Gingival crevicular fluid sampling brush quantification method based on image recognition

Abstract

A gingival crevicular fluid sampling brush quantification method based on image recognition comprises the steps that the tip of a gingival crevicular fluid sampling brush is pre-stained with a plaque staining agent, the pre-stained gingival crevicular fluid sampling brush is placed at a single periodontal pocket site for gingival crevicular fluid sampling, and after the gingival crevicular fluid sampling brush sucks gingival crevicular fluid of different volumes, the gingival crevicular fluid sampling brush is used for sampling gingival crevicular fluid. The dye attached to the tip of the gingival crevicular fluid sampling brush is bloomed upwards along with the fluid, and the dyeing boundary and the dyeing area of the gingival crevicular fluid sampling brush are changed; the method comprises the following steps: taking photos of different dimensions of a gingival crevicular fluid sampling brush, dividing the photos into a training set and a test set, training a neural network model through several stages of threshold segmentation, feature extraction and model training, and obtaining a correlation prediction result between an input image and a fluid volume so as to realize gingival crevicular fluid volume quantification. The method has the advantages of being simple in operation method, rapid in analysis and accurate in result.

Description

Technical Field

[0001] The invention relates to the technical field of periodontitis detection, and in particular to a gingival sulcus fluid sampling brush quantitative method based on image recognition. Background Art

[0002] Periodontitis is widely prevalent worldwide and is one of the world's sixth largest chronic diseases. Gingival crevicular fluid (GCF) has the advantages of non-invasive sampling and can accurately reflect the level of periodontitis sites, which is of great significance in the diagnosis of periodontitis. The amount of GCF can reflect the severity of periodontitis to a certain extent. Many studies have shown that the amount of GCF has potential value in indicating gingival inflammation. The value is low in healthy tissue and high in inflamed tissue. The relationship between the degree of inflammation and the amount of GCF may be linear. In some studies, researchers have also explored the relationship between other inflammatory index parameters and the amount of GCF: the gingival index and the amount of GCF are closely related. The higher the gingival index, the greater the amount of GCF; the density of inflammatory cells in pathological examination and the measurement results of GCF are well correlated; as the pocket depth increases, the GCF flow also increases. These research results show that the measurement of GCF volume is of great significance in reflecting the state of periodontitis. The existing methods for quantitative determination of gingival crevicular fluid mainly include the ninhydrin method, the weighing method, and the Periotron method corresponding to the Periopaper method. The ninhydrin staining method produces a blue-purple color reaction through the combination of ninhydrin and protein in the gingival crevicular fluid, and the liquid volume is evaluated according to the printing area of ​​the filter paper. However, the operation is complicated and may affect the analysis [Griffiths GS.Formation, collection and significance of gingival crevice fluid [J]. Periodontol 2000, 2003, 31: 32-42.]; the weighing method converts the volume of gingival crevicular fluid adsorbed on the paper strip according to the weight difference of the filter paper before and after collection. The amount of gingival crevicular fluid needs to be weighed before and after sampling, which is time-consuming and has a high risk of sample evaporation [Ali AF, Taha GI.Gingival crevicular fluid volume and protein concentration: A biomarker tool for predicting periodontal diseases progression and severity [J]. Natural and Life Sciences Communications, 2024, 23(4): e2024055.]. Periotron is an electronic instrument specifically used to quantify the volume of gingival crevicular fluid absorbed by PerioPaper filter paper. It is expensive and needs to be imported from abroad. These methods still have the above shortcomings for quantifying gingival crevicular fluid.The current internationally accepted method, Periopaper, has the disadvantages of slow sampling speed, long elution time, and high protein recovery rate, and there is no supporting quantitative measure. Currently, it can only be completed by weighing the sampling brush before and after sampling. It is inefficient and there is a risk of easy evaporation of trace gingival crevicular fluid [Zhang J, Li W, Lu H, et al. Interactions of periodontal pathogens with platelets in the gingival crevicular fluid of patients with periodontitis [J]. Journal of Clinical Periodontology, 2022, 49(9): 922-931.]. Summary of the invention

[0003] In order to overcome the problems existing in the above-mentioned prior art, the purpose of the present invention is to propose a quantification method of gingival sulcus fluid sampling brush based on image recognition. By pre-dying the tip of the sampling brush with a plaque stain, it can not only provide visual feedback for the doctor collecting the sample, but also provide image data for image recognition quantification; the collection intelligently identifies these changes, constructs the relationship between volume change and liquid aspirate amount, and realizes the quantification of trace liquids, which has the advantages of being simple, fast and accurate.

[0004] In order to achieve the above object, the technical solution adopted by the present invention is:

[0005] A quantification method for a gingival sulcus fluid sampling brush based on image recognition, which comprises the following steps: pre-dying the tip of the gingival sulcus fluid sampling brush with a plaque dye, placing the pre-dyed gingival sulcus fluid sampling brush at a single periodontal pocket site for gingival sulcus fluid sampling, and causing the dye attached to the tip of the gingival sulcus fluid sampling brush to spread upward with the liquid after the gingival sulcus fluid sampling brush absorbs gingival sulcus fluid of different volumes, and both the dyeing boundary and the dyeing area of ​​the gingival sulcus fluid sampling brush will change; immediately after sampling, photographs of the gingival sulcus fluid sampling brush in different dimensions are taken, and the photographs are divided into a training set and a test set; a neural network model is trained through threshold segmentation, feature extraction, and model training to obtain a correlation prediction result between the input image and the liquid volume, thereby achieving quantification of the gingival sulcus fluid volume.

[0006] A gingival sulcus fluid sampling brush quantitative method based on image recognition, specifically comprising the following steps:

[0007] Step 1: Pre-stain the tip of the gingival sulcular fluid sampling brush with a plaque stain;

[0008] Step 2, placing the pre-stained gingival sulcus fluid sampling brush at a single periodontal pocket site to sample gingival sulcus fluid; photographing the gingival sulcus fluid sampling brush after sampling from different directions, collecting images and performing RGB-HSV conversion on the images;

[0009] Step 3: Divide the collected images into training sets and test sets, train the neural network model through image segmentation, feature extraction, and model training, and obtain a correlation prediction model between the input image and the liquid volume reading to achieve quantification of gingival sulcular fluid volume.

[0010] The specific method of step 1 is:

[0011] Use plaque stain to pre-dye the tip of the gingival sulcular fluid sampling brush for a length of 2.8mm-3.2mm. The staining should be uniform and the staining depth should be uniform to the naked eye, and the dye should not move upward due to capillary action.

[0012] The bristles of the gingival sulcus fluid sampling brush are the bristles of a pointed soft-bristled brush, and the tip diameter of the sampling brush is 0.2-0.4 mm.

[0013] The specific method of step 2 is:

[0014] Use n pre-stained gingival sulcus fluid sampling brushes to quantitatively absorb gingival sulcus fluid according to the custom volume gradient, and take photos of the front, back, left and right sides of the n pre-stained gingival sulcus fluid sampling brushes that absorb gingival sulcus fluid; create folders and name them according to the volume gradient values; under each volume gradient, the four photos of each sampling brush are named "X-1-F", "X-1-B", "X-1-L", "X-1-L", where X is a number, representing the Xth sampling brush under the same volume gradient, 1 represents after absorption, F represents the front, B represents the back, L represents the left side, and R represents the right side. Finally, the image is converted from RGB to HSV.

[0015] The specific method of step 3 is:

[0016] Step 3.1 Divide the converted images in step 2 into training set and test set; perform image segmentation:

[0017] Step 3.1.1, select the target area and observe and compare the clarity and background interference of the target area at different hues (H), saturation (S), and brightness (V), and determine the optimal HSV threshold with the highest clarity and the least background interference of the target area;

[0018] Step 3.1.2, create a mask using the optimal HSV threshold, retain only the pixels whose HSV is within the optimal threshold range, extract the target area, and separate all the pixels that do not meet the optimal HSV threshold range;

[0019] Step 3.1.3, binarize the image differentiated in step 3.1.2, and convert the image into black and white mode using Otsu threshold method or adaptive threshold; then perform opening operation on the binarized image: based on the morphological structure element, use the morphological operation of first corrosion and then expansion to smooth the target edge and remove small noise; obtain a complete and clear image of the target area, and complete the image segmentation.

[0020] Step 3.2, extract features from the segmented image:

[0021] Step 3.2.1, respectively extract the number of all pixels of the three channels R, G, and B of the target area in the front, back, left, and right images, and the height, width, and area feature values ​​of the target area;

[0022] Step 3.2.2, using data analysis and visualization software to fit the correlation between each characteristic value extracted in step 3.2.1 and the gingival crevicular fluid volume;

[0023] Step 3.2.3, compare the goodness of fit between each eigenvalue in step 3.2.2 and the gingival sulcular fluid volume, remove the eigenvalues ​​with poor self-defined goodness of fit, multiply the remaining eigenvalues ​​by two, obtain n eigenvalues, and send the n eigenvalues ​​and the corresponding volume values ​​to the neural network model for learning;

[0024] Step 3.3, the image feature data extracted in step 3.2 is divided into a training set and a test set, and trained using a neural network model;

[0025] Step 3.3.1, construct a neural network model including an input layer, an output layer and multiple hidden layers. Batch Norm and Dropout are set before and after each layer of the input layer, output layer and multiple hidden layers, and residual connections are introduced between hidden layers to enhance feature learning capabilities;

[0026] Step 3.3.2, use the neural network model to learn and train the image feature data transmitted in step 3.2, monitor the change of the loss function (loss) value in real time during the training process, and use the preset neural network model parameters corresponding to the lowest loss value of the test set as the final convergence result; then, use the model with the lowest loss value of the test set to perform linear regression analysis on the predicted value and the true value of the gingival sulcular fluid volume in the test set, and use the regression goodness of fit and regression coefficient R 2 Evaluate the accuracy of the neural network model's predictions.

[0027] The regression goodness of fit and regression coefficient R 2 It is positively correlated with the accuracy of the prediction results of the neural network model, and the accuracy of the prediction results of the neural network model represents the quantitative accuracy of the gingival sulcular fluid volume.

[0028] Compared with the prior art, the present invention has the following beneficial effects:

[0029] The present invention pre-dyes the tip of the sampling brush with a plaque dye. When sampling, the tip of the brush absorbs gingival sulcus fluid and the dye on the tip is soaked. The dye dissolves in the gingival sulcus fluid and moves upward with the brush fibers. Different color distributions are generated on the sampling brush according to the amount of liquid absorbed, providing real-time visual feedback for doctors to judge the sampling amount, and generating high-contrast images for analysis by the neural network model.

[0030] The present invention uses a deep learning algorithm to segment and extract features of the stained area, combines color thresholds to judge the stained area and color depth, and accurately predicts the volume of the sampled liquid; the method of the present invention is simple to operate, fast to analyze and accurate in results, and can subsequently be combined with smartphone software development to take photos and upload them in real time to complete automatic recognition and data feedback, thereby realizing intelligent and standardized evaluation of sampling quality, greatly improving analysis efficiency and convenience, and being suitable for clinical and primary care scenarios.

[0031] In summary, the present invention detects the sampling amount of the gingival crevicular fluid sampling brush by combining a quantitative method of image processing and a neural network model. This scheme not only provides doctors with instant visual feedback of clinical sampling, but also provides a fast, accurate and simple scheme for gingival crevicular fluid quantification, which has strong clinical significance. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 Schematic diagram of pre-staining and liquid aspiration of the sampling brush according to an embodiment of the present invention.

[0033] Figure 2 This is a schematic diagram of image acquisition after the sampling brush is pre-stained and absorbs liquid in an embodiment of the present invention (scale: 5mm).

[0034] Figure 3 Schematic diagram of the image processing process according to an embodiment of the present invention.

[0035] Figure 4 Schematic diagram of the image processing process of an embodiment of the present invention (goodness of fit of 24 features and volume).

[0036] Figure 5 This is the construction of a quantitative neural network model for the volume of liquid absorbed by a sampling brush according to an embodiment of the present invention. DETAILED DESCRIPTION

[0037] The present invention will be further described in detail below in conjunction with the accompanying drawings.

[0038] The present invention provides a quantification method for gingival crevicular fluid sampling brush based on image recognition, which achieves the correlation determination coefficient R between the actual volume and the prediction of the neural network model in the range of 0-2 μL artificial saliva (simulated gingival crevicular fluid) in in vitro testing. 2=0.7190, confirming the feasibility of AI image recognition quantification. The present invention is conducive to obtaining the important diagnostic information of gingival sulcular fluid volume while collecting gingival sulcular fluid. The present invention is further described in detail below in conjunction with specific embodiments, which are explanations of the present invention rather than limitations.

[0039] A gingival sulcus fluid sampling brush quantitative method based on image recognition, specifically comprising the following steps:

[0040] Step 1: Prepare 250 gingival sulcus fluid sampling brushes with a tip diameter of 0.3 mm using the bristles of a pointed soft-bristle brush; dip the brushes into a commercially available dental plaque stain to pre-dye the tip 3 mm of the gingival sulcus fluid sampling brushes; the front, back and sides of the sampling brushes need to be evenly dyed, and the specific dyeing depth needs to be uniform and consistent to the naked eye; the sampling brushes that are dyed beyond the tip 3 mm due to operational errors or the capillary action of the dye need to be discarded, such as Figure 1 shown.

[0041] Step 2, use the pre-stained sampling brushes to absorb 0 μL, 0.2 μL, 0.4 μL, 0.6 μL, 0.8 μL, 1.0 μL, 1.2 μL, 1.4 μL, 1.6 μL, 1.8 μL and 2 μL of gingival crevicular fluid, with 20 sampling brushes for each volume gradient, and take photos of the front, back, left side and right side immediately after sampling;

[0042] The sampling brushes were used to absorb 0.3 μL, 0.5 μL, 0.7 μL, 0.9 μL, 1.1 μL, 1.3 μL, 1.5 μL, 1.7 μL and 1.9 μL of water respectively. Four sampling brushes were used for each volume gradient. Photos of the front, back, left and right sides were taken immediately after sampling. Figure 2 shown.

[0043] The above images are all 3840*2160 in size. Create folders and name them according to the volume gradient values. Under each volume gradient, the four photos of each sampling brush are named "X-1-F", "X-1-B", "X-1-L", and "X-1-L", where X is a number, representing the Xth sampling brush under the same volume gradient, 1 represents after liquid absorption, F represents the front side, B represents the back side, L represents the left side, and R represents the right side. The pictures are converted from RGB to HSV.

[0044] Step 3: Divide the collected images into training sets and test sets. Train the neural network model through image segmentation, feature extraction, and model training to obtain a correlation prediction model between the input image and the liquid volume reading, thereby achieving quantification of gingival sulcular fluid volume:

[0045] 3.1 Divide the converted images in step 2 into training set and test set, with a total of 256 sampled brushes and 1024 images, of which 220 brushes and 880 images are used as training set; 36 brushes and 144 images are used as test set; perform image segmentation:

[0046] like Figure 3 As shown, 3.1.1) Convert the photo image to HSV mode and observe the performance of the target area in terms of hue (H), saturation (S), and brightness (V). Set multiple groups of HSV ranges for testing, such as H in 10-20, S in 80-255, and V in 90-210, and test the segmentation effect of each combination on the target area. By comparing the regional clarity and background interference, determine the highest and lowest HSV thresholds with the best effect;

[0047] 3.1.2) Use the final threshold to create a mask, only retaining pixels within the HSV range, turning other pixels black, extracting the target area, and turning all pixels that do not meet the threshold range black;

[0048] 3.1.3) First, binarize the segmented image: the area that meets the threshold is set to white (255), and the area that does not meet the threshold is set to black (0) for easy subsequent operations. Then, the binary image is opened, that is, erosion first and then expansion, and a morphological transformation is performed using a rectangular structure element of size 15x15. This operation can eliminate small noise points, and the segmented image only retains the target area that meets the HSV threshold, the background is cleared, and small noise points are also removed, and finally a complete and clear image of the target area is obtained, completing the image segmentation.

[0049] 3.2, feature extraction of the segmented image:

[0050] 3.2.1) Extract 24 feature values, including the number of all pixels in the three channels of R, G, and B, the height, width, and area of ​​the target area in the front, back, left, and right images respectively;

[0051] 3.2.2) Fitting the correlation between the 24 eigenvalues ​​and the gingival sulcular fluid volume: Use the data analysis and visualization software Graphpad Prism to fit the correlation between each eigenvalue extracted in step 3.2.1 and the gingival sulcular fluid volume;

[0052] 3.2.3) If Figure 4 As shown: the goodness of fit between different features and liquid volume is compared, and 8 features with poor goodness of fit are removed. Finally, 16 features are left, and 256 eigenvalues ​​are obtained by multiplying them two by two.

[0053] (4) After the pre-stained sampling brush sucks 0 μL, 0.2 μL, 0.4 μL, 0.6 μL, 0.8 μL, 1.0 μL, 1.2 μL, 1.4 μL, 1.6 μL, 1.8 μL and 2 μL of gingival crevicular fluid, the characteristic value and volume value of the sampling brush are correspondingly transmitted to the neural network for learning.

[0054] 3.3, the image feature data extracted in step 3.2 is divided into a training set and a test set, and trained using a neural network model;

[0055] 3.3.1) Constructing a neural network model:

[0056] a) 1 input layer, 1 output layer, and 32 hidden layers;

[0057] b) Residual Connection is used between hidden layers;

[0058] c) BatchNorm and Dropout are set before and after each of the 1 input layer, 1 output layer, and 32 hidden layers.

[0059] 3.3.2) Model training:

[0060] The neural network model is used to train the image feature data delivered in step 3.2. During the training process, the change of the loss function (loss) value is monitored in real time, and the preset neural network model parameters corresponding to the lowest loss value of the test set are used as the final convergence result. Subsequently, a linear regression analysis is performed on the predicted value and the true value of the gingival sulcular fluid volume in the test set using the model when the loss value of the test set is the lowest. The goodness of fit and the regression coefficient R 2 Evaluate the accuracy of model prediction results, regression goodness of fit and regression coefficient R 2 It is positively correlated with the accuracy of prediction results.

[0061] Simulation experiment:

[0062] The training and testing environment for this study is: Intel i5-14600KF CPU, torch 2.1.1, python.

[0063] Operating environment;

[0064] The hyperparameter settings involved in the experiment can be seen in the following table:

[0065]

[0066] The dataset used in this study contains 256 sample brushes and 1024 images, of which 220 brushes and 880 images are used as training sets, and 36 brushes and 144 images are used as test sets.

[0067] During the training process, the change of loss value needs to be constantly monitored. This study selected 30 iterations (Epoch); the model parameters with the lowest loss value of the test set were taken as the final result;

[0068] Evaluation index: In the in vitro test, the correlation coefficient R between the actual volume and the prediction of the neural network model was achieved in the range of 0-2μL artificial saliva (simulated gingival sulcus fluid). 2 =0.7190, the predicted value and true value of the test set are linearly fitted when the loss value of the test set is the lowest, and the accuracy of the model prediction result is high, which confirms the feasibility of quantitative image recognition of neural network model. Figure 5 shown.

[0069] It can be seen from the embodiments and experimental results that, compared with the prior art, the present invention has developed a method for quantitative liquid aspiration based on a previous innovative invention: a gingival sulcus fluid sampling brush. This method has two specific advantages and innovations.

[0070] One is to pre-dye the tip of the sampling brush to make it colored. During the sampling process, the doctor can observe the color spread due to the aspiration of the liquid, which provides the doctor with good sampling feedback.

[0071] The second is to innovatively apply AI image recognition to the strategy of quantification of gingival crevicular fluid and verify that the developed AI model has good ability to predict fluid (R2=0.7190).

[0072] Based on the above two advantages, combined with subsequent software development and chairside system construction, a new, visual, fast and accurate solution will be provided for the rapid collection and quantification of clinical gingival crevicular fluid.

Claims

1. A quantification method of gingival crevicular fluid sampling brush based on image recognition, characterized in that: The tip of the gingival sulcus fluid sampling brush was pre-stained with plaque stain, and the pre-stained gingival sulcus fluid sampling brush was placed at a single periodontal pocket site for gingival sulcus fluid sampling. After the gingival sulcus fluid sampling brush absorbed different volumes of gingival sulcus fluid, the dye attached to the tip of the gingival sulcus fluid sampling brush spread upward with the liquid, and the staining boundary and staining area of ​​the gingival sulcus fluid sampling brush changed at the same time; photos of the gingival sulcus fluid sampling brush in different dimensions were taken immediately after sampling, and these photos were divided into training and test sets. The neural network model was trained through threshold segmentation, feature extraction, and model training to obtain the correlation prediction result between the input image and the liquid volume, thereby realizing the quantification of gingival sulcus fluid volume.

2. A quantification method of gingival crevicular fluid sampling brush based on image recognition, characterized in that: The specific steps include: Step 1: Pre-stain the tip of the gingival sulcular fluid sampling brush with a plaque stain; Step 2, placing the pre-stained gingival sulcus fluid sampling brush at a single periodontal pocket site to sample gingival sulcus fluid; photographing the gingival sulcus fluid sampling brush after sampling from different directions, collecting images and performing RGB-HSV conversion on the images; Step 3: Divide the collected images into training sets and test sets, train the neural network model through image segmentation, feature extraction, and model training, and obtain a correlation prediction model between the input image and the liquid volume reading to achieve quantification of gingival sulcular fluid volume.

3. The quantification method of gingival crevicular fluid sampling brush based on image recognition according to claim 2, characterized in that: The specific method of step 1 is: Use plaque stain to pre-dye the tip of the gingival sulcular fluid sampling brush for a length of 2.8mm-3.2mm. The staining should be uniform and the staining depth should be uniform to the naked eye, and the dye should not move upward due to capillary action.

4. The quantification method of gingival crevicular fluid sampling brush based on image recognition according to claim 3, characterized in that: The bristles of the gingival sulcus fluid sampling brush are the bristles of a pointed soft-bristled brush, and the tip diameter of the sampling brush is 0.2-0.4 mm.

5. The quantification method of gingival crevicular fluid sampling brush based on image recognition according to claim 2, characterized in that: The specific method of step 2 is: Use n pre-stained gingival sulcus fluid sampling brushes to quantitatively absorb gingival sulcus fluid according to the custom volume gradient, and take photos of the front, back, left and right sides of the n pre-stained gingival sulcus fluid sampling brushes that absorb gingival sulcus fluid; create folders and name them according to the volume gradient values; under each volume gradient, the four photos of each sampling brush are named "X-1-F", "X-1-B", "X-1-L", "X-1-L", where X is a number, representing the Xth sampling brush under the same volume gradient, 1 represents after absorption, F represents the front, B represents the back, L represents the left side, and R represents the right side. Finally, the image is converted from RGB to HSV.

6. The quantification method of gingival crevicular fluid sampling brush based on image recognition according to claim 2, characterized in that: The specific method of step 3 is: Step 3.1 Divide the converted images in step 2 into training set and test set; perform image segmentation: Step 3.1.1, select the target area and observe and compare the clarity and background interference of the target area at different hues (H), saturation (S), and brightness (V), and determine the optimal HSV threshold with the highest clarity and the least background interference of the target area; Step 3.1.2, create a mask using the optimal HSV threshold, retain only the pixels whose HSV is within the optimal threshold range, extract the target area, and separate all the pixels that do not meet the optimal HSV threshold range; Step 3.1.3, binarize the image differentiated in step 3.1.2, and convert the image into black and white mode using Otsu threshold method or adaptive threshold; then perform opening operation on the binarized image: based on morphological structural elements, use morphological operations of first corrosion and then expansion to smooth the target edge and remove small noise; Get a complete and clear image of the target area and complete image segmentation; Step 3.2, extract features from the segmented image: Step 3.2.1, respectively extract the number of all pixels of the three channels R, G, and B of the target area in the front, back, left, and right images, and the height, width, and area feature values ​​of the target area; Step 3.2.2, using data analysis and visualization software to fit the correlation between each characteristic value extracted in step 3.2.1 and the gingival crevicular fluid volume; Step 3.2.3, compare the goodness of fit between each eigenvalue in step 3.2.2 and the gingival sulcular fluid volume, remove the eigenvalues ​​with poor self-defined goodness of fit, multiply the remaining eigenvalues ​​by two, obtain n eigenvalues, and send the n eigenvalues ​​and the corresponding volume values ​​to the neural network model for learning; Step 3.3, the image feature data extracted in step 3.2 is divided into a training set and a test set, and trained using a neural network model; Step 3.3.1, construct a neural network model including an input layer, an output layer and multiple hidden layers. Batch Norm and Dropout are set before and after each layer of the input layer, output layer and multiple hidden layers, and residual connections are introduced between hidden layers to enhance feature learning capabilities; Step 3.3.2, use the neural network model to conduct learning and training on the image feature data transmitted in step 3.2, monitor the change of the loss function (loss) value in real time during the training process, and take the preset neural network model parameters corresponding to the lowest loss value of the test set as the final convergence result; Subsequently, a linear regression analysis was performed on the predicted and true values ​​of the gingival sulcular fluid volume in the test set using the model with the lowest loss value in the test set. The goodness of fit and regression coefficient R 2 Evaluate the accuracy of the model's predictions.

7. The quantification method of gingival crevicular fluid sampling brush based on image recognition according to claim 6, characterized in that: The regression goodness of fit and regression coefficient R 2 It is positively correlated with the accuracy of the prediction results of the neural network model, and the accuracy of the prediction results of the neural network model represents the quantitative accuracy of the gingival sulcular fluid volume.

Images