AI imaging diagnostic device and dental OCT imaging diagnostic device

Abstract

AI image diagnostic device (1) into which three-dimensional dental image data acquired by a dental OCT device are input and which analyzes the input three-dimensional dental image data, comprising: a model execution device (11) that sequentially inputs two-dimensional tomographic image data, which constitutes the 3D diagnostic target dental image data, into a trained model (10) in order to obtain, as an execution processing result of the trained model (10), lesion information data, which are data relating to a part that is identified as a characteristic part such as a lesion in the input tomographic image data, for each part of tomographic image data in the three-dimensional dental image data, and to detect lesions from the input three-dimensional dental image data using the obtained lesion information data.wherein the trained model (10) is constructed by training on three-dimensional dental image data of several subjects previously acquired by a dental OCT device.

Description

AI diagnostic imaging device and dental OCT diagnostic imaging device

[0001] The present invention relates to an AI image diagnostic device, and more particularly to an AI image diagnostic device that uses images captured by a dental OCT device and a dental OCT image diagnostic device.

[0002] Currently, MI treatment, proposed by the FDI (Fédération Internationale de Dentistry) in 2000, is becoming widely accepted in the dental world. MI stands for Minimal Intervention, and refers to minimally invasive caries treatment. Dental OCT (Optical Coherence Tomography) devices can obtain high-resolution, high-sensitivity tomographic images of teeth without X-ray exposure. For this reason, dental OCT imaging diagnostic devices are considered to be effective diagnostic devices for practicing MI.

[0003] For example, the OCT device described in Patent Document 1 has a measurement imaging mode that assumes the storage of high-resolution subject images and a preview imaging mode that quickly displays low-resolution subject images as real-time videos on a display device. Furthermore, the OCT device described in Patent Document 2 uses a two-dimensional scanning mechanism to switch the direction of the first scan on the light-irradiated surface of the tooth between a horizontal scan that scans horizontally (horizontally) and a vertical scan that scans vertically (vertically).

[0004] Patent No. 5827024 Patent No. 6712106

[0005] OCT devices are characterized by their high resolution and sensitivity, allowing them to image the inside of teeth. However, due to their high sensitivity, changes in the internal characteristics of teeth other than caries are also imaged, making diagnosis using OCT images difficult without an understanding of the structure of teeth and the characteristics of dental OCT images. In other words, only dentists with extensive experience in diagnosing with OCT images can effectively utilize OCT images for dental diagnosis and treatment. Meanwhile, to realize treatment for MI by providing early treatment before symptoms worsen or as a preventative measure, it is desirable to enable dentists, including those with little experience in diagnosing with OCT images, to effectively utilize OCT images for dental diagnosis and treatment.

[0006] The present invention has been made in consideration of the above-mentioned circumstances, and its objective is to provide an AI image diagnostic device and a dental OCT image diagnostic device that allow dentists to effectively utilize OCT images for dental diagnosis and treatment.

[0007] In order to solve the above problem, the AI ​​image diagnostic device of the present invention is an AI image diagnostic device that receives three-dimensional dental image data taken with a dental OCT device and analyzes the input three-dimensional dental image data, and by sequentially inputting each two-dimensional tomographic image data that constitutes the three-dimensional dental image data of the diagnosis target into a trained model, as a result of executing the trained model, injury and disease information data, which is data regarding areas in the input tomographic image data that are identified as characteristic parts of injuries, diseases, etc., is obtained for each tomographic image data of the three-dimensional dental image data, and injuries and diseases are detected from the input three-dimensional dental image data using the acquired injury and disease information data, and the trained model is characterized in that it has been constructed by learning from three-dimensional dental image data of multiple patients that were previously taken with a dental OCT device.

[0008] According to the present invention, dentists, including those with little experience in diagnosing with OCT images, can effectively utilize dental OCT imaging diagnostic devices during dental diagnosis and treatment. Furthermore, according to the present invention, by using AI image diagnosis and having a hygienist perform screening in advance with the dental OCT imaging diagnostic device, dentists can reduce their consultation time.

[0009] 1 is a functional block diagram of an AI image diagnostic device according to an embodiment of the present invention. FIG. 1 is a conceptual diagram of the operation of generating training data to be used for learning. FIG. 2 is a conceptual diagram of the operation of constructing a trained model in the learning stage. FIG. 3 is a flowchart showing the flow of image analysis processing of an AI image diagnostic device according to an embodiment of the present invention. FIG. 4 is a flowchart showing the flow of the trained model construction process. FIG. 5 is a diagram showing an example of training data. FIG. 6 is a diagram showing an example of training data. FIG. 7 is a diagram showing an example of training data. FIG. 8 is a diagram showing an example of tomographic image data of a diagnosis target. FIG. 9 is a schematic diagram of an analyzed image. FIG. 10 is a diagram showing an example of output injury and disease information data. FIG. 11 is a diagram showing input data and two output data of a second trained model. FIG. 12 is a diagram showing examples of tomographic image data of plane A of 3D image data acquired by a dental OCT device. FIG. 13 is a diagram showing an example of a 3D image displayed on a screen. FIG. 14 is a schematic diagram of a molar in the case of horizontal scanning. FIG. 15 is a schematic diagram of a molar in the case of vertical scanning. FIG. 16 is a diagram showing example injury and disease information data acquired by a first trained model. FIG. 17 is a graph showing an example of the relationship between the tomographic position and the similarity of a feature part. FIG. 18 is a diagram showing an area where the similarity exceeds a predetermined threshold. FIG. 19 is a diagram showing the detection result of injury and disease in three-dimensional dental image data. FIG. 20 is a diagram showing an image of a tooth displayed on a screen before transitioning. FIG. 1 is a diagram showing an image of a tooth displayed on a screen after a transition; FIG. 2 is a diagram showing input data and three output data of a second trained model; FIG. 3 is a diagram showing reconstructed volume data; FIG. 4 is a diagram showing an image of a tooth displayed on a screen; FIG. 5 is a diagram showing an image of a tooth with volume data superimposed; FIG. 6 is a hardware configuration diagram showing an example of a computer that realizes the functions of an AI image diagnostic device according to an embodiment of the present invention; FIG. 7 is a functional block diagram of a dental OCT image diagnostic device according to an embodiment of the present invention.

[0010] The present invention will be described in detail with reference to the accompanying drawings, in which the size and positional relationship of components shown in the drawings may be exaggerated for clarity.

[0011] [Configuration of AI Image Diagnostic Device] The configuration of an AI image diagnostic device according to an embodiment of the present invention will be described with reference to Fig. 1. The AI ​​image diagnostic device 1 is a device that receives three-dimensional tooth image data captured by a dental OCT device and analyzes the input three-dimensional tooth image data. Hereinafter, the three-dimensional tooth image data (volume data) captured by the dental OCT device will be referred to as an OCT 3D image. The OCT 3D image input to the AI ​​image diagnostic device 1 is a diagnostic image.

[0012] The AI ​​image diagnostic device 1 includes a model execution unit 11. The model execution unit 11 sequentially inputs each of the two-dimensional tomographic image data constituting the three-dimensional dental image data of the diagnosis target into the trained model 10, thereby acquiring injury / illness information data for each of the tomographic image data of the three-dimensional dental image data as a result of the execution process of the trained model 10, and detects injuries / illnesses from the input three-dimensional dental image data using the acquired injury / illness information data. The injury / illness information data is data related to areas identified as characteristic parts of injuries / illnesses, etc. in the input tomographic image data.

[0013] The trained model 10 is constructed by learning 3D dental image data of multiple examinees that were previously captured using a dental OCT device. As described below, the trained model 10 is constructed in a learning stage using training data. In the utilization stage of the trained model 10, the AI ​​image diagnostic device 1 performs image diagnosis based on the input OCT 3D images (diagnostic images). Note that, as will be described in detail later, the trained model 10 may output injury / illness information data by outputting numerical values ​​(position coordinates of injuries, etc.) or images (visualized positions of injuries, etc.).

[0014] The model execution unit 11 inputs one tomographic image data of an OCT 3D image (diagnostic image) into the trained model 10 and obtains, as an analysis result, injury and disease information data for the tomographic image data. If the model execution unit 11 detects multiple injuries or diseases in one tomographic image data, it obtains injury and disease information data for each of them. If the model execution unit 11 does not detect an injury or disease in one tomographic image data, it obtains injury and disease information data having information such as a similarity of 0, indicating that no injury or disease was detected. The model execution unit 11 inputs all tomographic image data of the OCT 3D image (diagnostic image) into the trained model 10 and obtains, as an analysis result, injury and disease information data corresponding to each tomographic image data. Note that a diagnostic image is composed of, for example, several hundred pieces of tomographic image data. The analysis result 13A shown in FIG. 1 indicates the analysis result for all of this tomographic image data. The detection result 13B is information on injuries or diseases detected by the model execution unit 11 from the diagnostic image based on the analysis result 13A. Since the analysis results 13A and detection results 13B differ depending on the output contents of the trained model 10, when explaining specific examples of trained models, the analysis results 13A and detection results 13B corresponding to those models will also be explained.

[0015] The AI ​​diagnostic imaging device 1 may include a display control unit 12 that functions as a viewer that displays an OCT 3D image (diagnostic image) on a display device 15, such as a liquid crystal display. The display control unit 12 receives information (hereinafter referred to as "selected information") determined or selected by a user from an input device 14, such as a mouse or keyboard operated by the user. The display control unit 12 functions as a viewer that displays a predetermined image based on the OCT 3D image (diagnostic image) and the detection results 13B on a display device 15, such as a liquid crystal display. Note that the selected information operated by the user and the predetermined image displayed on the display device 15 differ depending on the output content of the trained model 10. When describing specific examples of the trained model 10, the selected information and images corresponding to that model will also be described.

[0016] [Learning Phase] Next, the generation of training data and model construction as the learning phase will be described with reference to FIGS. 2 and 3. FIG. 2 is a conceptual diagram of the process of generating training data used for learning. The computer 20 shown in FIG. 2 includes a central processing unit (CPU) 21 and a storage device 22 and functions as a viewer. The CPU 21 operates based on a program stored in the storage device 22 and controls an input device 24 and a display device 25. The storage device 22 includes a read-only memory (ROM), a random access memory (RAM), a hard disk drive (HDD), etc. The storage device 22 stores various programs, such as software required for labeling tomographic image data. The input device 24 is used to input user instructions and includes a mouse, keyboard, etc. The display device 25 displays tomographic images and includes, for example, a liquid crystal display. Windows, icons, buttons, etc. are displayed on the screen of the display device 25, and an operator (dentist) can select them using the input device 24.

[0017] The operator (dentist) uses a viewer (computer 20) to search for images of lesions (characteristic features of lesions) while reviewing tomographic images of teeth (OCT 3D images) captured by a dental OCT device. The operator (dentist) labels the tomographic images in which characteristic features of lesions are found to classify the characteristic features of lesions. That is, the operator (dentist) creates and saves training data by adding labels created by inputting, for example, the name of the lesion or a label selected from multiple pre-prepared labels to the tomographic image data. This labeling process is called annotation. The training data 30 shown in FIGS. 2 and 3 schematically represents a collection of labeled tomographic images, for example, a collection of 1,000 tomographic images.

[0018] The training data can include at least one of tomographic image data of an A-section, tomographic image data of an L-section, tomographic image data of an S-section, en-face image data, and 3D image data consisting of several consecutive frames of tomographic image data. Here, A, L, and S represent different cross-sectional directions. Specifically, the A-section is a cross-section parallel to a plane defined by both the B-axis direction, which is perpendicular to the A-axis direction, which is the irradiation direction of the OCT laser light on the tooth, and the A-axis direction. The L-section is a cross-section parallel to a plane defined by both the V-axis direction, which is perpendicular to the A-axis direction and the B-axis direction, and the A-axis direction. The S-section is a cross-section parallel to a plane defined by both the B-axis direction and the V-axis direction. The en-face image data is image data that combines information about the surface of the tooth irradiated with the OCT laser light and information about the A-axis direction. This en-face image data also combines internal information that is not normally visible on the outer surface. Furthermore, the training data preferably includes at least tomographic image data of an A-section and tomographic image data of an L-section. For each lesion, there are three cross-sectional images: A-section, L-section, and S-section. These three tomographic image data can be labeled and all three can be used as training data.

[0019] The name of the injury or disease is input as a label to be added to the training data. The characteristic part of the injury or disease in the tomographic image (hereinafter referred to as the injury or disease characteristic part) is an image of the injury or disease, such as an image showing at least one injury or disease among early caries (Ce), caries (C1 or more), secondary caries, root caries, cracks, fractures, and attrition.

[0020] In addition to the name of the injury or disease, the type of tooth may also be input into the label added to the training data. For example, tooth types can be broadly classified into molars, incisors, and canines. From another perspective, tooth types can also be classified into permanent teeth and baby teeth. Dental formulas can also be used to represent these tooth types. For example, with a two-digit FDI (Flat-Digit Imaging) system, simply entering a two-digit number makes it possible to identify whether a tooth is a permanent or baby tooth, and whether it is a molar, incisor, or canine. Because teeth have different external shapes and internal structures depending on the type, inputting the type of tooth into the label added to the training data makes it possible to more accurately determine whether a tooth is injured or diseased.

[0021] The type of image data may also be input into the label added to the training data. Examples of image data types include A-section tomographic images, L-section tomographic images, S-section tomographic images, en-face images, and 3D images. These can be identified by two letters or less, such as A, L, S, en, and 3D.

[0022] Furthermore, names other than diseases may be entered into the labels added to the training data. In this case, when the operator (dentist) finds, for example, an image of dental plaque as a characteristic feature other than diseases in the tomographic image (hereinafter referred to as a non-disease characteristic feature), the operator enters "plaque" as a name other than diseases into the label. A non-disease characteristic feature in a tomographic image is an image other than diseases, such as an image showing at least one of metal, ceramic, resin, dental plaque, and saliva bubbles. Although these are images other than diseases, entering a name other than diseases into the labels added to the training data makes it possible to more accurately determine the presence or absence of diseases.

[0023] FIG. 3 is a conceptual diagram of the work of constructing a trained model in the learning stage. The model builder 40 shown in FIG. 3 includes a CPU 41 and a storage device 42 and functions as a learner. The CPU 41 operates based on a model construction processing program 43 stored in the storage device 42. The model construction processing program 43 is a schematic representation of a program that executes a machine learning algorithm such as a neural network. A trained model 10 can be constructed by inputting training data 30 into the model builder 40 and training it. When creating the training data 30, the operator (dentist) finds feature areas (images of injuries, etc.) in the tomographic images and labels them. In response, the trained model finds feature areas (images of injuries, etc.) in the input tomographic images and outputs injury information data.

[0024] [Operation of AI Image Diagnostic Device] Next, the flow of image analysis processing by the AI ​​image diagnostic device 1 will be described with reference to FIG. 4 (and also with reference to FIG. 1 as appropriate). FIG. 4 is a flowchart showing the flow of image analysis processing by the AI ​​image diagnostic device 1. As shown in FIG. 4, first, a dental OCT device (not shown) acquires a diagnostic image (volume data) (step S1). Then, the AI ​​image diagnostic device 1 inputs one piece of tomographic image data from the diagnostic image to the trained model 10 via the model execution unit 11 (step S2). Note that the tomographic image data from the diagnostic image may be acquired from an external device each time, or may be stored in the storage device of the AI ​​image diagnostic device 1.

[0025] The trained model 10 then analyzes the tomographic images and outputs, as analysis results, information such as the name, center coordinates, and similarity of the lesions found in the tomographic images (step S3). The model execution unit 11 then acquires the analysis results (step S4). The model execution unit 11 then determines whether all tomographic images have been analyzed (step S5). If not all tomographic images have been analyzed (step S5: No), the model execution unit 11 returns to step S2. On the other hand, if all tomographic images have been analyzed (step S5: Yes), the model execution unit 11 determines the lesion detection results 13B in the diagnostic image (volume data) based on the analysis results 13A acquired so far (step S6). If the AI ​​image diagnostic device 1 includes a display control unit 12, the AI ​​image diagnostic device 1 receives user selection information from the input device 14. The display control unit 12 then searches the diagnostic image (volume data) for the lesion selected by the user based on the detection results 13B and displays the tomographic image containing the lesion on the display device 15 (step S7).

[0026] [Trained Model Construction Process] Next, as an example, a process flow in which an operator generates training data and constructs a trained model as a series of tasks will be described with reference to FIG. 5 (and also with reference to FIGS. 2 and 3 as appropriate). FIG. 5 is a flowchart showing the process flow for constructing a trained model. As shown in FIG. 5, first, the computer 20 functioning as a viewer acquires an OCT image (step S11). This OCT image is raw data of a tomographic image in an OCT 3D image captured in advance to generate training data. This tomographic image data may be data acquired each time from an external device or data stored in the storage device 22 of the computer 20. Then, the operator (dentist) searches for lesions in the tomographic image while checking the tomographic image using the viewer (step S12). Then, for the tomographic image in which a lesion or disease is found, the operator (dentist) inputs a label classifying each lesion and the tooth type (dental formula), and saves the labeled tomographic image data as training data 30 (step S13). The operator then inputs the training data 30 into the model builder 40 to train the model (step S14). When model construction is complete (step S15: Yes), the model at that time is constructed as the trained model 10, and the process ends. On the other hand, if model construction is not complete (step S15: No), the process returns to step S11. The process of step S14 is repeated until training is completed for, for example, 1,000 labeled tomographic images.

[0027] [Method of outputting trained model] Hereinafter, the trained model 10 that outputs numerical values ​​(position coordinates of injuries, illnesses, etc.) as injury and illness information data will be referred to as the first trained model. Also, the trained model 10 that outputs images (visualized positions of injuries, illnesses, etc.) as injury and illness information data will be referred to as the second trained model. The first trained model and the second trained model will be described in order.

[0028] <First Trained Model> The first trained model is constructed by performing machine learning using tomographic image data containing disease characteristic features from tomographic image data constituting three-dimensional tooth image data previously captured with a dental OCT device as first training data, searching for disease characteristic features from the input tomographic image data, and outputting disease information data. The disease information data includes at least the name of the characteristic feature, a central position indicating the coordinates of the point with the highest similarity to the characteristic feature, which is assigned to each point constituting the image in the tomographic image data, and a similarity assigned to the point indicating the central position. The "name of the characteristic feature" in the disease information data is, for example, incipient caries or caries, and is the same as the name input as the label added to the training data. The "similarity" is the degree of similarity to the image of the disease or injury and indicates the probability of matching the disease or injury.

[0029] In the learning phase, training data 30 is input to the model builder 40 (FIG. 3), and the model is trained to construct a first trained model (trained model 10). FIGS. 6A to 6C are diagrams showing examples of training data used in the learning phase. The training data shown in FIG. 6A has oval marks 101 and 102 added to a tomographic image. Mark 101 was added by a dentist during the training data generation phase to surround a lesion found in a tomographic image of a molar, indicating that the dentist has entered the label "caries" for the found lesion. Mark 102 was added to a similarly found lesion, indicating that the dentist has entered the label "early caries." The training data shown in FIG. 6B is an example of an image other than a lesion. Mark 103 was added to an image other than a lesion found in a tomographic image of a molar, indicating that the dentist has entered the label "resin." The training data mark 104 shown in FIG. 6C is added to the lesion found in the tomographic image of the front tooth, and indicates that the label "root caries" has been input.

[0030] Next, the input and output data of the first trained model in the utilization phase will be described with reference to Figures 7A to 7C. Figure 7A is a diagram showing an example of tomographic image data of a diagnosis target. Figure 7B is a schematic diagram of an image analyzed when the tomographic image 110 shown in Figure 7A is input to the first trained model. Figure 7C shows an example of injury and disease information data output from the first trained model at this time. The conceptual diagram in Figure 7B is an enlarged view of the tomographic image 110 shown in Figure 7A, showing areas 111 and 112 where images estimated to be injuries or diseases were detected, as well as horizontal (X-axis) and vertical (Y-axis) axes. The scales on the vertical and horizontal axes indicate position coordinates. As an example, the total number of tomographic positions constituting the diagnostic image is 400. The injury and disease information data shown in Figure 7C indicates that the image in area 111 is determined to be incipient caries with a similarity of 70%, and the coordinates (310, 195) of the point with the highest incipient caries similarity are determined as the center position. Similarly, the image in area 112 indicates that it has been determined to be a carious object with a similarity of 90%.

[0031] If the first trained model finds multiple injuries or diseases in one piece of tomographic image data, it outputs the injury or disease information data for each injury or disease. If the first trained model finds the same type of injury or disease in a diagnostic image, it assigns a distinguishing name to each. For example, caries may be distinguished as "Caries_01" and "Caries_02." The first trained model can determine an injury or disease by searching within a single piece of tomographic image data, but it can also determine an injury or disease from the results of searching several consecutive pieces of tomographic image data. The model execution unit 11 of the AI ​​image diagnostic device 1 acquires injury or disease information data from the first trained model.

[0032] Although the above-described first trained model is constructed by performing machine learning using tomographic image data containing lesion characteristic portions as first training data, this is not limiting. The first trained model may also be constructed by performing machine learning using tomographic image data containing lesion characteristic portions and tomographic image data containing non-lesion characteristic portions as training data. In this variation, the first trained model uses, as second training data, tomographic image data containing lesion characteristic portions and tomographic image data containing non-lesion characteristic portions, which are tomographic image data constituting three-dimensional tooth image data previously captured by a dental OCT device. In this case, the first trained model is constructed by performing machine learning to search for lesion characteristic portions and non-lesion characteristic portions from input tomographic image data and output injury information data.

[0033] <Second Trained Model> The second trained model was constructed by performing machine learning using tomographic image data containing lesion features from 3D tooth image data previously captured with a dental OCT device as first training data. The model searches for lesion features from the input tomographic image data, generates an image representing a degree of match, which is image data in which the similarity between each point constituting the image in the tomographic image data and the feature is replaced with a pixel value, and visualizes and outputs the position of the feature. The injury / illness information data includes at least the name of the feature and an image representing the degree of match. The image representing the degree of match is image data in which the similarity between the injury / illness is replaced with a pixel value (a numerical value representing brightness). The second trained model increases (brightens) the pixel value as the similarity increases. In other words, the image representing the degree of match is an image in which parts estimated to be lesions are bright and parts not estimated to be lesions are dark.

[0034] In the learning stage, training data 30 is input to a model builder 40 (FIG. 3), and the model is trained to construct a second trained model (trained model 10). FIGS. 6A to 6C show examples of training data used in the learning stage.

[0035] Next, the input and output data of the second trained model in the utilization stage will be described with reference to Fig. 8 . The tomographic image 120 shown in Fig. 8 shows an example of tomographic image data of a diagnosis target. The images showing the degree of match output when this tomographic image 120 is input to the second trained model are the two images shown to the right of the tomographic image 120. The output image arranged at the top in Fig. 8 is an image showing the degree of match of early caries 121, and the output image arranged at the bottom is an image showing the degree of match of caries 122.

[0036] The second trained model can determine injury or disease by searching within a single piece of tomographic image data, but can also determine injury or disease from the results of searching several consecutive pieces of tomographic image data.

[0037] The model execution unit 11 of the AI ​​image diagnostic device 1 determines, as selection information, the names of feature parts included in the injury / illness information data based on the injury / illness information data for each slice of image data in the 3D dental image data obtained from the second trained model.The model execution unit 11 then generates 3D images of the feature parts by reconstructing images representing the degree of match included in the injury / illness information data for each feature part name.The model execution unit 11 then sets the generated 3D images of the feature parts as the injury / illness detection results 13B in the 3D dental image data.

[0038] When replacing the similarity with the feature with a pixel value, the second trained model may convert the pixel values ​​of pixels whose similarity is less than a preset threshold to 0 to generate an image representing the degree of match. In this case, the model execution unit 11 of the AI ​​image diagnostic device 1 acquires from the second trained model an image in which the pixel values ​​of pixels whose similarity is less than the threshold have been converted to 0 as an image representing the degree of match. By doing so, an image in which all areas of a tomographic image in which an injury or disease is found that have low similarity to the injury or disease are completely black is output as an image representing the degree of match. This makes the image representing the degree of match easier to see.

[0039] Although the second trained model has been described as being constructed by machine learning using tomographic image data containing lesion characteristic portions as first training data, this is not limiting. The second trained model may also be constructed by machine learning using tomographic image data containing lesion characteristic portions and tomographic image data containing non-lesion characteristic portions as training data. In this variation, the second trained model uses tomographic image data constituting three-dimensional dental image data previously captured with a dental OCT device, the tomographic image data containing lesion characteristic portions and the tomographic image data containing non-lesion characteristic portions, as second training data. In this case, the second trained model is constructed by machine learning to search for lesion characteristic portions and non-lesion characteristic portions in the input tomographic image data, and to generate an image representing a degree of similarity, which is image data in which the similarity between each point constituting the image in the tomographic image data and the corresponding feature is replaced with a pixel value, thereby visualizing and outputting the position of the feature portions.

[0040] Example 1 Next, an example (Example 1) will be described in which the AI ​​image diagnostic device 1 detects an injury or disease in an input OCT 3D image (diagnostic image) using the first trained model. The model execution unit 11 of the AI ​​image diagnostic device 1 according to Example 1 graphs the relationship between the tomographic position and the similarity of the feature for each feature based on the injury or disease information data for each tomographic image data of the 3D dental image data acquired from the first trained model. The model execution unit 11 then identifies regions on the graph where the similarity exceeds a predetermined threshold at consecutive tomographic positions. The model execution unit 11 then determines a name for each region as selection information. For each named region, the model execution unit 11 then associates the name of the feature, the tomographic position having the point with the highest similarity within the region, and the center position of the tomographic position having the point with the highest similarity, thereby generating an injury or disease detection result for the 3D dental image data.

[0041] When the AI ​​image diagnostic device 1 according to Example 1 includes a display control unit 12, the name of an area, which is selection information, is input by a user operation to the display control unit 12. Then, the display control unit 12 extracts tomographic image data corresponding to the input name of the area from the input three-dimensional dental image data based on the injury / disease detection result 13B generated by the model execution unit 11, and displays a dental image including the extracted tomographic image on the display device 15.

[0042] Specifically, the model execution unit 11 of the AI ​​image diagnostic device 1 according to Example 1 inputs multiple tomographic image data of A-sections in 3D image data acquired by a dental OCT device into the first trained model in order from front to back (in the anterior-posterior direction). FIG. 9 is a diagram showing an example of tomographic image data of A-sections input sequentially into the first trained model. The tomographic image data shown in FIG. 9 and the training data were captured in a measurement imaging mode that assumes the storage of high-resolution subject images in the dental OCT device. The model execution unit 11 may input all tomographic image data of A-sections sequentially into the first trained model, or may input every few images into the first trained model. In this manner, Example 1 can process 3D image data at high speed.

[0043] Fig. 10A shows an example of a screen displaying a 3D image captured by a dental OCT device. The image on the left in Fig. 10A is a 3D image. The four images shown divided into four on the right side of Fig. 10A are a tomographic image of the A cross section (upper left), a tomographic image of the L cross section (upper right), a tomographic image of the S cross section (lower right), and an en-face image (lower left).

[0044] If the images used as training data in the learning stage include a tomographic image of the A section and a tomographic image of the L section, in the usage stage the user can select either a tomographic image constructed by horizontal scanning or a tomographic image constructed by vertical scanning, thereby improving usability.

[0045] Horizontal scanning and vertical scanning will be described with reference to FIGS. 10B and 10C . The dental OCT device irradiates a tooth with laser light to acquire A-scan data 131, which has multiple data acquisition points, line by line. During horizontal scanning, the dental OCT device writes the multiple A-scan data 131 into a 3D coordinate space along the B-axis direction, as shown in FIG. 10B , to obtain tomographic image data 132A of the A cross section. The dental OCT device then constructs 3D image data 130 by stacking the tomographic image data 132A in the V-axis direction. During vertical scanning, the dental OCT device writes the multiple A-scan data 131 into a 3D coordinate space along the V-axis direction, as shown in FIG. 10C , to obtain tomographic image data 132L of the L cross section. The dental OCT device then constructs 3D image data 130 by stacking the tomographic image data 132L in the B-axis direction. For example, when a dentist diagnoses root caries or gum disease, it is more suitable to use tomographic image data constructed by vertical scanning rather than horizontal scanning. Details are described in Patent Document 2, so further explanation will be omitted.

[0046] The model execution unit 11 of the AI ​​image diagnostic device 1 according to Example 1 accepts an instruction for either a horizontal scan or a vertical scan, for example, via a user operation. At this time, when the model execution unit 11 accepts an instruction for a horizontal scan, it inputs tomographic image data of the A-plane into the first trained model and executes the computational processing of the first trained model. Furthermore, when the model execution unit 11 accepts an instruction for a vertical scan, it inputs tomographic image data of the L-plane into the first trained model and executes the computational processing of the first trained model. Alternatively, the model execution unit 11 may determine the scan direction based on the input information. In this case, the model execution unit 11 determines, for example, information on the scan direction when the image was captured with a dental OCT device, which is included in the input three-dimensional tooth image data. At this time, when the model execution unit 11 determines that the input three-dimensional tooth image data includes horizontal scan information, it inputs the tomographic image data of the A-plane into the first trained model and executes the computational processing of the first trained model. Furthermore, when the model execution unit 11 determines that the input 3D dental image data contains longitudinal scan information, it inputs L-section tomographic image data into the first trained model and executes the calculation process of the first trained model. The first trained model finds injuries and diseases from the input tomographic image data and outputs injury and disease information data (feature name, center position, similarity). Figure 11A shows an example of injury and disease information data obtained by analyzing all tomographic image data with the first trained model, assuming a total of 400 tomographic positions constituting a diagnostic image.

[0047] 11A shows injury and disease information data for the 20th, 100th, 200th, and 300th tomographic image data out of a total of 400 tomographic positions. In this example, early caries, caries, and cracks were found as injuries and diseases throughout all the tomographic image data, but these injuries and diseases were not found in, for example, the 20th tomographic image data.

[0048] Figure 11B is an example of a graph obtained by the model execution unit 11 of the AI ​​image diagnostic device 1 according to Example 1, based on the injury and disease information data for all of the tomographic image data, to determine the relationship between the tomographic position and the similarity of each feature. In Figure 11B, the horizontal axis indicates the tomographic position, and the vertical axis indicates the similarity (%). The thick line indicates early caries (Ce), the dashed line indicates caries (C1 or higher), and the dotted line indicates cracks. The dashed and dotted line graphs have a single peak. In contrast, the thick line (caries) graph has two peaks, indicating that caries may have been discovered in two locations.

[0049] As shown in FIG. 12A , the model execution unit 11 of the AI ​​image diagnostic device 1 according to Example 1 determines regions a1, a2, a3, and a4 where the similarity exceeds a predetermined threshold at consecutive cross-sectional positions on the graph. The model execution unit 11 then determines a name for each of the determined regions as selection information. In this example, the similarity threshold is set to 60, and the boundaries of the regions are determined. The name of region a1 is set to "Incipient Caries_1," the name of region a2 is set to "Caries_1," the name of region a3 is set to "Crack_1," and the name of region a4 is set to "Incipient Caries_2." These region names are used as selection information in the processing of the display control unit 12 of the AI ​​image diagnostic device 1 according to Example 1.

[0050] The model execution unit 11 of the AI ​​image diagnostic device 1 according to Example 1 determines a representative point for each region whose name has been determined. The representative point of a region is characterized by the cross-sectional position of the point in the region with the highest similarity and the center position of the cross-sectional position of the point with the highest similarity. The model execution unit 11 generates a disease detection result 13B (see FIG. 1 ) in the three-dimensional dental image data by associating the representative point of the region with the name of the region (selection information). FIG. 12B shows an example of the detection result 13B. In FIG. 12B , for example, data with the name "Incipient Caries_1" indicates that the center position of the incipient caries in region a1 is at coordinates (152, 102) in the cross-sectional image data of cross-sectional position 100.

[0051] When the AI ​​image diagnostic device 1 according to Example 1 includes a display control unit 12, OCT 3D image data (diagnostic image) and the detection result 13B (representative point data of the lesion area) shown in FIG. 12B are input to the display control unit 12 (OCT 3D viewer). The display control unit 12 of Example 1 then displays the diagnostic image on the display device 15, and lists the names of the areas as selection information and displays them on the display device 15. The names of the displayed areas are determined by the model execution unit 11 based on the names of the characteristic features (names of the lesions) in the lesion information data output by the first trained model. FIG. 13A shows the names of the areas and an image of a tooth displayed on the initial screen. In the viewer displaying OCT images, two orthogonal lines indicating the cross-sections are displayed on the A-plane, L-plane, and S-plane cross-sectional images, respectively, and the cross-sections are displayed three-dimensionally in the 3D image. Then, by changing the position of the lines indicating the cross-sections in the cross-sectional images, the tomographic image data of the desired cross-sectional position can be easily extracted. In FIG. 13A, the names of four regions are displayed on the right side of the tooth image.

[0052] The user can view this initial screen and use the input device 14, such as a mouse, to select (click) the name of the region (selection information), for example, "Caries 1." In this case, the display control unit 12 of the first embodiment searches for the tomographic position and center position of the region of the lesion (Caries 1) associated with the name of the region from the detection result 13B (representative point data of the lesion region), and displays a tomographic image in which the cut plane is moved to the tomographic position and center position obtained by the search, as shown in FIG. 13B. FIG. 13B is a diagram showing an image of a tooth displayed on the screen after transitioning from the initial screen of FIG. 13A. In FIGS. 13A and 13B, reference numeral 141 indicates the display area of ​​the tomographic image of the A-plane. The tomographic image of the A-plane in FIG. 13B is a tomographic image in which a caries is present, unlike the tomographic image of the A-plane in FIG. 13A (initial image). Similarly, tomographic images in FIG. 13B also show the presence of a caries in the tomographic images of the L-plane and the S-plane.

[0053] The user may be notified of which injury or illness the image being displayed on the display device 15 is of. On the display screen in Fig. 13B, the frame of the button 142 for the name of the area (selected information) selected by the user is highlighted with a thick line, but the color of the selected button may also be changed.

[0054] In this embodiment (Example 1), the OCT 3D image is described as being input as an OCT 3D image acquired in the measurement imaging mode of the dental OCT device. However, instead, an OCT 3D image acquired in the preview imaging mode of the dental OCT device can also be input. The dental OCT device described in Patent Document 1 displays a subject image as a real-time video on a display device at high speed in the preview imaging mode. Therefore, for example, the AI ​​image diagnostic device 1 according to Example 1 can acquire 3D image data (data in a video memory) being captured in the preview imaging mode from the dental OCT device. In this case, the AI ​​image diagnostic device 1 according to Example 1 can similarly acquire injury / disease information data from the first trained model by sequentially inputting the tomographic image data constituting the acquired 3D image data being captured into the first trained model by the model execution unit 11.

[0055] [Example 2] Next, an example (Example 2) will be described in which the AI ​​image diagnostic device 1 uses the second trained model to detect an injury or disease in an input OCT 3D image (diagnostic image). The model execution unit 11 of the AI ​​image diagnostic device 1 according to Example 2 inputs tomographic image data of multiple A-planes in 3D image data acquired by a dental OCT device to the second trained model in order from front to back (front-to-back direction). Figure 9 is a diagram showing an example of tomographic image data of A-planes input in order to the second trained model. Note that the model execution unit 11 of Example 2 may input all tomographic image data of A-planes in order to the second trained model, or may input every few images to the second trained model. In this way, Example 2 can process 3D image data at high speed.

[0056] The model execution unit 11 of the AI ​​image diagnostic device 1 according to Example 2 receives an instruction for either a horizontal scan or a vertical scan, for example, through a user operation. When the model execution unit 11 receives an instruction for a horizontal scan, it inputs tomographic image data of the A-section into the second trained model and executes the computational processing of the second trained model. When the model execution unit 11 receives an instruction for a vertical scan, it inputs tomographic image data of the L-section into the second trained model and executes the computational processing of the second trained model. Alternatively, the model execution unit 11 may determine the scan direction based on the input information. In this case, the model execution unit 11 determines, for example, information about the scan direction when the image was captured by the dental OCT device, contained in the input three-dimensional tooth image data. When the model execution unit 11 determines that the input three-dimensional tooth image data contains horizontal scan information, it inputs the tomographic image data of the A-section into the second trained model and executes the computational processing of the second trained model. Furthermore, when the model execution unit 11 determines that the input three-dimensional dental image data contains longitudinal scan information, it inputs the L-section tomographic image data to the second trained model and causes the second trained model to execute arithmetic processing. The second trained model finds injuries and diseases from the input tomographic image data and outputs injury and disease information data (the names of characteristic parts and an image showing the degree of match).

[0057] The model execution unit 11 of the AI ​​image diagnostic device 1 according to Example 2 analyzes all tomographic image data, for example, with a total of 400 tomographic positions constituting the diagnostic image, using the second trained model. Furthermore, the model execution unit 11 assigns similarity to feature parts to each point constituting the image in each tomographic image data, and detects lesions such as early caries, caries, and cracks throughout the entire tomographic image data. The names of these discovered feature parts are used as selection information in the processing of the display control unit 12 of the AI ​​image diagnostic device 1 according to Example 2.

[0058] For example, when the tomographic image 150 shown in Fig. 14A is input, the second trained model outputs an image showing the degree of match for each feature. That is, the second trained model outputs three images showing the degree of match: an image showing the degree of match for early caries 151, an image showing the degree of match for caries 152, and an image 153 showing the degree of match for cracks. Note that in Fig. 14A, image 153 showing the degree of match for cracks is simply a completely black image because no cracks were found in the input tomographic image 150.

[0059] The second trained model analyzes all tomographic image data and generates, for each name of injury or disease (each name of feature), images showing the degree of match equal to the total number of tomographic positions. The model execution unit 11 of the AI ​​image diagnostic device 1 according to Example 2 acquires images showing the degree of match from the second trained model and reconstructs the images showing the degree of match for each name of injury or disease (each name of feature) to generate three-dimensional image data. For example, when the model execution unit 11 reconstructs an image showing the degree of match of caries for all tomographic image data, it can generate volume data 154 (see Figure 14B) of caries only. The three-dimensional image of the feature generated by reconstruction is an example of the injury or disease detection result 13B (see Figure 1) in the input OCT 3D image data (diagnostic image).

[0060] When the AI ​​image diagnostic device 1 according to Example 2 is equipped with a display control unit 12, OCT 3D image data (diagnostic image) and volume data (detection result 13B) as shown in FIG. 14B are input to the display control unit 12 (OCT 3D viewer). As a result, the display control unit 12 according to Example 2 displays the diagnostic image on the display device 15, and also lists the names of the feature parts as selection information and displays them on the display device 15. The names of the feature parts displayed are the names of the feature parts detected by the second trained model. FIG. 15A is a diagram showing the names of the feature parts and an image of a tooth displayed on an initial screen. In FIG. 15A, three names of the feature parts are displayed to the right of the image of the tooth.

[0061] The user can view this initial screen and use the input device 14, such as a mouse, to select (click), for example, "dental caries" as the name (selection information) of the characteristic part. In this case, the display control unit 12 of Example 2 superimposes a three-dimensional image of the characteristic part (dental caries) on the OCT 3D image data (diagnostic image) as shown in FIG. 15B to visualize the caries. FIG. 15B is a diagram showing an image of a tooth displayed on the screen after transitioning from the initial screen of FIG. 15A. Reference numeral 161 indicates the display area of ​​the tomographic image of the A plane. FIG. 15B shows that volume data 154 (see FIG. 14B) of only caries is superimposed on the tooth in the 3D image display area. At this time, the display control unit 12 of Example 2 also superimposes the volume data of only caries on the tomographic image. That is, the display control unit 12 superimposes and displays images representing the degree of match on the tomographic image of the A plane, the tomographic image of the L plane, and the tomographic image of the S plane.

[0062] In the example shown in Fig. 15B, one feature is displayed on one screen, but multiple feature may be displayed on one screen, or each feature may be color-coded. The user may be notified of which injury or illness the three-dimensional image of the feature being displayed on the display device 15 represents. In the screen shown in Fig. 15B, the frame of the button 162 for the name of the feature (selected information) selected by the user is highlighted with a thick line, but the color of the selected button may be changed.

[0063] According to Example 2, by superimposing a 3D image of the characteristic area of ​​the lesion on the OCT 3D image data, the location and extent of the lesion can be easily grasped visually and quantitatively (area, volume). Furthermore, if a second trained model is constructed by inputting "plaque" as a name other than the lesion into the labels added to the training data during the learning phase, Example 2 can superimpose a 3D image of the characteristic area of ​​the plaque on the OCT 3D image data of the teeth. Generally, in tooth brushing instruction, after a patient brushes their teeth, the plaque is stained with a plaque staining agent, and the patient receives instruction while checking for areas that have not been brushed. Care must be taken because the staining agent will not come off if it gets on clothing. In contrast, Example 2 visualizes and quantitatively captures the state of plaque adhesion, allowing it to be used for tooth brushing instruction without the use of a staining agent.

[0064] [Hardware Configuration] Next, a hardware configuration showing an example of a computer that realizes the functions of the AI ​​diagnostic imaging device 1 according to this embodiment will be described with reference to Fig. 16. The computer 200 has a CPU 201, a storage device 202, an input / output IF (Interface) 203, and an IF 204 for image data input.

[0065] The CPU 201 operates based on a program stored in the storage device 202, and performs control by a control unit (the model execution unit 11 and the display control unit 12 shown in FIG. 1 ). The CPU 201 controls an input device 14 such as a mouse or a keyboard, and a display device 15 such as a liquid crystal display, via an input / output IF 203. The CPU 201 acquires data from the input device 14 via the input / output IF 203, and outputs generated data to the display device 15. Note that a GPU (Graphics Processing Unit) or the like may be used as a processor together with the CPU 201.

[0066] The storage device 202 includes a ROM, a RAM, a HDD, etc. The storage device 202 stores a boot program executed by the CPU 201 when the computer 200 starts up, programs related to the hardware of the computer 200, etc. The storage device 202 stores programs executed by the CPU 201 (trained model execution program, viewer program), data used by the programs, etc.

[0067] The image data input IF 204 includes a communication IF, a media IF, etc. The image data input IF 204 receives image data from another device via a communication network and outputs the image data to the CPU 201, for example, or reads image data stored on a recording medium and outputs the image data to the CPU 201 via the storage device 202. For example, when the computer 200 functions as the AI ​​image diagnostic device 1 according to the embodiment, the CPU 201 realizes the functions of the AI ​​image diagnostic device 1 by executing a trained model execution program and a viewer program loaded onto the RAM.

[0068] [Dental OCT Imaging Diagnostic Apparatus] Next, a dental OCT imaging diagnostic apparatus according to an embodiment of the present invention will be described with reference to FIG. 17 . The dental OCT imaging diagnostic apparatus 2 is configured by mounting an AI imaging diagnostic apparatus 1 on a dental OCT device 3. The AI ​​imaging diagnostic apparatus 1 has already been described, so its description will be omitted. The dental OCT device 3 includes various components necessary for capturing dental images. For example, the dental OCT device 3 includes an OCT light source and respective optical systems for measurement light and reference light branched from a laser beam emitted from the OCT light source. Furthermore, the dental OCT device 3 includes a two-dimensional scanning mechanism for scanning the measurement light, a detector for detecting light resulting from the combination of reflected light from the tooth and the reference light, a controller, and the like. Such a configuration is well known, and the dental OCT device 3 can be, for example, an OCT device described in Patent Document 1 or Patent Document 2. The dental OCT imaging diagnostic apparatus 2 enables a consistent process from imaging to image diagnosis.

[0069] The above describes an AI imaging diagnostic device and a dental OCT imaging diagnostic device according to embodiments of the present invention. However, the scope of the present invention is not limited to these descriptions and should be broadly interpreted based on the claims. It goes without saying that various modifications and alterations based on these descriptions are also within the scope of the present invention. For example, while the AI ​​imaging diagnostic device 1 in FIGS. 1 and 17 is shown to include a display control unit 12 and a display device 15, neither is a required component. If a viewer is provided separately from the AI ​​imaging diagnostic device, the viewer can display OCT 3D images (diagnostic images) input to the AI ​​imaging diagnostic device and the results of injury or disease detection by the AI ​​imaging diagnostic device.

[0070] REFERENCE SIGNS LIST 1 AI image diagnostic device 2 Dental OCT image diagnostic device 3 Dental OCT device 10 Trained model 11 Model execution unit 12 Display control unit 13A Analysis result 13B Detection result 14 Input device 15 Display device 20 Computer 21 CPU 22 Storage device 24 Input device 25 Display device 30 Teacher data 40 Model builder 41 CPU 42 Storage device 43 Model building processing program

Claims

1. An AI image diagnostic device that receives three-dimensional dental image data taken with a dental OCT device and analyzes the input three-dimensional dental image data, and is equipped with a model execution unit that sequentially inputs each of the two-dimensional tomographic image data that make up the three-dimensional dental image data of the diagnosis target into a trained model, and as a result of executing the trained model, obtains injury / illness information data for each tomographic image data of the three-dimensional dental image data, which is data related to areas identified as characteristic features of injuries, diseases, etc. in the input tomographic image data, and uses the obtained injury / illness information data to detect injuries and diseases from the input three-dimensional dental image data, and the trained model is constructed by learning from three-dimensional dental image data of multiple patients that were previously taken with the dental OCT device.

2. The trained model is constructed by performing machine learning using tomographic image data constituting three-dimensional dental image data previously taken with a dental OCT device, in which characteristic lesions that are characteristic of lesions are present, as training data, to search for characteristic lesions from the input tomographic image data, and to output the injury / illness information data, wherein the injury / illness information data includes at least the name of the characteristic lesion, a central position indicating the coordinates of the point with the highest similarity to the characteristic lesion that is assigned to each point constituting the image in the tomographic image data, and a similarity assigned to the point indicating the central position, The AI ​​image diagnostic device of claim 1, wherein the model execution unit graphs the relationship between the tomographic position and the similarity of the characteristic feature based on the injury and disease information data for each tomographic image data of the three-dimensional dental image data acquired from the trained model, determines, for each of the characteristic features, areas where the similarity exceeds a predetermined threshold at consecutive tomographic positions on the graph, determines a name of the area for each of the areas as selection information, and generates an injury and disease detection result in the three-dimensional dental image data by associating, for each area whose name has been determined, the name of the characteristic feature, the tomographic position having the point in the area to which the highest similarity is assigned, and the center position of the tomographic position having the point to which the highest similarity is assigned, with the name of the area.

3. The trained model is constructed by performing machine learning using training data of tomographic image data constituting three-dimensional dental image data previously taken with a dental OCT device, including tomographic image data containing injury characteristic parts that are characteristic of injuries and injuries, and tomographic image data containing non-injury characteristic parts that are characteristic of parts other than injuries, to search for injury characteristic parts and non-injury characteristic parts from the input tomographic image data, and output the injury information data, wherein the injury information data includes at least the name of the characteristic parts, a central position indicating the coordinates of the point that is most similar to the characteristic part and that is assigned to each point constituting the image in the tomographic image data, and a similarity assigned to the point indicating the central position, The AI ​​image diagnostic device of claim 1, wherein the model execution unit graphs the relationship between the tomographic position and the similarity of the characteristic feature for each characteristic feature based on the injury and disease information data for each tomographic image data of the three-dimensional dental image data acquired from the trained model, finds areas on the graph where the similarity exceeds a predetermined threshold at consecutive tomographic positions, determines a name of the area for each of the areas as selection information, and generates an injury and disease detection result in the three-dimensional dental image data by associating, for each area whose name has been determined, the name of the characteristic feature, the tomographic position having the point in the area to which the highest similarity has been assigned, and the center position of the tomographic position having the point to which the highest similarity has been assigned, with the name of the area.

4. The AI ​​image diagnostic device according to claim 2 or 3, characterized in that it comprises a display control unit that, when the name of the area, which is the selection information, is input by user operation, extracts tomographic image data corresponding to the name of the area from the input three-dimensional dental image data based on the injury / disease detection result generated by the model execution unit, and displays a dental image including the tomographic image on a display device.

5. The AI ​​image diagnostic device described in any one of claims 2 to 4, characterized in that the two-dimensional tomographic image data that the model execution unit sequentially inputs to the trained model is tomographic image data that constitutes three-dimensional tooth image data being captured by a dental OCT device.

6. The trained model is constructed by performing machine learning using tomographic image data constituting three-dimensional dental image data previously taken with a dental OCT device, containing characteristic lesions that are characteristic of an injury or disease, as training data, to search for the characteristic lesions in the input tomographic image data, and generate an image representing a degree of similarity, which is image data in which the similarity to the characteristic lesion assigned to each point constituting the image in the tomographic image data is replaced with a pixel value, thereby visualizing and outputting the position of the characteristic lesion; the injury and disease information data includes at least the name of the characteristic lesion and an image representing the degree of similarity; and the model execution unit determines the name of the characteristic lesion included in the injury and disease information data as selection information based on the injury and disease information data for each tomographic image data of the three-dimensional dental image data obtained from the trained model, reconstructs the image representing the degree of similarity included in the injury and disease information data for each name of the characteristic lesion, thereby generating a three-dimensional image of the characteristic lesion, and uses the generated three-dimensional image of the characteristic lesion as the detection result of the injury or disease in the three-dimensional dental image data.

7. The trained model is constructed by machine learning using training data of tomographic image data constituting three-dimensional dental image data previously taken with a dental OCT device, including tomographic image data containing lesion characteristic parts that are characteristic of lesions and tomographic image data containing non-lesion characteristic parts that are characteristic of parts other than lesions, to search for lesion characteristic parts and non-lesion characteristic parts from the input tomographic image data, and generate an image representing the degree of similarity, which is image data in which the similarity between each point constituting the image in the tomographic image data and the corresponding feature is replaced with a pixel value, to visualize and output the position of the feature part; and the injury information data includes at least the name of the feature part and the image representing the degree of similarity, The AI ​​image diagnostic device according to claim 1, characterized in that the model execution unit determines the name of the characteristic part included in the injury / illness information data as selection information based on the injury / illness information data for each tomographic image data of the three-dimensional dental image data obtained from the trained model, generates a three-dimensional image of the characteristic part by reconstructing an image representing the degree of match included in the injury / illness information data for each name of the characteristic part, and sets the generated three-dimensional image of the characteristic part as a detection result of the injury or disease in the three-dimensional dental image data.

8. The AI ​​image diagnostic device according to claim 6 or 7, characterized in that it comprises a display control unit that, when the name of a feature, which is the selection information, is input by user operation, superimposes a three-dimensional image of the feature generated corresponding to the name of the feature on the input three-dimensional dental image data based on the injury / disease detection result generated by the model execution unit, and displays the superimposed image on a display device.

9. The AI ​​image diagnostic device according to any one of claims 6 to 8, characterized in that, when replacing the similarity with the feature part with a pixel value, the trained model performs a process of converting the pixel values ​​of pixels whose similarity is smaller than a preset threshold to 0 to generate an image representing the degree of match, and the model execution unit obtains from the trained model an image in which the pixel values ​​of pixels whose similarity is smaller than the threshold have been converted to 0 as the image representing the degree of match.

10. The AI ​​imaging diagnostic device of any one of claims 2 to 9, characterized in that the training data includes at least one of: tomographic image data of an A-section parallel to a plane specified by both the B-axis direction, which is orthogonal to the A-axis direction, which is the irradiation direction in which the OCT laser light is irradiated on the tooth, and the A-axis direction; tomographic image data of an L-section parallel to a plane specified by both the V-axis direction, which are orthogonal to the A-axis direction and the B-axis direction, and the A-axis direction; tomographic image data of an S-section parallel to a plane specified by both the B-axis direction and the V-axis direction; en-face image data, which is image data combining information on the tooth surface irradiated with the OCT laser light and information in the A-axis direction; and three-dimensional image data consisting of several consecutive tomographic image data.

11. The AI ​​image diagnostic device according to claim 10, characterized in that the model execution unit receives an instruction for either a horizontal scan or a vertical scan through user operation, and when receiving an instruction for a horizontal scan, inputs the tomographic image data of the A cross-section into the trained model, and when receiving an instruction for a vertical scan, inputs the tomographic image data of the L cross-section into the trained model, thereby executing arithmetic processing of the trained model.

12. The AI ​​image diagnostic device described in claim 10, characterized in that the model execution unit determines information on the scan direction when the input three-dimensional tooth image data is captured using a dental OCT device, and if it determines that horizontal scan information is included, inputs the tomographic image data of the A cross section into the trained model, and if it determines that vertical scan information is included, inputs the tomographic image data of the L cross section into the trained model, and executes calculation processing for the trained model.

13. A dental OCT imaging diagnostic device characterized in that the dental OCT device is equipped with an AI imaging diagnostic device described in any one of claims 1 to 12.

Images