A lung CT visualization method and system

Abstract

The present invention discloses a lung CT visualization method and system, which includes a PACS system module, a TransU‑Net network module that introduces a gated axial attention mechanism, a pixel adjustment module, a dimension adjustment module, an agglomerative hierarchical clustering module, a feature fusion module, and a visualization module. The PACS system module is used to store, retrieve, transmit and display medical images; the improved TransU‑Net network module is used for basic image segmentation; the pixel adjustment module is used for image pixel multiplication and restoration; the dimension adjustment module is used for image dimension increase and reduction; the agglomerative hierarchical clustering module is used for image clustering segmentation after pixel multiplication and dimension increase; the feature fusion module is used for the fusion of basic segmented images and clustered segmented images; the visualization module is used for pseudo-color processing of fused images, and finally outputs a visualized image result. The image segmentation and visualization based on TransU‑Net and Agglomerative Clustering fusion in the present invention can be used for lung CT image segmentation, optimize the segmentation effect of the basic U‑Net, and achieve high-precision segmentation and visualization of the lung area.

Description

Technical Field

[0001] The present invention belongs to the technical field of medical image processing, and in particular relates to a lung CT visualization method and system based on improved TransU-Net supplemented by Agglomerative Clustering to optimize image segmentation. Background Art

[0002] With the continuous advancement of computer vision, deep learning has become the mainstream technology in the field of medical image processing. In particular, by 2015, the introduction of the U-Net network improved the segmentation performance of medical images and redefined the development direction of deep learning in the field of medical image segmentation. Most of the existing segmentation models for lung problems are improved on the U-Net network, but because the shape of the target area of ​​​​the lung CT image varies from person to person, and because of privacy issues, the data set that can train the U-Net network is limited, resulting in incomplete feature extraction and unsatisfactory segmentation accuracy. At the same time, most existing CT images are grayscale images, which are prone to visual fatigue, leading to misdiagnosis.

[0003] The existing TransU-Net-based image segmentation solves the local inductive bias and can provide sufficient local information representation for global modeling. However, when TransU-Net directly applies the ViT-16 model to process three-dimensional images, the computational complexity increases significantly, accompanied by the risk of overfitting, and small-scale data sets are difficult to support Transformer pre-training consumption, which can easily lead to reduced model performance, especially when processing 2D slices of three-dimensional images, contextual spatial information will be lost. TransU-Net, which introduces the gated axial attention mechanism, can reduce the computational complexity, but still loses contextual spatial information. The Agglomerative Clustering hierarchical clustering algorithm has significant advantages in image boundary segmentation and can make up for the problem of loose context connection between three-dimensional image boundaries, but it cannot extract internal features of the image. Summary of the invention

[0004] The present invention aims to effectively improve the problems existing in the above prior art, and proposes a lung CT visualization method and system based on improved TransU-Net supplemented by Agglomerative Clustering to optimize image segmentation. The present invention introduces a gated axial attention mechanism to improve TransU-Net to reduce computational complexity, and supplemented by Agglomerative Clustering to connect contextual spatial information, increase inter-slice pixel dependencies, improve the accuracy of boundary segmentation of the target area, and effectively suppress over-segmentation or under-segmentation. Finally, the result image is converted into RGB format through pseudo-color processing, which can provide more intuitive lung area information, and is more convenient to observe the location, size and morphological characteristics of the lesion than grayscale images.

[0005] A lung CT visualization method provided by the present invention is as follows:

[0006] Step 1: The PACS system module converts the medical images generated after CT scanning into a standard data format for storage;

[0007] Step 2: Convert the standard data format into an operable image format and proceed to step 3 and step 4 for different processing;

[0008] Step 3: Import the operable image format into the TransU-Net network module that introduces the gated axial attention mechanism to extract and process image features and complete basic segmentation;

[0009] Step 4: Import the operable image format into the pixel adjustment module for pixel multiplication; import the grayscale image after pixel multiplication into the dimension adjustment module for three-dimensional reconstruction to generate a three-dimensional grayscale image; then cluster and segment the target area of ​​the three-dimensional grayscale image through the agglomerative hierarchical clustering module; import the obtained three-dimensional cluster segmentation image into the dimension adjustment module for two-dimensional restoration; then import the cluster segmentation image after dimension restoration into the pixel adjustment module for pixel restoration;

[0010] Step 5: Import the basic segmentation image obtained in step 3 and the cluster segmentation image obtained in step 4 into the feature fusion module for image fusion. Take the segmentation line in the cluster segmentation image as a reference, optimize the boundary of the basic segmentation image, and obtain the optimized fused segmentation image.

[0011] Step 6: Import the fused segmented image into the visualization module for pseudo-color processing and convert it into a visual RGB image format;

[0012] Step 7: Output the visualized RGB image through the imaging workstation.

[0013] Specifically, in step 3, basic segmentation is performed on the imported operable image format, including:

[0014] Step 3.1: Embed the image into a convolutional neural network layer to extract hidden features;

[0015] Step 3.2: Put the extracted hidden features into the transformer layer encoding with the gated axial attention mechanism to obtain the processed hidden feature sequence;

[0016] Step 3.3: Reshape the features of the sequence;

[0017] Step 3.4: Restore the image through the upsampling layer.

[0018] In step 4, cluster segmentation is performed on the target area of ​​the three-dimensional grayscale image, including:

[0019] Step 4.1: Treat all data points as separate clusters;

[0020] Step 4.2: Calculate the distance between all adjacent data points and select the two clusters with the smallest distance to merge;

[0021] Step 4.3: Update the clustering tree and recalculate the distances between the new clusters;

[0022] Step 4.4: Repeat steps 4.2 and 4.3 until all data points reach the preset number of clusters.

[0023] In step 6, pseudo color processing is performed on the fused segmented image by converting the grayscale image into an index image through a grayscale to color transformation method, and then mapping the grayscale value of each pixel to a color in the RGB color space through color transformation, thereby achieving pseudo color processing.

[0024] In step 3.2, the extracted hidden features are put into the transformer layer that introduces the gated axial attention mechanism for encoding, including:

[0025] Step 3.2.1: Assume that the spatial resolution of image x is H×W and the number of channels is C; reshape the input x processed by the convolutional neural network layer into a series of two-dimensional slices Among them, x p represents slices, P represents the side length of each slice, and the size of each slice is P×P. is the number of image slices, i.e. the input sequence length;

[0026] Step 3.2.2: Vectorize the slice x by trainable linear projection p Mapped to the potential D-dimensional embedding space, the slice embedding formula is expressed as Among them, the size of the D-dimensional embedding space should be greater than P×P, is the slice embedding method, It is the position embedding method;

[0027] Step 3.2.3: After the slice is embedded, it is encoded by the Transformer encoder. For the Lth layer of the encoder, its input is recorded as Z l-1 , the output is Z l , the calculation process is:

[0028] z' l =MGAA(NORM(z l-1 ))+z l-1 , l=1,…,L

[0029] z l =MLP(NORM(z' l ))+z' l , l=1,…,L

[0030] Where MGAA is the gated axial attention mechanism layer, MLP is the multi-layer perceptron layer, and NORM represents the layer normalization operator, so Z L Represents a hidden feature sequence.

[0031] The present invention also proposes a lung CT visualization system, including a PACS system module, a TransU-Net network module improved by introducing a gated axial attention mechanism, a pixel adjustment module, a dimension adjustment module, a cohesive hierarchical clustering module, a feature fusion module, and a visualization module;

[0032] The PACS system module is used to convert the medical images generated after CT scanning into a standard data format for storage, and convert the standard data format into an operable image format, and send them to the TransU-Net network module and the pixel adjustment module respectively, which are improved by introducing a gated axial attention mechanism;

[0033] The TransU-Net network module improved by introducing the gated axial attention mechanism is used to extract and process image features, complete basic image segmentation, and send it to the feature fusion module;

[0034] The pixel adjustment module is used to multiply and restore the pixels of the original image; the dimension adjustment module is used to increase and reduce the dimension of the image; the agglomerative hierarchical clustering module is used to cluster and segment the target area of ​​the three-dimensional grayscale image;

[0035] Import the operable image format into the pixel adjustment module for pixel multiplication; import the grayscale image after pixel multiplication into the dimension adjustment module for three-dimensional reconstruction to generate a three-dimensional grayscale image; then cluster and segment the target area of ​​the three-dimensional grayscale image through the agglomerative hierarchical clustering module; import the obtained three-dimensional cluster segmentation image into the dimension adjustment module for two-dimensional restoration; then import the cluster segmentation image after dimension restoration into the pixel adjustment module for pixel restoration and send it to the feature fusion module;

[0036] The feature fusion module is used to fuse the basic segmentation image and the cluster segmentation image output by the pixel adjustment module, and optimize the boundary of the basic segmentation image based on the segmentation line in the cluster segmentation image to obtain an optimized fused segmentation image;

[0037] The visualization module is used to perform pseudo-color processing on the fused segmented image and convert it into a visualized RGB image format.

[0038] The beneficial effects of the present invention are:

[0039] The present invention extracts and processes image features by introducing the improved TransU-Net network with a gated axial attention mechanism to complete basic image segmentation. The TransU-Net before the improvement was limited by the size of the medical image dataset, which resulted in some degradation of the Transformer's performance. The present invention introduces a gated axial attention mechanism into the self-attention module, controls the information flow by adjusting the embedding positions of the query (Quary, Q), key (Key, K), and value (Value, V), limits irrelevant position information, reduces the computational complexity of the self-attention, and enables the model to automatically adjust the gate parameters according to the size of the dataset during training.

[0040] The present invention performs clustering segmentation on the three-dimensional grayscale image after pixel multiplication through Agglomerative Clustering. The advantage of this is that, compared with other clustering algorithms, Agglomerative Clustering can automatically determine the number of clusters and process noise points; compared with using TransU-Net alone, by using clustering segmentation for auxiliary segmentation in the three-dimensional grayscale image, the contextual spatial information can be better connected, which makes up for the disadvantage that Transformer loses the contextual spatial information when processing the 2D slices of the three-dimensional image, so that the segmentation of the target area is more accurate and the edge error is reduced.

[0041] The present invention uses a grayscale-color transformation method through pseudo-color processing to convert a grayscale image into an index image, and then maps the grayscale value of each pixel to a color in the RGB color space through color transformation, thereby realizing pseudo-color processing, which can provide more intuitive lung area information and is more convenient for observing the location, size and morphological characteristics of the lesion than grayscale images. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 It is a system block diagram of the present invention.

[0043] Figure 2 The present invention is a flow chart of the method.

[0044] Figure 3 Schematic diagram of feature fusion.

[0045] Figure 4 This is the network structure diagram of TransU-Net that introduces the gated axial attention mechanism.

[0046] Figure 5 This is the Agglomerative Clustering hierarchical clustering network structure diagram. DETAILED DESCRIPTION

[0047] The technical solutions in the embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0048] See also Figure 1 As shown, the lung CT visualization system provided by the present invention includes a PACS system module, an improved TransU-Net network module introducing a gated axial attention (MGAA) mechanism, a pixel adjustment module, a dimension adjustment module, a cohesive hierarchical clustering module, a feature fusion module, and a visualization module.

[0049] Among them, the PACS system module is used to convert the medical images generated after CT scanning into a standard data format for storage; and convert the standard data format into an operable image format, such as DICOM, NII, etc., and send it to the TransU-Net network module and pixel adjustment module that introduce the improved gated axial attention mechanism.

[0050] The TransU-Net network module with improved gated axial attention mechanism is introduced to extract and process image features and complete basic image segmentation;

[0051] The pixel adjustment module is used to multiply and restore the pixels of the original image; the dimension adjustment module is used to increase and reduce the dimension of the image; the pixel adjustment module and the dimension adjustment module are arranged in the upper and lower layers of the agglomerative hierarchical clustering module, and are used for cluster segmentation preprocessing and restoration processing after segmentation.

[0052] The feature fusion module is used to fuse the basic segmentation image and the clustering segmentation image to make the segmentation of the target area more accurate and reduce the edge error; the visualization module is used to perform pseudo-color processing on the fused segmentation image and convert it into a visual RGB image format.

[0053] The basic segmentation image output by the TransU-Net network module improved by the gated axial attention mechanism and the cluster segmentation image output by the pixel adjustment module are both sent to the feature fusion module; the feature fusion module fuses the two to obtain an optimized fused segmentation image.

[0054] The visualization module is used to perform pseudo-color processing on the fused segmented image and convert it into a visualized RGB image format.

[0055] The following is a detailed introduction to a lung CT visualization method based on the gated axial attention mechanism to improve TransU-Net and supplemented by Agglom erative Clustering to optimize image segmentation. Please refer to Figure 2 shown.

[0056] Step 1: Convert the medical images generated after CT scanning into a standard data format for storage through the PACS system. This includes: first, obtaining the patient's medical images through CT; then, extracting the images from CT through the image collector, converting them into digital information, and standardizing the data format; finally, storing the images and related data through the storage system.

[0057] Step 2: Convert the standard data format into an operable image format and proceed to step 3 and step 4 for different processing.

[0058] Step 3: Import the operable image format into the improved TransU-Net network module that introduces the gated axial attention mechanism to extract and process image features and complete the basic segmentation of the image. Figure 4 , cited in https: / / arxiv.org / abs / 2102.04306.

[0059] Suppose given an image The spatial resolution is H×W and the number of channels is C.

[0060] Step S3.1, Figure 4 The CNN layer in the middle acts as a feature extractor to generate feature maps for the input and embed the image into the CNN layer to extract hidden features.

[0061] Step S3.2, the Transformer layer reshapes the input x processed by the CNN layer into a series of two-dimensional slices Among them, the size of each slice is P×P, is the number of image slices, i.e., the input sequence length; the vectorized slices x are transformed by trainable linear projection p Mapped to the potential D-dimensional embedding space, the slice embedding formula is expressed as Among them, the size of the D-dimensional embedding space should be greater than P×P, is the slice embedding method, is the position embedding method; the improved Transformer encoder consists of gated axial attention (MGAA) and multi-layer perceptron (MLP). After the slice is embedded, it is encoded by the Transformer encoder. For the Lth layer of the encoder, its input is recorded as Z 1-1 , the output is Z1, then the calculation process is:

[0062] z' l =MGAA(NORM(z l-1 ))+z l-1 , l=1,…,L

[0063] z l =MLP(NORM(z' l ))+z' l , l=1,…,L

[0064] Among them, MGAA(·) is the gated axial attention mechanism layer, MLP(·) is the multi-layer perceptron layer, NORM(·) represents the layer normalization operator, and finally Z is obtained L Represents a hidden feature sequence.

[0065] Compared with the improved self-attention mechanism (MSA), the gated axial attention mechanism controls the information flow by adjusting the embedding positions of the query (Quary, Q), key (Key, K), and value (Value, V), limiting irrelevant position information, and reducing the complexity of self-attention calculation. This allows the model to automatically adjust the gating parameters according to the data set during training. The calculation process is:

[0066] S323.1: First, the input quantity X is obtained through three parameter matrices (query matrix Q, key matrix K, value matrix V) to obtain the corresponding matrix W Q , W K , W V , Q=W Q X,K=W K X, V = W V X, where X is the image slice matrix after the input x is mapped to the D-dimensional embedding space. The matrix has N elements, which is the number of image slices.

[0067] S323.2: W Q Perform matrix multiplication with WK to calculate the similarity;

[0068] S323.3:W Q With positional encoding Q Perform matrix multiplication;

[0069] S323.4: Combine the result obtained in S323.3 with the gate control unit GQ Matrix multiplication;

[0070] S323.5:W K With positional encoding K Perform matrix multiplication;

[0071] S323.6: Combine the result obtained in S323.5 with the gate control unit G K Matrix multiplication;

[0072] S323.7: Add the results obtained in S323.2, S323.4, and S323.6;

[0073] S323.8: Send the result obtained in S323.7 to the Softmax activation function for processing;

[0074] S323.9: Combine the result obtained in S323.8 with the position code r v Perform matrix multiplication;

[0075] S323.10: Perform matrix multiplication of the result obtained in S323.9 and the gate control unit Gv1;

[0076] S323.11: Add the result of S323.8 to the matrix W v Perform matrix multiplication;

[0077] S323.12: Combine the result obtained in S323.11 with the gate control unit G V2 Perform matrix multiplication;

[0078] S323.13: Add the results obtained in S323.10 and S323.12;

[0079] S323.14: Use the result obtained in S323.13 as the final output Y of the gated axial attention mechanism layer.

[0080] Among them, the position code is the relative position code of the embedded position information Q, K, and V. The gating unit G is used to change the weight of the position code. If the model cannot learn the accurate position code from a smaller medical data set, the gating unit G will reduce the weight of these position codes.

[0081] Step S3.3, hidden feature sequence After feature reshaping, we get shape.

[0082] Step S3.4, the cascade upsampler (CUP) decodes the hidden features to output the final segmentation mask, by cascading multiple upsampling blocks to achieve to full H×W resolution.

[0083] Among them, each block consists of two upsampling operators, a 3×3 convolution layer, and a ReLU function activation layer in sequence.

[0084] Step 4 includes the following operations:

[0085] Multiply the pixels of the original image through a pixel adjustment module;

[0086] The image is reconstructed in three dimensions through a dimension adjustment module to generate a three-dimensional grayscale image;

[0087] Then, the target area of ​​the three-dimensional grayscale image is clustered and segmented through the agglomerative hierarchical clustering module;

[0088] The image is restored in two dimensions through the dimension adjustment module;

[0089] The pixel adjustment module is used to restore the original image pixels.

[0090] See also Figure 5 As shown, the above clustering segmentation of the target area of ​​the three-dimensional grayscale image through the agglomerative hierarchical clustering module includes:

[0091] Step 4.1: Treat all data points as separate clusters, i.e., initial clusters;

[0092] Step 4.2: Calculate the distance between all adjacent data points △m, △n, △r represent the distances between two points in three directions in the three-dimensional coordinate system, and the two clusters with the smallest distance are selected and merged;

[0093] Step 4.3: Update the clustering tree and recalculate the distances between the new clusters;

[0094] Step 4.4: Repeat steps 4.2 and 4.3 until all data points reach the preset number of clusters.

[0095] Step 5: Please refer to Figure 3 As shown, the basic segmentation image and the clustering segmentation image are fused through the feature fusion module, so that the segmentation of the target area is more accurate and the edge error is reduced.

[0096] Step 6: Perform pseudo-color processing on the fused segmented image and convert it into a visual RGB image format through the visualization module, including:

[0097] Step S6.1, using grayscale-color conversion method through pseudo color processing, the function conversion relationship is as follows:

[0098]

[0099] Where a and b are grayscale value indices, R, G, and B are the three primary color components, and f(a, b) represents the grayscale value of the grayscale image at a specific point. R (a,b),f G (a,b),f B (a, b) represent the color values ​​of R, G, and B channels respectively, and F(R, G, B) represents the color value of the RGB image after the specific grayscale image is transformed.

[0100] Step S6.2, converting the grayscale image into an indexed image, and then mapping the grayscale value of each pixel to a color in the RGB color space through color transformation, thereby achieving pseudo color processing.

[0101] Step 7: Use the PACS system to transmit the optimized digital images and related data to the imaging workstation through the existing imaging transmission system for medical staff to view, diagnose, and analyze.

[0102] The embodiments described above are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

Claims

1. A lung CT visualization method, characterized in that: The following steps are involved: Step 1: The PACS system module converts the medical images generated after CT scanning into a standard data format for storage; Step 2: Convert the standard data format into an operable image format and proceed to step 3 and step 4 for different processing; Step 3: Import the operable image format into the TransU-Net network module that introduces the gated axial attention mechanism to extract and process image features and complete basic segmentation; Step 4: Import the operable image format into the pixel adjustment module for pixel multiplication; The grayscale image after pixel multiplication is imported into the dimension adjustment module for three-dimensional reconstruction to generate a three-dimensional grayscale image; then the target area of ​​the three-dimensional grayscale image is clustered and segmented through the agglomerative hierarchical clustering module; The obtained three-dimensional cluster segmentation image is imported into the dimension adjustment module for two-dimensional restoration; and the dimensionally restored cluster segmentation image is then imported into the pixel adjustment module for pixel restoration; Step 5: Import the basic segmentation image obtained in step 3 and the cluster segmentation image obtained in step 4 into the feature fusion module for image fusion. Take the segmentation line in the cluster segmentation image as a reference, optimize the boundary of the basic segmentation image, and obtain the optimized fused segmentation image. Step 6: Import the fused segmented image into the visualization module for pseudo-color processing and convert it into a visual RGB image format; Step 7: Output the visualized RGB image through the imaging workstation; In step 3, basic segmentation is performed on the imported operable image format, including: Step 3.1: Embed the image into a convolutional neural network layer to extract hidden features; Step 3.2: Put the extracted hidden features into the transformer layer encoding with the gated axial attention mechanism to obtain the processed hidden feature sequence; Step 3.3: Reshape the features of the sequence; Step 3.4: Restore the image through the upsampling layer; In step 3.2, the extracted hidden features are put into the transformer layer that introduces the gated axial attention mechanism for encoding, including: Step 3.2.1: Assume that the spatial resolution of image x is H×W and the number of channels is C; reshape the input x processed by the convolutional neural network layer into a series of two-dimensional slices Among them, x p represents slices, P represents the side length of each slice, and the size of each slice is P×P. is the number of image slices, i.e. the input sequence length; Step 3.2.2: Vectorize the slice x by trainable linear projection p Mapped to the potential D-dimensional embedding space, the slice embedding formula is expressed as Among them, the size of the D-dimensional embedding space should be greater than P×P, is the slice embedding method, It is the position embedding method; Step 3.2.3: After the slice is embedded, it is encoded by the Transformer encoder. For the Lth layer of the encoder, its input is recorded as Z l-1 , the output is Z l , the calculation process is: from' l =MGAA(NORM(z l-1 ))+z l-1 ,l=1,…,L With l =MLP(NORM(z' l ))+z' l ,l=1,…,L Where MGAA is the gated axial attention mechanism layer, MLP is the multi-layer perceptron layer, and NORM represents the layer normalization operator, so Z L Represents a hidden feature sequence.

2. The lung CT visualization method according to claim 1, characterized in that: In step 4, cluster segmentation is performed on the target area of ​​the three-dimensional grayscale image, including: Step 4.1: Treat all data points as separate clusters; Step 4.2: Calculate the distance between all adjacent data points and select the two clusters with the smallest distance to merge; Step 4.3: Update the clustering tree and recalculate the distances between the new clusters; Step 4.4: Repeat steps 4.2 and 4.3 until all data points reach the preset number of clusters.

3. The lung CT visualization method according to claim 1, characterized in that: In step 6, pseudo color processing is performed on the fused segmented image by converting the grayscale image into an index image through a grayscale to color transformation method, and then mapping the grayscale value of each pixel to a color in the RGB color space through color transformation, thereby achieving pseudo color processing.

4. The lung CT visualization method according to claim 1, characterized in that: In step 3.2.3, the calculation process of the gated axial attention mechanism layer includes: S323.1: First, the input quantity X passes through three parameter matrices, namely the query matrix Q, the key matrix K, and the value matrix V, and obtains the corresponding matrix W respectively. Q , W K , W V , Q=W Q X, K = W K X, V = W V X, where X is the image slice matrix after the input x is mapped to the D-dimensional embedding space; S323.2: W Q and W K Perform matrix multiplication to calculate similarity; S323.3:W Q With positional encoding Q Perform matrix multiplication; S323.4: Combine the result obtained in S323.3 with the gate control unit G Q Matrix multiplication; S323.5:W K With positional encoding K Perform matrix multiplication; S323.6: Combine the result obtained in S323.5 with the gate control unit G K Matrix multiplication; S323.7: Add the results obtained in S323.2, S323.4, and S323.6; S323.8: Send the result obtained in S323.7 to the Softmax activation function for processing; S323.9: Combine the result obtained in S323.8 with the position code r v Perform matrix multiplication; S323.10: Combine the result obtained in S323.9 with the gate control unit G V1 Perform matrix multiplication; S323.11: Add the result of S323.8 to the matrix W v Perform matrix multiplication; S323.12: Combine the result obtained in S323.11 with the gate control unit G v2 Perform matrix multiplication; S323.13: Add the results obtained in S323.10 and S323.12; S323.14: Use the result obtained in S323.13 as the final output Y of the gated axial attention mechanism layer; Among them, the position code r Q , r K , r V It is the relative position encoding of the embedded position information Q, K, V, and the gating unit is used to adjust the weight of the position encoding.

5. A lung CT visualization system for implementing a lung CT visualization method as claimed in claim 1, comprising a PACS system module, characterized in that: It also includes the TransU-Net network module that introduces the improved gated axial attention mechanism, the pixel adjustment module, the dimension adjustment module, the agglomerative hierarchical clustering module, the feature fusion module, and the visualization module; The PACS system module is used to convert the medical images generated after CT scanning into a standard data format for storage, and convert the standard data format into an operable image format, and send them to the TransU-Net network module and the pixel adjustment module respectively, which are improved by introducing a gated axial attention mechanism; The TransU-Net network module improved by introducing the gated axial attention mechanism is used to extract and process image features, complete basic image segmentation, and send it to the feature fusion module; The pixel adjustment module is used to multiply and restore the pixels of the original image; the dimension adjustment module is used to increase and reduce the dimension of the image; the agglomerative hierarchical clustering module is used to cluster and segment the target area of ​​the three-dimensional grayscale image; Import the operable image format into the pixel adjustment module for pixel multiplication; The grayscale image after pixel multiplication is imported into the dimension adjustment module for three-dimensional reconstruction to generate a three-dimensional grayscale image; then the target area of ​​the three-dimensional grayscale image is clustered and segmented through the agglomerative hierarchical clustering module; The obtained three-dimensional cluster segmentation image is imported into the dimension adjustment module for two-dimensional restoration; the cluster segmentation image after dimension restoration is then imported into the pixel adjustment module for pixel restoration and sent to the feature fusion module; The feature fusion module is used to fuse the basic segmentation image and the cluster segmentation image output by the pixel adjustment module, and optimize the boundary of the basic segmentation image based on the segmentation line in the cluster segmentation image to obtain an optimized fused segmentation image; The visualization module is used to perform pseudo-color processing on the fused segmented image and convert it into a visualized RGB image format.

6. The lung CT visualization system according to claim 5, characterized in that: The TransU-Net network module improved by introducing the gated axial attention mechanism performs basic segmentation on the imported operable image format by: Embed the image into a convolutional neural network layer to extract hidden features; The extracted hidden features are put into the transformer layer encoding with the gated axial attention mechanism introduced to obtain the processed hidden feature sequence; Reshape the sequence features; Restore the image through the upsampling layer.

Images