Computer-implemented method for training a machine learning model to complete an incomplete digital impression of a dental feature
A machine learning model trained on dental impression datasets addresses incomplete reconstructions by completing digital impressions in real time, improving accuracy and efficiency.
Patent Information
- Application Number
- PCT/IB2026/054393
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2025-05-05
- Filing Date
- 2026-05-04
- Publication Date
- 2026-11-12
AI Technical Summary
Existing digital dental impression techniques suffer from incomplete reconstructions due to occluded areas, light reflection, and scanner movement, necessitating re-scanning and consuming time, while geometric algorithms are inaccurate and resource-intensive.
A machine learning model trained on datasets of incomplete and complete dental impressions to rapidly complete digital impressions in real time, reducing the need for re-scanning and improving accuracy.
The model enables real-time completion of digital impressions, enhancing accuracy and reducing overall time required for obtaining a complete digital impression.
Abstract
Description
[0001] COMPUTER- IMPLEMENTED METHOD FOR TRAINING A MACHINE LEARNING MODEL TO COMPLETE AN INCOMPLETE DIGITAL IMPRESSION OF A DENTAL FEATURE
[0002] The present invention relates to the technical field of dental technology, and refers to a computer-implemented method for training a machine learning model to complete an incomplete digital impression of a dental feature . In the dental industry, the necessity of obtaining an estimate of the three-dimensional structure of a dental feature of a patient is well-known, for example in order to manufacture dental appliances (such as dental braces, clear aligners, retainers, implants, dentures and other prosthetics) especially tailored for the patient, or in order to help a dentist to diagnose dental diseases and devise corresponding treatment plans . Examples of dental features are one or more teeth, one or more parts of a tooth, one or more parts of a dental arch, and an entire dental arch.
[0003] Traditionally, the dental industry has relied heavily on physical dental impressions for obtaining an estimate of the three-dimensional structure of the dental features of a patient . These dental impressions are usually obtained by placing a semi-solid impression material (such as alginate) on a tray shaped so as to fit over the dental arch of the patient, and then placing the tray in the mouth over the arch: as soon as the impression material solidifies, it bears the physical negative imprint of the dental arch, from which a positive reproduction (such as a cast) can be formed obtaining the estimate of the three-dimensional structure of the dental features of the dental arch.This solution, however, is relatively invasive, since the tray should be kept in the mouth for several minutes in order to allow the impression material to at least partially solidify, creating discomfort for the patient and the risk of gag reflex . In addition, the retrieval of the impression material without deformation is a complicated operation which is difficult to perform properly and requires specific training. Finally, each physical dental impression requires a fresh amount of impression material, the preparation of which is time consuming .
[0004] As an alternative, digital dental impressions are becoming more and more common in the dental industry. This technology usually involves the use of an intraoral scanner which is inserted into the mouth of the patient in one or more swipes and captures a set of images that are used to reconstruct a digital impression of the observed dental feature : the digital impression contains three-dimensional structures that together form the estimated three-dimensional structure of the observed dental feature .
[0005] Actually, the digital dental impressions could be obtained by means of many different solutions, including active wavefront sampling, "quasi" confocal microscopy and active or passive stereo cameras .
[0006] These "digital" solutions allow to obtain the impressions in a less invasive and more rapid way. However, the digital impressions that are obtained with the first swipe of the scanner are often incomplete, having "holes" corresponding for example to occluded areas, dark areas, areas of light reflection or areas blurred as a consequence of patient movement or scannermovement . In fact, all the mentioned digital solutions rely on light patterns that are proj ected on the observed dental feature : in the case of a direct reflection, too much light is reflected to the scanner, which is therefore "blinded" and cannot distinguish the proj ected light pattern on the observed surface, creating a hole in the digital impression. Further, in the case of steep surfaces in the observed dental feature, too little light is reflected to the scanner to distinguish the proj ected light pattern, creating a hole in the digital impression. Other holes can appear in digital impressions as a consequence of a low resolution thereof, which could be caused for example by a proj ected light pattern that is too coarse for a particular detail of the observed dental feature, or by a scanned area that is too large .
[0007] Other holes can appear in digital impressions when the scanner is moved too quickly inside the mouth for correctly reconstructing the digital impression, for example if the scanner relies on a reconstruction software that needs images taken at different depths . Other holes can appear in digital impressions obtained by means of stereo cameras, in particular when a particular detail of the dental feature can be observed with only one camera : in fact, since stereo cameras systems rely on the combination of two or more images taken from different angles to reconstruct the digital impression, a hole appears where a detail is visible by only one camera . This often happens when scanning for example interdental spaces, abutments and inlays .
[0008] In any case, when the first obtained impression has one or more holes, further swipes of the scanner need to be performed to capture further images of the observeddental feature, for example from different angles, in order to fill the holes . This procedure is obviously time consuming.
[0009] As an alternative, geometric algorithms might be used on incomplete digital impressions to fill the holes . However, these algorithms are not accurate enough and require large computational resources, making it impossible to accurately complete digital impressions in real time .
[0010] An obj ect of the present invention is to overcome the above-mentioned drawbacks and in particular to provide a machine learning model specifically trained to rapidly complete an incomplete digital impression, therefore avoiding the need for a re-scanning of the observed dental feature, reducing the total time needed to obtain a complete digital impression, and even allowing to show the complete digital impression in real time to the operator .
[0011] Another obj ect of the present invention is to provide a computer-implemented method for estimating the three-dimensional structure of a dental feature, using the above-mentioned machine learning model .
[0012] Another obj ect of the present invention is to provide a device for estimating the three-dimensional structure of a dental feature, making use of the above-mentioned computer-implemented method.
[0013] Another obj ect of the present invention is to provide a method for producing a dental appliance suited for a patient, by first estimating the three-dimensional structure of a dental feature making use of the above-mentioned computer-implemented method.
[0014] These and other results are achieved according to thepresent invention providing a computer-implemented method for training a machine learning model according to claim 1, a computer-implemented method for estimating the three-dimensional structure of a dental feature according to claim 7, a device for estimating the three-dimensional structure of a dental feature according to claim 10, a method for producing a dental appliance according to claim 11, a computer program according to claim 12 and a computer program according to claim 13. Further preferred characteristics are the obj ect of the dependent claims .
[0015] The present invention will be now described, in an illustrative and non-limiting way.
[0016] In a first embodiment, the invention refers to a computer-implemented method for training a machine learning model to complete an incomplete digital impression of a dental feature .
[0017] First of all, this method comprises the step of providing a plurality of datasets, wherein each dataset comprises one or more incomplete digital impressions of the same dental feature . If the dataset comprises more than one incomplete digital impression, these incomplete digital impressions preferably have holes located in different positions . For example, one or more of the incomplete digital impressions of each dataset can be obtained by removing different portions from a corresponding complete digital impression.
[0018] Preferably, however, one or more of the incomplete digital impressions of the datasets are obtained taking incomplete digital impressions of a real dental feature of a real patient by means of a scanner, in particular in such a way that these incomplete digital impressionsdeliberately contain holes ( for example as a consequence of intrinsic limitations of the scanner) . In this way, the machine learning model will be trained with incomplete digital impressions corresponding to real situations, and the dimension of the datasets could be greatly reduced, expediting the training process and reducing the computational resources needed.
[0019] A complete digital impression of the corresponding dental feature is then associated to each dataset . For example, one or more of the complete digital impressions can be obtained taking a complete digital impression of a real dental feature of a real patient, by means of a scanner, in particular in such a way that these complete digital impressions do not contain holes . The digital impressions could be obtained with any suitable technique, including active wavefront sampling, "quasi" confocal microscopy, active or passive stereo cameras and monocular depth estimation. As described above, the incomplete digital impressions might then be obtained by removing different portions from the complete digital impression .
[0020] As an alternative, one or more of the complete digital impressions can be obtained digitizing a physical imprint of the dental feature, for example a physical imprint obtained with the traditional technique that uses a semi-solid impression material .
[0021] As a further alternative, one or more of the complete digital impressions can be obtained using CAD (computer-assisted design) . This alternative can be used for training the machine learning model also with dental features comprising metallic parts that are difficult to accurately scan, such as abutments . In this way, thetrained machine learning model will be able to complete an incomplete digital impression that comprises also these particular dental features, avoiding the need for a further post-processing step where the scanned abutment is compared with a library of known abutments . Once the corresponding complete digital impression is associated to each dataset, the machine learning is trained in a supervised way, on the basis of each dataset (which comprises incomplete digital impressions) associated with the corresponding complete digital impression, which acts as ground truth. In this way, the machine learning model is trained to associate to each incomplete digital impression of a dental feature a corresponding complete digital impression of the dental feature .
[0022] The resulting trained machine learning model is able to rapidly complete any incomplete digital impression provided, avoiding the need for a re-scanning of the observed dental feature, reducing the total time needed to obtain a complete digital impression, and even allowing to show the complete digital impression in real time to the operator .
[0023] The trained machine learning model could complete any incomplete digital impression in real time while the digital impression itself is taken with an intra-oral scanner, in order to provide the operator with a live complete estimated three-dimensional structure of the observed dental feature . As an alternative, the trained machine learning model could be used in a post processing step to complete any incomplete digital impression previously taken.
[0024] For example, the trained machine learning model, in use,may compare a new incomplete digital impression with the training incomplete digital impressions of the training datasets, determine which training incomplete digital impression is most similar to the new digital impression, and complete the new digital impression using the complete training digital impression associated to the most similar training incomplete digital impression. The present invention also relates to a computer-implemented method for estimating the three-dimensional structure of a dental feature, in which an incomplete digital impression of the dental feature is provided and completed using the trained machine learning model .
[0025] In an embodiment, the computer-implemented method completes the incomplete digital impression only in the non-reconstructed areas that could be completed by the trained machine learning model with sufficient accuracy. In general, in fact, an incomplete digital impression comprises a plurality on non-reconstructed areas (also known as "holes") corresponding for example to occluded areas, dark areas, areas of light reflection or areas blurred as a consequence of patient movement or scanner movement . Further, not all the holes are equally easy to complete, for example as a consequence of different dimensions and shapes of the holes .
[0026] In this regard, a completion accuracy estimating software could be used to determine an expected completion accuracy score for each hole in the incomplete digital impression, wherein such score estimates the accuracy with which the trained machine learning model is able to complete the hole . For example, the completion accuracy estimating software may include a completion accuracy estimating machine learning model, trainedusing incomplete digital impressions and corresponding completed digital impressions conveniently labelled on the basis of the accuracy of the completion as judged by a training operator .
[0027] Once the expected completion accuracy scores are determined for all the holes of the incomplete digital impression, the trained machine learning model completes only the holes whose expected completion accuracy scores are higher than a predefined completion accuracy score threshold. In this way, the trained machine learning model completes only the holes of the incomplete digital impression that could be completed with sufficient accuracy, improving the overall accuracy of the completed digital impressions .
[0028] In addition, it is possible to include particular technical constraints in the determination of the expected completion accuracy scores . For example, the completion accuracy estimating software may be built to assign slightly higher expected completion accuracy scores to holes which could be completed with reconstructed dental features which are slightly smaller than the corresponding real dental features, and to assign slightly lower expected completion accuracy scores to holes which could be completed with reconstructed dental features which are slightly bigger than the corresponding real dental features : in fact, when a dental crown is to be fabricated it is always preferable to underestimate the dimensions of the corresponding tooth than to overestimate them.
[0029] The present invention also relates to a device for estimating the three-dimensional structure of a dental feature, comprising a scanner configured to take adigital impression of the dental feature and a central control unit programmed to complete it using the trained machine learning model .
[0030] The present invention also relates to a method for producing a dental appliance tailored for a patient, for example dental braces, clear aligners, retainers, implants, dentures or other prosthetics . In particular, this method comprises the steps of estimating the three-dimensional structure of a dental feature inside the mouth of the patient, by using the trained machine learning model to complete an incomplete digital impression of the dental feature, and then creating digital data associated to the estimated three-dimensional structure, transferring the digital data to a manufacturing system, and manufacturing the dental appliance by means of the manufacturing system on the basis of the digital data . In this way, the dental appliance can be manufactured to be perfectly tailored for the patient .
[0031] The manufacturing system could be a CAD / CAM system or a 3D printing system.
[0032] The present invention has been described in an illustrative and non-limiting way according to relative favourite embodiments, but it is clear that the person skilled in the art could perform many variations and modifications, all of which are within the scope of the invention .
Claims
CLAIMS1 ) Computer-implemented method for training a machine learning model to complete an incomplete digital impression of a dental feature, the method comprising the following steps :providing a plurality of datasets, each dataset comprising one or more incomplete digital impressions of the same dental feature;associating to each dataset a complete digital impression of the corresponding dental feature;- training the machine learning model in a supervised way, on the basis of each dataset associated with the corresponding complete digital impression as ground truth .2 ) Method according to claim 1, wherein, for each dataset, at least one incomplete digital impression is obtained taking an incomplete digital impression of the dental feature by means of a scanner .3) Method according to any of claims 1 or 2, wherein, for each dataset, at least one incomplete digital impression is obtained removing a different portion of the corresponding complete digital impression.4 ) Method according to any of claims 1 to 3, wherein at least one complete digital impression is obtained taking a complete digital impression of the dental feature by means of a scanner .5) Method according to any of claims 1 to 4, wherein at least one complete digital impression is obtained digitizing a physical imprint of the dental feature . 6) Method according to any of claims 1 to 5, wherein at least one complete digital impression is obtained using CAD.7 ) Computer-implemented method for estimating the three-dimensional structure of a dental feature, comprising the following steps :- providing an incomplete digital impression of the dental feature;- completing the incomplete digital impression by means of a machine learning model trained according to any of claims 1 to 6 .8 ) Method according to claim 7, wherein the incomplete digital impression is completed in real time while it is being taken.9) Method according to claim 7, wherein the incomplete digital impression is completed as a post processing step performed on an incomplete digital impression already taken.10) Method according to any of claims 7 to 9, wherein the incomplete digital impression comprises a plurality of non-reconstructed areas, wherein a corresponding expected completion accuracy score is determined for each non-reconstructed area to estimate an accuracy with which the machine learning model is able to complete the non-reconstructed area, and wherein only the nonreconstructed areas for which the determined expected completion accuracy score is higher than a predefined completion accuracy score threshold are completed.11 ) Device for estimating the three-dimensional structure of a dental feature, comprising a scanner configured to take a digital impression of the dental feature and a central control unit programmed to carry out a method according to any of claims 7 to 10.12 ) Method for producing a dental appliance tailored for a patient, comprising the following steps :- estimating the three-dimensional structure of a dental feature inside the mouth of the patient, by means of a method according to any of claims 7 to 10;creating digital data associated to the estimated three-dimensional structure;transferring the digital data to a manufacturing system;- manufacturing the dental appliance by means of the manufacturing system on the basis of the digital data .13) Computer program comprising instructions which, when the program is executed by a computer, cause the computer to carry out the method of any of claims 1 to 6 .14 ) Computer program comprising instructions which, when the program is executed by a computer, cause the computer to carry out the method of any of claims 7 to 10.
