Acupuncture point positioning method and system based on artificial intelligence

Abstract

The invention discloses an acupuncture point positioning method and system based on artificial intelligence, and relates to the technical field of acupuncture, first, RGB images of a to-be-positioned acupuncture point area are collected, LabelImg is used for marking to obtain a data set, pre-processing such as graying, multi-coordinate conversion, singular value decomposition, edge detection and morphological processing is carried out, and the RGB images of the to-be-positioned acupuncture point area are obtained; and inputting the data into an acupuncture point positioning model consisting of a multi-scale Transform feature extraction backbone network, a pyramid multi-scale feature fusion module and a positioning output layer for positioning. By means of the artificial intelligence technology, the problems that traditional acupuncture point positioning depends on doctor experience, efficiency is low, errors are large and the like are solved, by enhancing image features, the accuracy of subsequent positioning is improved, features of different levels are effectively extracted and fused, positioning is more accurate, automatic processing is achieved, and the positioning efficiency is improved. The efficiency, objectivity and consistency of acupoint positioning are improved, and the method has important significance on clinical application of acupuncture and moxibustion.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of acupuncture, more particularly to an acupuncture point positioning method and system based on artificial intelligence. BACKGROUND

[0002] At present, acupuncture is an important part of traditional medicine and has a wide range of applications in clinical treatment. Accurate point positioning is a key prerequisite for acupuncture treatment. At present, acupuncture point positioning mainly relies on the experience and manual touch of doctors. This approach has the following problems: first, it requires high experience of doctors, and the positioning results of different doctors may differ, resulting in unstable treatment effect; second, manual touch positioning is low in efficiency and difficult to achieve rapid and accurate positioning; third, for some deep or surface features not obvious points, positioning is difficult and errors are prone to occur. With the development of artificial intelligence technology, it has important research significance and practical application value to apply it to acupuncture point positioning.

[0003] Therefore, how to realize the rapid, accurate and objective positioning of acupuncture points is a problem that needs to be solved by those skilled in the art. SUMMARY

[0004] Therefore, the present application provides an acupuncture point positioning method and system based on artificial intelligence to solve the problems in the background art.

[0005] In order to achieve the above purpose, the present application adopts the following technical scheme:

[0006] An acupuncture point positioning method based on artificial intelligence, comprising:

[0007] Collecting an RGB image of a point to be positioned;

[0008] Labeling the RGB image of the point to be positioned using LabelImg labeling tool to obtain a point to be positioned image dataset;

[0009] Preprocessing the images in the point to be positioned image dataset to obtain a preprocessed point to be positioned image dataset;

[0010] Inputting the preprocessed point to be positioned image dataset into a pre-trained acupuncture point positioning model, positioning the points in the point to be positioned image through the acupuncture point positioning model, and outputting the acupuncture point positioning result.

[0011] Optionally, preprocessing the images in the point to be positioned image dataset to obtain a preprocessed point to be positioned image dataset, comprising:

[0012] The images in the to-be-positioned acupoint image dataset are subjected to grayscale processing to obtain a grayscale image dataset;

[0013] The image sequence in the grayscale image dataset is converted from a three-dimensional matrix to a two-dimensional coordinate system matrix by using a multi-coordinate conversion method;

[0014] The data in the converted image pixels are arranged in sequence, and the obtained column vector is subjected to matrix conversion processing to obtain a one-dimensional matrix;

[0015] The one-dimensional sequence is sequentially sorted, a new two-dimensional matrix is constructed, and the new two-dimensional matrix is subjected to singular value decomposition theorem decomposition standardization processing;

[0016] The two-dimensional matrix after the singular value decomposition theorem decomposition standardization processing is converted to obtain an image-enhanced to-be-positioned acupoint image dataset.

[0017] Optionally, the images in the to-be-positioned acupoint image dataset are preprocessed to obtain a preprocessed to-be-positioned acupoint image dataset, and the method further comprises:

[0018] The images in the image-enhanced to-be-positioned acupoint image dataset are read, and edge detection is performed on the images by using a Canny edge detection algorithm to obtain an edge image;

[0019] The edge features in the edge image are enhanced by applying a Laplacian filter to obtain an enhanced edge image;

[0020] The gradient direction and size of each pixel point in the enhanced edge image are traversed, and an image thinning algorithm is used to perform edge thinning on the enhanced edge image to obtain a thinned edge image;

[0021] The thinned edge image is processed by using a morphological dilation and erosion method to obtain the preprocessed to-be-positioned acupoint image dataset.

[0022] Optionally, the acupuncture acupoint positioning model is composed of a multi-scale Transformer feature extraction backbone network, a pyramid multi-scale feature fusion module, and a positioning output layer.

[0023] Optionally, the multi-scale Transformer feature extraction backbone network is composed of a stem module and four feature extraction stage modules, and the four feature extraction stage modules are sequentially a first feature extraction stage module, a second feature extraction stage module, a third feature extraction stage module, and a fourth feature extraction stage module.

[0024] The first feature extraction stage module sequentially comprises a PatchEmbedding module, a SwinTransformer-Encoder module and a down-sampling layer, the input of the PatchEmbedding module is the output of the stem module, the SwinTransformer-Encoder module obtains the output of the PatchEmbedding module, and the feature processed by the SwinTransformer-Encoder module is processed by the down-sampling layer to obtain a first feature map M1;

[0025] The second feature extraction stage module comprises a SwinTransformer-Encoder module and a down-sampling layer; the input of the SwinTransformer-Encoder module is the first feature map M1 obtained by the first feature extraction stage module.

[0026] The third feature extraction stage module comprises a SwinTransformer-Encoder module and a down-sampling layer; the input of the SwinTransformer-Encoder module is the second feature map M2 obtained by the second feature extraction stage module.

[0027] The fourth feature extraction stage module comprises a SwinTransformer-Encoder module and a down-sampling layer; the input of the SwinTransformer-Encoder module is the third feature map M3 obtained by the third feature extraction stage module; and the fourth feature extraction stage module outputs a fourth feature map M4.

[0028] Optionally, the pyramid multi-scale feature fusion module fuses the first feature map M1 output by the first feature extraction stage module and the second feature map M2 output by the second feature extraction stage module to obtain a first fused feature map M2'; the first fused feature map M2' is fused with the third feature map M3 output by the third feature extraction stage module after being processed by a down-sampling layer to obtain a second fused feature map M3'; the second fused feature map M3' is fused with the fourth feature map M4 output by the fourth feature extraction stage module after being processed by a down-sampling layer to obtain a third fused feature map M4'; and the third fused feature map M4' is processed by a down-sampling layer to obtain a fifth feature map M5; the first fused feature map M2', the second fused feature map M3', the third fused feature map M4' and the fifth feature map M5 are input into the positioning output layer.

[0029] Optionally, the positioning output layer is formed by cascading two convolution layers, and outputs an acupuncture point positioning result.

[0030] Optionally, the SwinTransformer-Encoder module is composed of two consecutive SwinTransformerBlock, which are sequentially sent to the window self-attention mechanism after the normalization layer, and then input to the MLP module after the normalization layer; then sent to the offset window self-attention mechanism after the normalization layer, and then input to the MLP module after the normalization layer, wherein there are residual links after the self-attention mechanism and the MLP module.

[0031] An acupuncture point positioning system based on artificial intelligence, comprising:

[0032] A data acquisition module acquires an RGB image of a region to be positioned;

[0033] A data labeling module labels the RGB image and the depth image of the region to be positioned using a LabelImg labeling tool to obtain an image dataset of the region to be positioned;

[0034] A data preprocessing module pre-processes the images in the image dataset of the region to be positioned to obtain a pre-processed image dataset of the region to be positioned;

[0035] A positioning result output module inputs the pre-processed image dataset of the region to be positioned into an acupuncture point positioning model pre-trained to position the acupuncture point in the image of the region to be positioned, and outputs an acupuncture point positioning result.

[0036] According to the above technical solution, compared with the prior art, the present application provides an acupuncture point positioning method and system based on artificial intelligence, which first acquires an RGB image of a region to be positioned, obtains a dataset using LabelImg labeling, and then pre-processes the dataset through grayscale, multi-coordinate conversion, singular value decomposition, edge detection, morphological processing, etc., and then inputs the data into an acupuncture point positioning model composed of a multi-scale Transformer feature extraction backbone network, a pyramid multi-scale feature fusion module, and a positioning output layer for positioning. The present application uses artificial intelligence technology to solve the problems of traditional acupuncture point positioning, such as reliance on doctor's experience, low efficiency, large error, etc. By enhancing image features, the accuracy of subsequent positioning is improved, and different levels of features are effectively extracted and fused, making the positioning more accurate and achieving automatic processing, improving the efficiency, objectivity and consistency of acupuncture point positioning, and having important significance for acupuncture clinical application. BRIEF DESCRIPTION OF DRAWINGS

[0037] In order to make the technical solutions in the embodiments of the present application or the prior art clearer, the accompanying drawings needed in the embodiments or prior art description will be briefly introduced below. Obviously, the accompanying drawings in the following description only aim to explain part of the embodiments of the present application, and not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the protection scope of the present application.

[0038] Figure 1 The method flowchart provided by the present application is shown in the figure. DETAILED DESCRIPTION

[0039] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, and not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the protection scope of the present application.

[0040] The embodiment of the present application discloses an acupuncture point positioning method based on artificial intelligence, as shown in the figure, comprising: Figure 1

[0041] Collecting an RGB image of a region to be positioned;

[0042] Labeling the RGB image of the acupuncture point to be positioned using LabelImg labeling tool to obtain an acupuncture point image data set to be positioned;

[0043] Preprocessing the images in the acupuncture point image data set to be positioned to obtain a preprocessed acupuncture point image data set to be positioned;

[0044] Inputting the preprocessed acupuncture point image data set to be positioned into the acupuncture point positioning model pre-trained to position the acupuncture point in the acupuncture point image to be positioned, and outputting the acupuncture point positioning result.

[0045] In one specific embodiment, preprocessing the images in the acupuncture point image data set to be positioned to obtain a preprocessed acupuncture point image data set to be positioned, comprising:

[0046] Gray processing the images in the acupuncture point image data set to be positioned to obtain a gray image data set;

[0047] Converting the image sequence in the gray image data set from a three-dimensional matrix to a two-dimensional coordinate system matrix by using a multi-coordinate conversion method;

[0048] Arranging the data in the converted image pixels in sequence, and performing matrix conversion processing on the obtained column vector to obtain a one-dimensional matrix;

[0049] ​The one-dimensional sequence is sequentially sorted to form a new two-dimensional matrix, and the new two-dimensional matrix is subjected to singular value decomposition theorem decomposition standardization processing;

[0050] The two-dimensional matrix subjected to singular value decomposition theorem decomposition standardization processing is transformed to obtain an image-enhanced to-be-positioned acupoint image dataset.

[0051] In one specific embodiment, the images in the to-be-positioned acupoint image dataset are preprocessed to obtain a preprocessed to-be-positioned acupoint image dataset, which further comprises:

[0052] The images in the image-enhanced to-be-positioned acupoint image dataset are read, and edge detection is performed on the images by using a Canny edge detection algorithm to obtain an edge image;

[0053] The edge features in the edge image are enhanced by applying a Laplacian filter to obtain an enhanced edge image;

[0054] The gradient direction and size of each pixel point in the enhanced edge image are traversed, and an image thinning algorithm is used to thin the edges of the enhanced edge image to obtain a thinned edge image;

[0055] The thinned edge image is processed by using a morphological dilation and erosion method to obtain a preprocessed to-be-positioned acupoint image dataset.

[0056] In one specific embodiment, the acupuncture acupoint positioning model is composed of a multi-scale Transformer feature extraction backbone network, a pyramid multi-scale feature fusion module, and a positioning output layer.

[0057] A part of the preprocessed to-be-positioned acupoint image dataset is extracted as a training set and fed into the acupuncture acupoint positioning model for training: the multi-scale Transformer feature extraction backbone network performs feature extraction, the pyramid multi-scale feature fusion module performs scale fusion, the obtained fused feature map is input into the positioning output layer, and the positioning output layer completes the final acupuncture acupoint positioning result output; the training error is subjected to gradient descent during the training process, the trainable weight parameter learning is completed, and the trained acupuncture acupoint positioning model is obtained.

[0058] In one specific embodiment, the multi-scale Transformer feature extraction backbone network is composed of a stem module and four feature extraction stage modules, and the four feature extraction stage modules are sequentially a first feature extraction stage module, a second feature extraction stage module, a third feature extraction stage module, and a fourth feature extraction stage module.

[0059] The first feature extraction stage module sequentially comprises a PatchEmbedding module, a SwinTransformer-Encoder module and a down-sampling layer, the input of the PatchEmbedding module is the output of the stem module, the SwinTransformer-Encoder module obtains the output of the PatchEmbedding module, and the feature processed by the SwinTransformer-Encoder module is processed by the down-sampling layer to obtain the first feature map M1;

[0060] The second feature extraction stage module comprises a SwinTransformer-Encoder module and a down-sampling layer; the input of the SwinTransformer-Encoder module of the second feature extraction stage module is the first feature map M1 obtained by the first feature extraction stage module.

[0061] The third feature extraction stage module comprises a SwinTransformer-Encoder module and a down-sampling layer; the input of the SwinTransformer-Encoder module of the third feature extraction stage module is the second feature map M2 obtained by the second feature extraction stage module.

[0062] The fourth feature extraction stage module comprises a SwinTransformer-Encoder module and a down-sampling layer; the input of the SwinTransformer-Encoder module of the fourth feature extraction stage module is the third feature map M3 obtained by the third feature extraction stage module; and the fourth feature extraction stage module outputs a fourth feature map M4.

[0063] In one specific embodiment, the pyramid multi-scale feature fusion module fuses the first feature map M1 output by the first feature extraction stage module and the second feature map M2 output by the second feature extraction stage module to obtain a first fused feature map M2'; the first fused feature map M2' is fused with the third feature map M3 output by the third feature extraction stage module after being processed by a down-sampling layer to obtain a second fused feature map M3'; the second fused feature map M3' is fused with the fourth feature map M4 output by the fourth feature extraction stage module after being processed by a down-sampling layer to obtain a third fused feature map M4'; and the third fused feature map M4' is processed by a down-sampling layer to obtain a fifth feature map M5; the first fused feature map M2', the second fused feature map M3', the third fused feature map M4' and the fifth feature map M5 are input into the positioning output layer.

[0064] In one specific embodiment, the positioning output layer is formed by cascading two convolution layers, and outputs an acupuncture point positioning result.

[0065] In a specific embodiment, the SwinTransformer-Encoder module consists of two consecutive SwinTransformerBlocks, which are sequentially fed into the window self-attention mechanism after passing through the normalization layer, and then fed into the normalization layer and input into the MLP module; then fed into the offset window self-attention mechanism after passing through the normalization layer, and then fed into the normalization layer and input into the MLP module, where there are residual links after the self-attention mechanism and the MLP module.

[0066] In the present invention, the PatchEmbedding module cuts the pre-processed RGB image of the acupuncture point area to be located into windows of size "image block" * "image block" and then embeds them. The PatchEmbedding module mainly includes a downsampling layer and a 1×1 convolution, which adjusts the channel dimension to twice the original. Here, a two-dimensional convolution layer is used to set the convolution kernel and stride to the size of the "image block". The output channel is set to determine the size of the embedding vector. Finally, the height (H) and width (W) dimensions are expanded and moved to the first dimension to generate a series of "image blocks".

[0067] In the present invention, the receptive field of the first feature extraction stage (Stage 1) module is relatively small. The pre-processed RGB image of the acupoint area to be located is processed by the stem module, which is conducive to the stability of training. The stem module is essentially a 7×7 convolution layer with a step size of 2.

[0068] The SwinTransformer-Encoder module consists of two consecutive SwinTransformerBlocks, which are fed into the window self-attention mechanism (W-MSA) after passing through the normalization layer, and then into the MLP module after passing through the normalization layer; then, they are fed into the offset window self-attention mechanism (SW-MSA) after passing through the normalization layer, and then into the MLP module after passing through the normalization layer. There are residual links after the self-attention mechanism and the MLP module.

[0069] In the present invention, the downsampling layer is a convolution with a convolution kernel size of 1×1 and a stride of 2.

[0070] After each down-sampling, the same scale feature map extracted by the corresponding feature extraction network is fused, the four feature maps output by the first feature extraction stage module, the second feature extraction stage module, the third feature extraction stage module and the fourth feature extraction stage module are selected for multi-scale feature fusion, and the down-sampling layer generates an effective feature map, and the specific operation is as follows: assuming that the original feature map WxHxC (W, H and C represent width, height and channel number respectively), the first feature extraction stage (Stage1) module is input to obtain the first feature map M1W / 2xH / 2x2C, and the first feature map M1 becomes the input of the pyramid multi-scale feature fusion module, and after passing through a down-sampling layer, the first feature map M1 is down-sampled to W / 4xH / 4x4C; the first feature map M1 is also down-sampled to the second feature map M2W / 4xH / 4x4C after passing through the second feature extraction stage (Stage2) module, the first feature map M1 is fused with the second feature map M2 to obtain the first fused feature map M2', and the specific fusion method is addition operation; the first fused feature map M2' is fused with the third feature map M3 after down-sampling to obtain the second fused feature map M3', and the second fused feature map M3' is fused with the fourth feature map M4 after down-sampling to obtain the second fused feature map M4', which can retain more detailed feature information and increase the representation ability of the feature map.

[0071] In the application, different prior boxes are set for the first fused feature map M2', the second fused feature map M3', the third fused feature map M4' and the fifth feature map M5, the number of prior boxes is set as [3, 4, 4, 4], and the prior boxes are set in terms of scale and aspect ratio. Different sizes of prior boxes are obtained by using the K-Means clustering algorithm, for the first fused feature map M2', the second fused feature map M3', the third fused feature map M4' and the fifth feature map M5, according to the different number of prior box settings, a total of 15 prior boxes can be obtained by the clustering algorithm, and the regression box used for prediction is determined by using the logistic regression.

[0072] The first fused feature map M2', the second fused feature map M3', the third fused feature map M4' and the fifth feature map M5 are input into the positioning output layer to generate a prediction box, and the prediction box is compared and verified with a plurality of prior boxes established for each grid center to determine whether the selected region contains the final recognized and positioned acupuncture point; after continuous adjustment between the prior box and the prediction box, the region edge is fine-tuned to obtain an accurate bounding box.

[0073] The prediction box generated by the positioning output layer is adjusted to obtain an accurate bounding box, a confidence threshold is set, each category is judged, and a result with a confidence score greater than the threshold is output, including the prediction box and the score; local maximum values are found by using a non-maximum suppression algorithm, and the remaining values in the neighborhood are filtered out (suppressed); the highest scoring prediction box is selected from the stacked bounding boxes using non-maximum suppression, and a final acupuncture point positioning result is generated, including the prediction box coordinate position, confidence, and category probability.

[0074] An artificial intelligence-based acupuncture point positioning system, comprising:

[0075] A data acquisition module acquires an RGB image of an acupuncture point region to be positioned.

[0076] A data labeling module labels the RGB image and depth image of the acupuncture point to be positioned using a LabelImg labeling tool to obtain an acupuncture point image data set to be positioned.

[0077] A data preprocessing module pre-processes the images in the acupuncture point image data set to be positioned to obtain a pre-processed acupuncture point image data set to be positioned.

[0078] A positioning result output module inputs the pre-processed acupuncture point image data set to be positioned into an acupuncture point positioning model trained in advance, positions the acupuncture point in the acupuncture point image to be positioned through the acupuncture point positioning model, and outputs an acupuncture point positioning result.

[0079] The various embodiments in the specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be mutually referred to. For the device disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the method part.

[0080] The above description of the disclosed embodiments enables a person skilled in the art to implement or use the present application. Various modifications to the embodiments will be apparent to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. An artificial intelligence-based acupuncture point positioning method, characterized in that: include: Collect RGB images of the acupoint area to be located; The LabelImg labeling tool is used to label the RGB image of the acupoint to be located to obtain an image dataset of the acupoint to be located; Preprocessing the images in the acupoint image dataset to be located to obtain a preprocessed acupoint image dataset to be located; The preprocessed acupoint image data set to be located is input into a pre-trained acupuncture point location model, the acupoints in the acupoint image to be located are located by the acupuncture point location model, and an acupuncture point location result is output.

2. The acupuncture point positioning method based on artificial intelligence according to claim 1, characterized in that: Preprocessing the images in the acupoint image dataset to be located to obtain a preprocessed acupoint image dataset to be located, including: grayscale the images in the acupoint image dataset to be located to obtain a grayscale image dataset; The multi-coordinate transformation method is used to transform the image sequence in the grayscale image data set from a three-dimensional matrix into a two-dimensional coordinate system matrix; Arrange the data in the converted image pixels in sequence, and perform matrix transformation on the obtained column vector to obtain a one-dimensional matrix; After sorting the one-dimensional sequences in sequence to form a new two-dimensional matrix, the new two-dimensional matrix is ​​decomposed and standardized using the singular value decomposition theorem; The two-dimensional matrix after normalization is decomposed according to the singular value decomposition theorem and transformed to obtain the image dataset of the acupoints to be located after image enhancement.

3. The acupuncture point positioning method based on artificial intelligence according to claim 1, characterized in that: Preprocessing the images in the acupoint image dataset to be located to obtain a preprocessed acupoint image dataset to be located further includes: Read the image in the acupoint image dataset to be located after image enhancement, and perform edge detection using the Canny edge detection algorithm to obtain an edge image; enhancing edge features in the edge image by applying a Laplace filter to obtain an enhanced edge image; Traversing the gradient direction and magnitude of each pixel in the enhanced edge image, and using an image thinning algorithm to thin the edge of the enhanced edge image to obtain a thinned edge image; The morphological dilation and erosion methods are used to process the edge image after thinning to obtain the preprocessed acupoint image dataset to be located.

4. The artificial intelligence-based acupuncture point positioning method according to claim 1, characterized in that: The acupuncture point positioning model consists of a multi-scale Transformer feature extraction backbone network, a pyramid multi-scale feature fusion module, and a positioning output layer.

5. The artificial intelligence-based acupuncture point positioning method according to claim 4, characterized in that: The multi-scale Transformer feature extraction backbone network consists of a stem module and four feature extraction stage modules, which are the first feature extraction stage module, the second feature extraction stage module, the third feature extraction stage module, and the fourth feature extraction stage module. The first feature extraction stage module includes a PatchEmbedding module, a SwinTransformer-Encoder module and a downsampling layer in sequence. The input of the PatchEmbedding module is the output of the stem module. The SwinTransformer-Encoder module obtains the output of the PatchEmbedding module. The features processed by the SwinTransformer-Encoder module are processed by the downsampling layer to obtain the first feature map M1; The second feature extraction stage module includes a SwinTransformer-Encoder module and a downsampling layer; the input of the SwinTransformer-Encoder module is the first feature map M1 obtained by the first feature extraction stage module; The third feature extraction stage module includes a SwinTransformer-Encoder module and a downsampling layer; the input of the SwinTransformer-Encoder module is the second feature map M2 obtained by the second feature extraction stage module; The fourth feature extraction stage module includes a SwinTransformer-Encoder module and a downsampling layer; the input of its SwinTransformer-Encoder module is the third feature map M3 obtained by the third feature extraction stage module; the fourth feature extraction stage module outputs a fourth feature map M4.

6. The artificial intelligence-based acupuncture point positioning method according to claim 4, characterized in that: The pyramid multi-scale feature fusion module fuses the first feature map M1 output by the first feature extraction stage module and the second feature map M2 output by the second feature extraction stage module to obtain a first fused feature map M2′; the first fused feature map M2′ is fused with the third feature map M3 output by the third feature extraction stage module after passing through a downsampling layer to obtain a second fused feature map M3′; the second fused feature map M3′ is fused with the fourth feature map M4 output by the fourth feature extraction stage module after passing through a downsampling layer to obtain a third fused feature map M4′; the third fused feature map M4′ is fused with the fifth feature map M5 after passing through a downsampling layer; the first fused feature map M2′, the second fused feature map M3′, the third fused feature map M4′ and the fifth feature map M5 are input into the positioning output layer.

7. The artificial intelligence-based acupuncture point positioning method according to claim 4, characterized in that: The positioning output layer is formed by cascading two convolutional layers to output the acupuncture point positioning results.

8. The artificial intelligence-based acupuncture point positioning method according to claim 5, characterized in that: The SwinTransformer-Encoder module consists of two consecutive SwinTransformerBlocks, which are fed into the window self-attention mechanism after passing through the normalization layer, and then into the MLP module after passing through the normalization layer; then, they are fed into the offset window self-attention mechanism after passing through the normalization layer, and then into the MLP module after passing through the normalization layer. There are residual links after the self-attention mechanism and the MLP module.

9. An artificial intelligence-based acupuncture point positioning system, characterized in that: An artificial intelligence-based acupuncture point positioning method according to any one of claims 1 to 8 is applied, comprising: Data acquisition module, collecting RGB images of the acupoint area to be located; A data labeling module uses the LabelImg labeling tool to label the RGB image and depth image of the acupoint to be located to obtain an image dataset of the acupoint to be located; a data preprocessing module, which preprocesses the images in the acupoint image dataset to be located to obtain a preprocessed acupoint image dataset to be located; The positioning result output module inputs the preprocessed acupoint image data set to be positioned into the pre-trained acupuncture point positioning model, locates the acupoints in the acupoint image to be positioned using the acupuncture point positioning model, and outputs the acupuncture point positioning results.

Images