PO) (step S4). In particular, this involves processing the predicted toothed and edentulous sections to deduce a percentage of bone loss by calculation (see below).

[0158] Bone loss is defined as the loss of bone height measured between the bone height before extraction and the bone height at the end of healing after tooth extraction. This bone loss is restricted to the alveolar bone, which is defined in the premolar-molar regions of the lower jaw as the bone located above the alveolar nerve. In the upper jaw as the bone located below the maxillary sinus in the premolar-molar regions and in the incisor-canine region regardless of the jaw as the bone surrounding the roots of the teeth.

[0159] Bone loss is inferred by means of image processing (step S4) between the edentulous CedM2 slices from the M2a or M2b model and the dentate CdM1 slices from the M1 model or extracted directly from the CBCT, Cd. Such processing allows the calculation of measurements mainly of height and width of the bone table in its entirety when the point M (alveolar nerve) cannot be located on the sagittal slice or only in its alveolar part when the alveolar nerve is annotated. This image processing is done on the 2D dentate slices of the tooth to be extracted and the predicted edentulous CedM2 sagittal slices as described for the M2a, M2b models.

[0160] When point M is annotated, the calculation of the percentage of bone loss (PO) is as follows: PO = (MC dentate - MC edentate) / MC dentate where MC is the height of the alveolar part of the bone table (M is the superior point of the alveolar nerve) and C (the Crestal point at the extraction site).

[0161] If point M cannot be identified, the calculation of PO is as follows: PO = (BC toothed side - BC edentulous side) / BC toothed side where BC is the total height of the bone table (B is the Basal point of the mandible at the extraction site and C the Crestal point at the extraction site).

[0162] The calculation of the PO, which represents the difference in MC measurements between the dentate and edentulous sectors, is only done at the level of the premolars and molars in the lower jaw. In the other sectors, the difference in BC measurement is used (Figure 12).

[0163] At the level of the upper jaw, the measurement is done identically except that BC is replaced by the measurement of SC where point S is the point vertical to point C at the level of the floor of the sinus at the level of the tooth to be extracted and at the symmetrical edentulous site. The determination of the PO measurements at the level of the maxilla can only be calculated for teeth related to the sinus.

[0164] Classification of bone loss (see Figure 12)

[0165] According to one aspect of the present disclosure, it is possible to obtain a class that corresponds to bone loss (step S5).

[0166] In particular, using a classification model M3a, a classification of the CedM2 toothed sagittal slices according to the extent of bone loss and the risk of VOR insufficiency for implant placement is obtained (step S5a). Alternatively, a similar result can be obtained by classification of a panoramic Rx (step S5b).

[0167] Indeed, the calculation described above for bone loss makes it possible to annotate 2D toothed slices associated with symmetrical edentulous 2D slices to classify the 2D toothed slices according to a risk of bone loss thus annotated using a trained M3a model for bone loss prediction classification (step AM3a). Such an M3a model makes it possible to classify each sagittal toothed slice (Cd) into several classes according to the importance of this predicted bone loss. This classification is associated with a clinical risk of insufficient radiological height of the bone table for implant placement: Low risk class PO < 5%; Moderate risk class 5% < PO < 20%; Severe risk class PO > 20%.

[0168] Alternatively, an M3b model can directly classify a dentate region of interest from an Rx panoramic. To do this, the M3b model directly associates for its training, an annotated dentate panoramic section of the measured PO / 2D dentate slices.

[0169] For the upper jaw, this classification is identical to what is done at the level of the lower jaw or mandible.

[0170] In this presentation and for reasons of simplification and clarity, only elements of description of the treatments used at the level of the lower jaw are described but all these treatments are applied in an identical way to the upper jaw. The only difference lies on the one hand in the axis of symmetry which is in the upper jaw, the axis passing through the nasal spine of the upper jaw whereas at the level of the lower jaw or mandible, this axis passes through the middle of the mental symphysis and on the other hand at the level of the premolar-molar sector of the measurement SC of bone loss where "C" is the most inclined point of the dental crest at the extraction site and the point S is the point vertical to the point C at the level of the floor of the sinus at the level of the tooth symmetrical to this edentation.

[0171] Architecture

[0172] The different aspects described are implemented within an architecture illustrated in Figure 13.

[0173] The learning (AM1, AM2, AM2a, AM2b, AM3a, AM3b) can be implemented on a first server 1 and the processing (steps S2, S3, S4, S5) can be implemented on a second server 2 (or even more). Each server 1, 2 advantageously comprises at least one processing unit 3, 4, for example a processor. Each server 1, 2 can also comprise a communication interface 7, 8 allowing communication between them. Such communication can for example be implemented by means of a wired or wireless link via any type of communication network, for example the Internet. As regards the first server 1, it can have access via the communication interface 7 to one or more remote database(s) useful for learning.

[0174] Also advantageously, each server 1, 2 may comprise a storage unit 5, 6, for example a hard disk. Typically, the first learning server 1 may store in the storage unit 5 one or more databases used for learning or have access to this or these databases. The architecture may advantageously be supplemented by a memory which makes it possible to store the intermediate calculations during learning or processing.

[0175] A person skilled in the art will readily understand that other architectures for implementing the methods are possible, the two servers 1, 2 being able in particular to be merged.

Claims

Tl CLAIMS 1. Computer-implemented method for processing an image of a patient's toothed region of interest, the method comprising: a) obtaining (SO, S1) an image of a patient's jaws on which a region of interest (ROI) is defined comprising a toothed sector with one or more teeth to be extracted;b) processing (S2, S3) of the region of interest (ROI) to predict at least one 2D sagittal edentulous slice (CedM2), called 2D edentulous slice, the 2D edentulous slice being a sagittal slice perpendicular to the occlusion curve representative of a residual bone volume (VOR) of a bone table at an extraction site after extraction and healing, the processing implementing at least one model (M2a, M2b) for predicting an edentulous slice trained from pairs of 2D toothed and edentulous slices symmetrical with respect to the axis of symmetry passing through the middle of the mental symphysis of a reference patient for the lower maxilla and by the median axis passing through the nasal spine for the upper maxilla.; 2. Method according to claim 1, wherein in step a) the image is a 2D sagittal section of the region of interest of the patient's jaws, called a 2D dentate section (Cd), the 2D dentate section (Cd) being a sagittal section perpendicular to an occlusion curve defined on the region of interest (ROI) and wherein in step b) the 2D dentate sagittal section is processed to predict the 2D edentate section (CedM2), the 2D dentate section (Cd) being representative of a dentate bone volume (VOD).

3. Method according to claim 2, comprising before step a) a step of obtaining (SO) a 3D tomographic examination of a patient, obtained by an imaging technique based on the digital analysis of the absorption of a conical beam called 3D CBCT or a scanner, the 3D CBCT showing the patient's jaws with symmetrical toothed and edentulous sectors, in step a) the 2D toothed section (Cd) is extracted (S22, SEL) from the 3D CBCT.

4. Method according to claim 2, comprising before step a) a step of obtaining (SO) a 2D radiographic dental panoramic of a patient on which the region of interest of a toothed sector comprising a tooth to be extracted and a curve occlusion are defined, in step a) the 2D toothed slice (CdM1) is obtained by processing the dental panoramic using a general prediction model (M1) trained to predict from the region of interest of the 2D radiographic dental panoramic of the toothed sector, a 2D slice representative of a toothed volume (VOD) called toothed slice (CdM1), the general prediction model (M1) having been trained from reference 2D radiographic dental panoramics to which are associated 2D slices of the reference CBCT of the same reference patient.

5. Method according to one of claims 2 to 4, wherein in step b) the prediction model (M2a) of the edentulous cut (CedM2) is trained to predict the edentulous cut (CedM2) from the toothed cut (Cd, CdM1) obtained in step a).

6. Method according to one of claims 1 to 5, in which the model for predicting an edentulous section (M2a, M2b) is trained using reference images from the following two patient populations: patients with unilateral and / or embedded edentulism of one or more teeth with an edentulous sector symmetrical to this toothed sector.

7. Method according to one of claims 2 to 6, comprising a classification (M3a) by means of a classification model of the 2D toothed section (Cd, CdM1) in order to define a risk of bone loss characterizing an insufficiency of residual bone volume to place an implant according to a percentage of bone loss obtained.

8. Method according to one of claims 4 to 7, in which the general prediction model (M1) was trained from the following reference images for each reference patient: - a reference 2D dental radiographic panoramic; and - a reference 3D tomographic examination associated with the same patient, obtained by an imaging technique based on the digital analysis of the absorption of a conical beam called 3D CBCT or a scanner, the reference 3D CBCT showing the individual's jaws with toothed and edentulous sectors according to the examinations; and / or - a virtual panoramic view from the 3D CBCT; and - a set of 2D sagittal slices taken perpendicular to the plane passing through the occlusion curve, modified or not, so that it is aligned with that of the reference 2D radiographic dental panoramic of this patient.

9. Method according to claim 8, in which the learning comprises a step of re-aligning, for the same reference patient, the virtual panoramic with the reference 2D radiographic panoramic so as to reposition the reference 2D radiographic panoramic within the 3D volume of the 3D CBCT of this patient, allowing the obtaining of sagittal sections correctly located and oriented relative to the defined region of interest of the Rx panoramic of this patient.

10. Method according to claim 8, in which the learning comprises a step of re-aligning the reference 2D radiographic panoramic in the volume of the corresponding reference 3D CBCT.

11. Method according to one of claims 1 to 10, in which the prediction model (M2a) of the edentulous section (CedM2) was trained during a learning step from the following reference images for each reference patient: - an associated reference 3D tomographic examination, obtained by an imaging technique based on the digital analysis of the absorption of a conical beam called 3D CBCT or a scanner, the reference 3D CBCT showing the individual's jaws with symmetrical toothed and edentulous sectors whose extent of the sectors is variable; - a set of 2D sagittal slices taken perpendicular to the plane passing through the occlusion curve, modified or not, so that it is aligned with that of the reference 2D radiographic dental panoramic of this patient when an Rx panoramic is used as an input image.

12. Method according to one of the preceding claims, in which the prediction model of the 2D edentulous section (M2a, M2b) is trained using reference images from patients presenting unilateral and / or embedded edentulism of one or more teeth with a symmetrical toothed sector.

13. Method according to claim 1, comprising before step a) a step of obtaining (S0) a 2D radiographic dental panoramic of a patient on which the region of interest of a toothed sector comprising a tooth to be extracted and an occlusion curve are defined, in step a) the image is the region of interest of the 2D radiographic dental panoramic Rx of the patient, in step b) the prediction model (M2b) of the edentulous section (CedM2) is, said model (M2b) being trained to predict the edentulous section (CedM2) from the region of interest of the 2D radiographic dental panoramic of the toothed sector comprising the tooth to be extracted.

14. A computer-implemented method for determining a percentage of bone loss (PL), comprising the following steps: - determination of a 2D toothed cut and an edentulous cut by means of the method according to one of claims 2 to 13; - processing (S4) of the 2D toothed section (Cd) and the edentulous section (CedM2) to calculate a percentage of bone loss (PO).

15. Method according to the preceding claim, in which the bone loss (PO) is defined by PO = (MC toothed side - MC edentulous side) / MC toothed side when the point M locating the alveolar nerve can be determined, or by PO = (BC toothed side - BC edentulous side) / BC dentate side, when point M is not identified, for regions of interest symmetrical with respect to the axis of symmetry passing through the middle of the mental symphysis at the level of a lower maxilla, one being a dentate region and the other being an edentulous region, with MC the total height of an alveolar table and BC a total height of a maxillary bone table where BC is replaced by the measurement of SC at the level of the upper maxilla, where point S is the point vertical to point C at the level of the sinus floor at the level of the tooth to be extracted and at the symmetrical edentulous site.

Images