Oral mucosa disease focus segmentation system and method

Abstract

The invention discloses an oral mucosa disease focus segmentation system and method, and the method comprises the steps: collecting oral mucosa disease clinical white light images, and dividing the images into a training set, a verification set and a test set according to a certain proportion; then, Label software is used for carrying out focus labeling on the training set images, a deep learning model is constructed based on Pixel SIFT-UNet, the deep learning model is optimized, and then the optimized deep learning model is used for segmenting focus parts; a deep learning model is constructed by adopting Pixel SIFT-UNet, the recognition of a boundary part of a focus part is optimized, and the recognition precision of the focus part is improved; optimizing parameters of the deep learning model by using a Dice coefficient; the method is used for balancing the accuracy of segmentation boundary optimization and overall global classification; compared with the prior art, the accuracy of boundary segmentation and overall recognition is improved, the detection precision of the deep learning model is enhanced, and compared with the prior art, accurate recognition of the focus part in the oral mucosal disease is improved, and distinguishing between the focus part and a normal part is enhanced.

Description

Technical Field

[0001] The present invention relates to oral mucosal lesion segmentation, and in particular to an oral mucosal lesion segmentation system and method. Background Art

[0002] In the clinical diagnosis of oral mucosal diseases (such as leukoplakia, ulcers, and precancerous lesions), lesion segmentation is a key technology that helps doctors quickly locate lesions. Traditional medical image segmentation methods, such as threshold segmentation and region growing, often perform poorly when processing oral image data due to the irregular shapes, variable textures, and complex background interference of lesions. With the rise of deep learning technology, convolutional neural network architectures such as U-Net have achieved some success in the field of medical image segmentation. However, they still suffer from insufficient segmentation accuracy when dealing with the subtle features and scale variations of oral mucosal lesions. Summary of the Invention

[0003] The present invention overcomes the deficiencies of the prior art and provides a system and method for segmenting oral mucosal lesions.

[0004] To achieve the above-mentioned object, the technical solution adopted by the present invention is: a system and method for segmenting oral mucosal lesions, comprising: a data preparation module for acquiring and segmenting oral image data, marking the lesion site, and defining a lesion area 1 and a normal area 0;

[0005] The PixelSIFT-UNet model building module is used to build a learning model. It uses the Sobel operator and dilated convolution to enhance lesion boundary information and improve the model's sensitivity to blurred boundaries.

[0006] Training strategy module, used to optimize learning model parameters;

[0007] Evaluation and testing module, used to measure the overlap between the predicted area and the true label;

[0008] The reasoning module inputs the oral image data and obtains the final result.

[0009] In a preferred embodiment of the present invention, the oral cavity image data is a white light image of the oral cavity.

[0010] In a preferred embodiment of the present invention, the oral image data is divided into a training set, a validation set, and a test set according to a certain ratio, and the certain ratio is 7:2:1.

[0011] In a preferred embodiment of the present invention, the lesion site is marked by using Labelme software to mark the lesion on the training set image, and the marked lesion site is recorded as Y mask , specifically: use the binary label map Y, where,

[0012]

[0013] In a preferred embodiment of the present invention, a deep learning model is constructed based on PixelSIFT-UNet using lesion data. Specifically, this model generates a lesion region segmentation probability map; specifically,

[0014]

[0015] Among them, Encode(x) is the multi-scale feature extracted by the encoder;

[0016] PixelSIFT(*) is a multi-scale gradient enhancement module in skip connections;

[0017] Decode(*) is the upsampling and feature fusion process of the decoder;

[0018] σ is the sigmoid function.

[0019] In a preferred embodiment of the present invention, for the multi-scale features extracted by the encoder, the oral image data x∈R H×W×3 Input into the backbone network of VGG-16 to generate a set of feature maps:

[0020] in

[0021] The specific steps of using the multi-scale gradient enhancement module are:

[0022] Feature map features of the input encoder Calculate the horizontal and vertical gradients of the encoder's feature map:

[0023]

[0024] Among them, Sobel x , Sobel y Represents the Sobel operator (horizontal / vertical direction);

[0025] Then, dilated convolution is used to extract the gradient features under different receptive fields:

[0026] F grad =Conv 3×3,dilation＝1 (G x +G y )+Conv 3×3,dilation＝2 (G x +G y );

[0027] After concatenating the original encoder features with the gradient features, enhanced features are generated through convolution:

[0028]

[0029] In a preferred embodiment of the present invention, the specific steps of the upsampling and feature fusion process of the decoder are as follows:

[0030] Multi-scale features of the input encoder and deep low-resolution features;

[0031] Starting from the deepest features, the resolution is gradually restored through transposed convolution:

[0032]

[0033] Concatenate the upsampled features with the PixelSIFT features at the same level:

[0034]

[0035] Refine the fused features through the convolutional layer:

[0036]

[0037] Final output: The decoder end generates a segmentation probability map through 1×1 convolution and Sigmoid function.

[0038] In a preferred embodiment of the present invention, in step S4,

[0039] Substitute the validation set and adjust the learning model parameters through Adam optimization:

[0040] m t =β1m t-1 +(1-β1)g t ,

[0041]

[0042] Adjusted parameters

[0043] Output optimized model parameters θ t ;

[0044] Among them, η=0.001, β1=0.9, β2=0.999, L is the mixing function.

[0045] In a preferred embodiment of the present invention, the specific steps of inputting oral image data and obtaining the final result are:

[0046]

[0047] When the output When it is greater than 0.5, it is the lesion area, and when it is less than 0.5, it is the normal area.

[0048] A method for segmenting oral mucosal lesions, according to any one of the oral mucosal lesion segmentation systems,

[0049] S1. Obtain oral image data and divide it into training set, validation set and test set according to a certain ratio;

[0050] S2. Label the lesion locations on the training set data to obtain lesion data;

[0051] S3. Build a deep learning model based on PixelSIFT-UNet using lesion data;

[0052] S4. Use the validation set to optimize the deep learning model parameters and evaluate the deep learning model performance;

[0053] S5. Input the oral image data into the optimized learning model and obtain the results.

[0054] The present invention solves the defects existing in the background technology and has the following beneficial effects:

[0055] (1) Clinical white light images of oral mucosal diseases were collected and divided into training set, validation set and test set according to a certain ratio. The training set images were then annotated with Labelme software, and a deep learning model was constructed based on PixelSIFT-UNet. The deep learning model was optimized and then the optimized deep learning model was used to segment the lesion sites. The deep learning model was constructed using PixelSIFT-UNet to optimize the recognition of the boundary of the lesion site and improve the recognition accuracy of the lesion site. Compared with the existing technology, the accurate recognition of the lesion site in oral mucosal diseases was improved, and the distinction between the lesion site and the normal site was strengthened.

[0056] (2) Use the Dice coefficient to optimize the parameters of the deep learning model; to balance the optimization of the segmentation boundary and the accuracy of the overall global classification; to improve the accuracy of the segmentation boundary and overall recognition, and enhance the detection accuracy of the deep learning model. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments described in the present invention. Those skilled in the art can also derive other drawings based on these drawings without inventive efforts.

[0058] Figure 1is a method flow chart of a preferred embodiment of the present invention;

[0059] Figure 2 is a system flow chart of a preferred embodiment of the present invention; DETAILED DESCRIPTION

[0060] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0061] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Therefore, the scope of protection of the present invention is not limited to the specific embodiments disclosed below.

[0062] like Figure 1 、 Figure 2 As shown, a system and method for segmenting oral mucosal lesions include:

[0063] S1. Obtain oral image data and divide it into training set, validation set and test set according to a certain ratio;

[0064] S2. Label the lesion locations on the training set data to obtain lesion data;

[0065] S3. Build a deep learning model based on PixelSIFT-UNet using lesion data;

[0066] S4. Use the validation set to optimize the deep learning model parameters and evaluate the deep learning model performance;

[0067] S5. Input the oral image data into the optimized learning model and obtain the results.

[0068] In step S1, the oral cavity image data is a white light image of the oral cavity;

[0069] Among them, the oral image data is obtained: the oral image data obtained by visible light shooting can clearly show the surface structure, color, texture and other macroscopic features of the oral mucosa.

[0070] In step S1, the certain ratio is 7:2:1.

[0071] In step S1, the oral image data is set to N, the training set N train , validation set N val and the test set N text Set as Ntrain =[0.7N], N val =[0.2N],N text =[0.1N].

[0072] In step S2, the specific steps of marking the lesion boundary are as follows: using Labelme software to mark the lesion on the training set image, and the marked lesion location is recorded as Y mask , specifically: use the binary label map Y, where,

[0073]

[0074] In step S3, the established learning model is the PixelSIFT-UNet model, which generates the lesion area segmentation probability map; specifically:

[0075]

[0076] Among them, Encode(x) is the multi-scale feature extracted by the encoder;

[0077] PixelSIFT(*) is a multi-scale gradient enhancement module in skip connections;

[0078] Decode(*) is the upsampling and feature fusion process of the decoder;

[0079] σ is the sigmoid function.

[0080] Specifically, for the multi-scale features extracted by the encoder, the oral image data x∈R H×W×3 Input into the backbone network of VGG-16 to generate a set of feature maps:

[0081] in H is the height of the feature map, W is the width of the feature map;

[0082] Specifically, the specific steps of using the multi-scale gradient enhancement module are:

[0083] Feature map features of the input encoder Calculate the horizontal and vertical gradients of the encoder's feature map:

[0084]

[0085] Among them, Sobel x , Sobel y Represents the Sobel operator (horizontal / vertical direction);

[0086] Then, dilated convolution is used to extract the gradient features under different receptive fields:

[0087] F grad =Conv 3×3,dilation＝1 (G x +G y )+Conv 3×3,dilation＝2 (G x +G y );

[0088] After concatenating the original encoder features with the gradient features, enhanced features are generated through convolution:

[0089]

[0090] More specifically, the specific steps for the decoder upsampling and feature fusion process are:

[0091] Multi-scale features of the input encoder and deep low-resolution features;

[0092] Starting from the deepest features, the resolution is gradually restored through transposed convolution:

[0093]

[0094] Concatenate the upsampled features with the PixelSIFT features at the same level:

[0095]

[0096] Refine the fused features through the convolutional layer:

[0097]

[0098] Final output: The decoder end generates a segmentation probability map through 1×1 convolution and Sigmoid function.

[0099] More specifically, in step S3, multi-scale features (such as edges and textures) are extracted through the VGG-16 backbone network to generate a feature map set, and then the high-resolution features of the encoder are fused with the low-resolution features of the decoder, and then the spatial resolution is gradually restored through transposed convolution upsampling to generate a segmentation probability map.

[0100] Furthermore, after the model is confirmed, the learning model is used to predict and a hybrid function is used to balance the predicted and actual lesion segmentation boundaries and classification accuracy: the specific calculation process is:

[0101] L=α·L Dice Loss +(1-α)·L Loss ;

[0102] Among them, LDice Loss The calculation formula is: By L Dice Loss Calculate the overlap between the predicted lesion area and the actual lesion area;

[0103] L Loss The calculation formula is: L Loss Constrained pixel-level classification probability distribution.

[0104] The use of DiceLoss is sensitive to category imbalance (such as a small number of lesion pixels) and can optimize the segmentation boundary; while the use of cross entropy Loss can punish the misclassification of a single pixel and improve the overall accuracy.

[0105] like Figure 2 As shown, in step S4,

[0106] Substitute the validation set and adjust the learning model parameters through Adam optimization:

[0107] m t =β1m t-1 +(1-β1)g t ,

[0108]

[0109] Adjusted parameters

[0110] Output optimized model parameters θ t ;

[0111] Among them, η=0.001, β1=0.9, β2=0.999, L is the mixing function.

[0112] Furthermore, the test set is used to quantify the learning model performance through the Dice coefficient;

[0113]

[0114] Among them, DSC is used to measure the overlap between the predicted area and the real area. The value is between 0 and 1. The closer it is to 1, the higher the overlap. |*| is used to calculate the number of pixels.

[0115] In step S6, the oral image data is input into the optimized learning model and the specific steps of obtaining the results are:

[0116]

[0117] When the output When it is greater than 0.5, it is the lesion area, and when it is less than 0.5, it is the normal area;

[0118] Among them, Softmax is the Softmax activation function, which is used to convert the original score output by the model into a probability distribution (the sum of the probabilities of all pixels is 1), N is the input oral image data, and PixelSIFT-UNet is the optimized learning model.

[0119] An oral mucosal lesion segmentation system, based on any oral mucosal lesion segmentation system, a data preparation module is used to obtain high-quality oral image data, mark the lesion site, and define the lesion area 1 and the normal area 0;

[0120] The PixelSIFT-UNet model building module is used to build a learning model. It uses the Sobel operator and dilated convolution to enhance lesion boundary information and improve the model's sensitivity to blurred boundaries.

[0121] Training strategy module, used to optimize learning model parameters;

[0122] Evaluation and testing module, used to measure the overlap between the predicted area and the true label;

[0123] The reasoning module inputs the oral image data and obtains the final result.

[0124] The above description is based on the ideal embodiment of the present invention. Based on the above description, relevant personnel can make various changes and modifications without departing from the technical scope of this invention. The technical scope of this invention is not limited to the content of the specification and must be determined according to the scope of the claims.

Claims

1. An oral mucosal lesion segmentation system, characterized in that: include: The data preparation module is used to obtain and divide oral image data, mark the lesion area, and define the lesion area 1 and the normal area 0; The PixelSIFT-UNet model building module is used to build a learning model. It uses the Sobel operator and dilated convolution to enhance lesion boundary information and improve the model's sensitivity to blurred boundaries. Training strategy module, used to optimize learning model parameters; Evaluation and testing module, used to measure the overlap between the predicted area and the true label; The reasoning module inputs the oral image data and obtains the final result.

2. The oral mucosal lesion segmentation system according to claim 1, characterized in that: The oral cavity image data is a white light image of the oral cavity.

3. The oral mucosal lesion segmentation system according to claim 1, characterized in that: Partitioning is to divide the oral image data into training set, validation set and test set according to a certain ratio of 7:2:

1.

4. The oral mucosal lesion segmentation system according to claim 1, characterized in that: The lesion site is marked by using Labelme software to mark the lesion on the training set images, and the marked lesion site is recorded as Y mask , specifically: use the binary label map Y, where, 5. The oral mucosal lesion segmentation system according to claim 1, characterized in that: The deep learning model is constructed based on PixelSIFT-UNet through lesion data. Specifically, this model generates a segmentation probability map of the lesion area; specifically: Among them, Encode(x) is the multi-scale feature extracted by the encoder; PixelSIFT(*) is a multi-scale gradient enhancement module in skip connections; Decode(*) is the upsampling and feature fusion process of the decoder; σ is the sigmoid function.

6. The oral mucosal lesion segmentation system according to claim 5, characterized in that: For the multi-scale features extracted by the encoder, the oral image data x∈R H×W×3 Input into the backbone network of VGG-16 to generate a set of feature maps: in The specific steps of using the multi-scale gradient enhancement module are: Feature map features of the input encoder Calculate the horizontal and vertical gradients of the encoder's feature map: Among them, Sobel x , Sobel y Represents the Sobel operator (horizontal / vertical direction); Then, dilated convolution is used to extract the gradient features under different receptive fields: F grad =Conv 3×3,dilation＝1 (G x +G y )+Conv 3×3,dilation＝2 (G x +G y ); After concatenating the original encoder features with the gradient features, enhanced features are generated through convolution:

7. The oral mucosal lesion segmentation system according to claim 6, characterized in that: The specific steps of the decoder upsampling and feature fusion process are: Multi-scale features of the input encoder and deep low-resolution features; Starting from the deepest features, the resolution is gradually restored through transposed convolution: Concatenate the upsampled features with the PixelSIFT features at the same level: Refine the fused features through the convolutional layer: Final output: The decoder end generates a segmentation probability map through 1×1 convolution and Sigmoid function.

8. The oral mucosal lesion segmentation system according to claim 7, characterized in that: Substitute the validation set and adjust the learning model parameters through Adam optimization: m t =β1m t-1 +(1-β1)g t Adjusted parameters Output optimized model parameters θ t ; Among them, η=0.001, β1=0.9, β2=0.999, L is the mixing function.

9. The oral mucosal lesion segmentation system according to claim 8, characterized in that: The specific steps to input oral image data and obtain the final result are: When the output When it is greater than 0.5, it is the lesion area, and when it is less than 0.5, it is the normal area.

10. A method for segmenting oral mucosal lesions, according to the oral mucosal lesion segmentation system according to any one of claims 1 to 9, characterized in that: S1. Obtain oral image data and divide it into training set, validation set and test set according to a certain ratio; S2. Label the lesion locations on the training set data to obtain lesion data; S3. Build a deep learning model based on PixelSIFT-UNet using lesion data; S4. Use the validation set to optimize the deep learning model parameters and evaluate the deep learning model performance; S5. Input the oral image data into the optimized learning model and obtain the results.

Images