Method for hidden crack image segmentation of cracked tooth based on FDB-DeepLabv3+ model

Abstract

The application relates to the technical field of software communication, and discloses a method for hidden crack tooth crack image segmentation based on an FDB-DeepLabv3+ model. A healthy tooth is processed by using a temperature difference method to generate a crack simulation hidden crack tooth; a normal tooth is switched in 100-degree boiling water and minus 196-degree liquid nitrogen, a crack of the hidden crack tooth is simulated through a quenching and heating process; an electron microscope is used to shoot a region with a crack on a tooth surface, the picture size is 1920*1080, the picture is cut into small pictures with a size of 512*512, and small pictures with cracks are manually marked with semantic level pixels; a marked data set is put into an improved FDB-DeepLabv3+ network for training, so that the network has the ability of crack segmentation, and a DeepLabV3+ image semantic segmentation model is formed. The method for hidden crack tooth crack image segmentation based on the FDB-DeepLabv3+ model realizes segmentation of hidden crack tooth cracks in a clinic, does not need manual participation, has no requirement on the diagnosis experience of doctors, and can improve the diagnosis accuracy.

Description

Technical Field

[0001] This invention relates to the field of hidden crack tooth crack recognition technology, specifically a method for hidden crack tooth crack image segmentation based on the FDB-DeepLabv3+ model. Background Technology

[0002] Cracked teeth, referring to tiny cracks on the tooth surface, are the third leading cause of tooth loss in adults and a challenging problem in current oral clinical examinations. Clinical diagnosis of cracked teeth is usually based on clinical symptoms, such as... For suspected teeth, clinicians can confirm the diagnosis using several traditional clinical tests, such as the bite test, probe method, staining method, cold stimulation method, transillumination, periapical radiograph (PR), and computed tomography (CBCT).

[0003] However, relying on the aforementioned imaging methods and visual observation to identify and diagnose cracks places a high demand on clinicians' clinical experience. Furthermore, visual fatigue or misjudgment by clinicians can lead to inaccurate diagnoses of hidden cracks. Using artificial intelligence for image processing and automatic crack identification and assessment can address these two issues to some extent. Therefore, improvements are urgently needed. Summary of the Invention

[0004] (a) Technical problems to be solved

[0005] To address the shortcomings of existing technologies, this invention provides a method for image segmentation of microcracked teeth based on the FDB-DeepLabv3+ model. This method enables clinical segmentation of microcracked teeth without manual intervention, reducing the need for experienced dentists and improving diagnostic accuracy. It is applicable to images generated by any imaging method (X-ray imaging, ultrasound imaging, optical imaging, etc.), solving the problem that clinicians currently rely on visual observation to identify and diagnose cracks using these imaging methods. This not only places high demands on clinicians' clinical experience but also leads to inaccurate diagnoses due to visual fatigue or misjudgment.

[0006] (II) Technical Solution

[0007] To achieve the goal of segmenting microcracked tooth cracks in clinical practice without manual intervention, requiring less diagnostic experience from doctors, and improving diagnostic accuracy; and applicable to images generated by any imaging method (X-ray imaging; ultrasound imaging; optical imaging, etc.), this invention provides the following technical solution: a method for segmenting microcracked tooth crack images based on the FDB-DeepLabv3+ model, comprising the following steps:

[0008] S1. Fabricating a cracked tooth

[0009] The temperature difference method is used to treat healthy teeth to create cracks to simulate hidden cracks; normal teeth are taken and switched between boiling water at 100 degrees and liquid nitrogen at -196 degrees to simulate the cracks of hidden cracks through a process of rapid cooling and heating.

[0010] S2. Creating a semantic segmentation dataset for cracked teeth.

[0011] The cracked areas on the tooth surface were photographed using an electron microscope. The images were 1920×1080 pixels in size. They were then cropped into smaller images of 512×512 pixels in size, and the cracked images were manually labeled with semantic-level pixels.

[0012] S3. Perform Deeplabv3+ network training.

[0013] The labeled dataset from step S2 above is fed into the improved FDB-DeepLabv3+ network for training, enabling it to segment cracks and forming the DeepLabV3+ image semantic segmentation model.

[0014] Preferably, in step S3, the DeepLabV3+ image semantic segmentation model adopts an encoder-decoder structure, and the backbone network is changed to the ResNet50 model.

[0015] Preferably, the ResNet50 model combines the feature maps from the first three stages of the ResNet50 model using FPN in the encoding part to compensate for lost detailed features. The fourth stage of the ResNet50 model employs the BAM attention mechanism module, which effectively guides model training by refining crack features, thereby obtaining better detection results.

[0016] Preferably, in step S3, inspired by the DenseNet network, the DeepLabV3+ image semantic segmentation model uses a dense connection method to connect the originally parallel attention mechanism modules in the densely connected module, and connects different dilated convolutional layers in a dense connection manner, thereby generating multi-scale features, which can significantly increase the model's receptive field and feature extraction capability.

[0017] Preferably, the DenseNet expression formula is as follows:

[0018] First, we use the traditional network output formula at layer l:

[0019] x l =H l (x l -1)

[0020] Adding the identity function from the input of the previous layer yields the ResNet formula:

[0021] x l =H l (x l -1)+x l -1

[0022] At this point, by connecting all the preceding layers as input, we can obtain the DenseNet formula:

[0023] x l =H l ([x0,x1,...,x l -1]).

[0024] Preferably, H1(·) represents a non-linear transformation function, which is a combination operation that may include a series of Batch Normalization (BN), ReLU, Pooling, and Conv operations. Note that there may actually be multiple convolutional layers between layer 1 and layers 1, -1.

[0025] (III) Beneficial Effects

[0026] Compared with existing technologies, this invention provides a method for image segmentation of hidden tooth cracks based on the FDB-DeepLabv3+ model, which has the following beneficial effects:

[0027] 1. This method for segmenting hidden crack tooth crack images based on the FDB-DeepLabv3+ model can achieve clinical segmentation of hidden crack tooth cracks without human intervention, reducing the need for doctors' diagnostic experience, and improving the accuracy of diagnosis.

[0028] 2. The method for segmenting hidden tooth crack images based on the FDB-DeepLabv3+ model is applicable to images generated by any imaging means (X-ray imaging; ultrasonic imaging; optical imaging, etc.). Attached Figure Description

[0029] Figure 1 It is a simulated tooth with a hidden crack;

[0030] Figure 2 It is the FDB-DeepLabv3+ model network model. Detailed Implementation

[0031] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0032] Please see Figure 1-2 The present invention provides the following technical solutions:

[0033] The method for image segmentation of hidden tooth cracks based on the FDB-DeepLabv3+ model includes the following steps:

[0034] S1. Fabricating a cracked tooth

[0035] A temperature difference method was used to treat healthy teeth to create cracks, simulating microcracks in teeth. Normal teeth were then subjected to a rapid cooling and heating process, switching between boiling water at 100 degrees Celsius and liquid nitrogen at -196 degrees Celsius, to simulate microcracks. Figure 1 As shown;

[0036] S2. Creating a semantic segmentation dataset for cracked teeth.

[0037] The cracked areas on the tooth surface were photographed using an electron microscope. The images were 1920×1080 pixels in size. They were then cropped into smaller images of 512×512 pixels in size, and the cracked images were manually labeled with semantic-level pixels.

[0038] S3. Perform Deeplabv3+ network training.

[0039] The labeled dataset from step S2 above is fed into the improved FDB-DeepLabv3+ network for training, enabling it to segment cracks and forming the DeepLabV3+ image semantic segmentation model.

[0040] The DeepLabV3+ image semantic segmentation model employs an encoder-decoder structure, with the backbone network replaced by a ResNet50 model. In the encoding phase, the feature maps from the first three stages of the ResNet50 model are combined using FPN to compensate for lost detail features. The fourth stage of the ResNet50 model utilizes a BAM attention mechanism module, which effectively guides model training by refining crack features, resulting in better detection results. Inspired by the DenseNet network, the DeepLabV3+ image semantic segmentation model uses dense connections to link the previously parallel attention mechanism modules, connecting different dilated convolutional layers to generate multi-scale features, significantly increasing the model's receptive field and feature extraction capabilities. The overall structure diagram of the FDB-DeepLabv3+ model is shown below. Figure 2 As shown.

[0041] The DenseNet expression formula is shown below:

[0042] First, we use the traditional network output formula at layer l:

[0043] x l =H l (x l -1)

[0044] Adding the identity function from the input of the previous layer yields the ResNet formula:

[0045] x l =H l (x l -1)+x l -1

[0046] At this point, by connecting all the preceding layers as input, we can obtain the DenseNet formula:

[0047] x l =H l ([x0,x1,...,x l -1]).

[0048] Here, H1(·) represents a non-linear transformation function, which is a combination operation that may include a series of Batch Normalization (BN), ReLU, Pooling, and Conv operations. Note that there may actually be multiple convolutional layers between layer 1 and layers 1 and -1.

[0049] This method for segmenting microcracked tooth images based on the FDB-DeepLabv3+ model enables clinical segmentation of microcracked teeth without manual intervention, reducing the need for experienced dentists and improving diagnostic accuracy. It is applicable to images generated by any imaging method (X-ray imaging, ultrasound imaging, optical imaging, etc.).

[0050] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for hidden crack image segmentation of cracked teeth based on FDB-DeepLabv3+ model, characterized in that, Comprising the following steps: S1, making a cracked tooth; The healthy teeth are treated by using the temperature difference method to produce cracks to simulate cracked teeth; normal teeth are taken and switched between 100-degree boiling water and minus 196-degree liquid nitrogen, through the process of rapid cooling and heating, to simulate cracked tooth cracks; S2, making a cracked tooth semantic segmentation dataset; An electron microscope is used to take pictures of the areas of the tooth surface with cracks, the picture size is 1920x1080, the small pictures with a size of 512x512 are cut, and the small pictures with cracks are manually marked at the semantic level; S3, Deeplabv3+ network training; The dataset marked in the above step S2 is put into the improved FDB-DeepLabv3+ network for training, so that it has the ability to segment cracks, and a DeepLabV3+ image semantic segmentation model is formed; In the step S3, the DeepLabv3+ image semantic segmentation model adopts the encoder-decoder structure, and the main network is changed to ResNet50 model; The ResNet50 model combines the feature maps of the first three stages of the ResNet50 model in the encoding part in the form of FPN to make up for the loss of detailed features; the fourth stage of the ResNet50 model adopts a BAM attention mechanism module, which refines the crack features to effectively guide the model training, so as to obtain better detection results; In the step S3, inspired by the DenseNet network, in the densely connected module of the DeepLabv3+ image semantic segmentation model, the originally parallel attention mechanism modules are connected in a densely connected manner, and different dilated convolution layers are connected in a densely connected manner, so as to generate multi-scale features, which can significantly increase the receptive field and feature extraction capability of the model.

2. The method for FDB-DeepLabv3+ model-based hidden crack tooth crack image segmentation according to claim 1, characterized in that, The DenseNet expression formula is as follows: First, the output formula of the traditional network at layer l: Add the identity function from the input of the previous layer to get the ResNet formula: At this time, connect all the previous layers as input to get the DenseNet formula: x l =H l ([x0,x1,..., ]) wherein the H l (·) represents a non-linear transformation, which is a combination operation including a series of BN (Batch Normalization), ReLU, Pooling and Conv operations; 1 layer and The plurality of convolutional layers are included between the layers.

Images