A medical 2D image segmentation method

Abstract

The present invention discloses a medical 2D image segmentation method, which includes the following steps: Step 1, extract the hierarchical slices of each of the to-be-processed images, and randomly sample and construct a data set according to the ratio of training set: validation set being 3:1; Step 2, sequentially perform feature extraction through four pyramid spatial attention modules, and simultaneously save the extraction result of the second time as low-level features; Step 3, obtain a primary fuzzy segmentation feature after a total of four upsamplings; Step 4, input the primary fuzzy segmentation feature in Step 3 and the low-level features in Step 2 into a fuzzy region refinement module to obtain a final optimized high-precision segmentation image. Compared with the prior art, it is more efficient, the segmented region is more complete, and the segmentation accuracy is higher, providing a more accurate assessment for actual medical image analysis.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image processing, and in particular to a medical 2D image segmentation method. Background Art

[0002] Medical image segmentation is routinely used for quantifying the size and shape of target organs, disease quantification, and treatment planning and intervention. Traditional medical image segmentation methods primarily include thresholding, watershed methods, and atlas methods. In recent years, deep learning methods have emerged as a highly competitive alternative and demonstrated significant superiority. They have proven capable of handling most current medical image segmentation datasets, offering advantages such as small model size and high accuracy. Consequently, deep learning methods have achieved tremendous success in medical image segmentation.

[0003] In the existing technology, most medical image segmentation methods ignore feature fusion in multiple dimensions, resulting in insufficient medical feature information. In addition, due to the limitations of convolutional neural networks themselves, there is a problem of blurred edge segmentation in most medical image segmentation tasks, which is reflected in the final result that the segmentation accuracy is not high enough and does not meet the level of actual medical applications. Summary of the Invention

[0004] The purpose of the present invention is to provide a medical 2D image segmentation method that improves the efficiency of global background modeling while maintaining a strong grasp of low-level information, ultimately improving the accuracy of medical image segmentation. The method mainly includes: an encoder-decoder structure with a large convolution kernel, which is mainly based on a tower-shaped spatial attention module and a fuzzy region refinement module. The tower-shaped spatial attention module can effectively keep the target in the reasoning process within the label-calibrated area; the fuzzy region refinement module can use the main features generated by the network to generate local reference features, and by judging the similarity between the local features and the main features, it can reclassify some uncertain pixels and eliminate the uncertainty in the segmentation process. The above two modules promote each other to play a role, which greatly optimizes the uncertain pixels in medical image segmentation and improves the accuracy and efficiency of organ segmentation. It includes the following steps:

[0005] Step 1: obtain multiple segmented images to be processed, extract hierarchical slices of each of the images to be processed, further remove samples without labels in the slices, then aggregate the slices and construct a dataset by randomly sampling in a ratio of 3:1 between the training set and the validation set;

[0006] Step 2: Input the dataset into the encoder-decoder network with a large convolution kernel, and extract features through the pyramid spatial attention module four times in sequence, while saving the second extraction result as the low-level feature;

[0007] Step 3: Upsample the features extracted in the fourth time, fuse the features after each upsampling with the features extracted by the pyramid spatial attention module, and perform the next upsampling on the fused features. After a total of four upsamplings, a primary fuzzy segmentation feature is obtained.

[0008] Step 4: Input the primary fuzzy segmentation feature in Step 3 and the low-level features in Step 2 into the fuzzy region refinement module to obtain the finally optimized high-precision segmentation image.

[0009] To achieve the above object, the present invention provides the following technical solution: A medical 2D image segmentation method. Compared with the prior art, the beneficial effects of the present invention are as follows:

[0010] By combining the encoder-decoder structure with large convolution kernels, as well as the pyramid spatial attention module and the fuzzy region refinement module, cross-scale semantic information fusion is obtained. First, input the image to be processed into the encoder structure, and respectively extract features through four pyramid spatial attention modules, and simultaneously generate four basic feature maps of different scales; input the last generated basic feature into the encoder structure, and sequentially fuse the four feature maps generated by the encoder to output a fuzzy segmentation feature map; extract the second basic feature and the fuzzy segmentation feature and input them into the fuzzy region refinement module together to generate the final segmentation image. Compared with the prior art, it is more efficient, the segmented region is more complete, the segmentation accuracy is higher, and it provides a more accurate evaluation for actual medical image analysis. Description of the Drawings

[0011] Figure 1 It is a schematic flow chart of a method for efficiently extracting medical image features provided by an embodiment of the present invention;

[0012] Figure 2 It is a schematic structural diagram of a medical 2D image segmentation method provided by an embodiment of the present invention;

[0013] Figure 3 It is a schematic structural diagram of the pyramid spatial attention module provided by an embodiment of the present invention;

[0014] Figure 4 It is a schematic structural diagram of the fuzzy region refinement module provided by an embodiment of the present invention; Detailed Embodiments

[0015] Next, the technical solutions of the present invention in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0016] Embodiment 1. A medical 2D image segmentation method includes the following steps:

[0017] Step 1: Obtain multiple segmentation images to be processed, extract the hierarchical slices of each image to be processed, further remove the samples without labels in the slices, and then summarize the slices and randomly sample them according to the ratio of training set: validation set of 3:1 to construct a data set;

[0018] Step 2: Input the data set into an encoder-decoder network with a large convolution kernel, and sequentially perform feature extraction through four pyramid spatial attention modules, and simultaneously save the extraction result of the second time as low-level features;

[0019] Step 3: Upsample the features extracted for the fourth time, fuse the features after each upsampling with the features extracted by the pyramid spatial attention module, and perform the next upsampling on the fused features. After a total of four upsamplings, a primary fuzzy segmentation feature is obtained.

[0020] Step 4: Input the primary fuzzy segmentation feature in Step 3 and the low-level features in Step 2 into the fuzzy region refinement module to obtain the finally optimized high-precision segmentation image.

[0021] The pyramid spatial attention module first weights the importance of each channel selectively through a channel attention module to obtain more hierarchical semantic feature information; the channel attention module includes bidirectional squeeze and excitation; in the squeeze stage, global average pooling and global max pooling are performed on the original input features to obtain two intermediate feature weights Gc and Mc

[0022] The attention weight calculation of the C-th channel in the channel attention module is as follows

[0023] w c = σ(W1δ(W0(G c )) + W1δ(W0(M c ))) (1)

[0024] Where W0 and W1 represent two different fully connected layers, δ represents the ReLU function symbol, and σ represents the activation function. The entire channel attention module is as Figure 3as shown in a;

[0025] Secondly, the result w of the above channel attention module c is input into the above-mentioned tower-shaped spatial attention module to obtain four feature maps of different scales. The tower-shaped spatial attention module is as Figure 3 shown in b.

[0026] The tower-shaped spatial attention module consists of four parts: The first part divides each data in the dataset according to channels to obtain feature maps on different channels, and respectively passes through three convolutional kernels of different sizes to obtain feature maps of different scales, and then through feature fusion to obtain the original feature map; The second part uses the channel attention module to obtain feature weights of different scales; The third part is to use Softmax to recalibrate the obtained weights to obtain the recalibrated feature weights; The fourth part multiplies the obtained original feature map and feature weights by the corresponding elements to obtain the feature with refined multi-scale information.

[0027] In the fuzzy region refinement module mentioned in step 4, for each pixel x, the uncertainty U(x) of the pixel is generated according to the confidence, and the uncertainty of the pixel is quantitatively represented; A residual structure Res is designed to generate the reference feature y i (x) on the neighborhood of the pixel x, y(x) = {y i (x)|i ∈ [1, L]}, where L represents the number of segmentation objects; In addition, the pixel x is processed by global average and global max pooling to generate the local reference feature local(x), and finally by judging the closeness of the local reference feature local(x) and the reference feature y i (x), if the local feature is very close to the reference feature y i (x), the pixel x is classified as the target i; The reference feature y i (x) is obtained by weighted average calculation within the small neighborhood N(x),

[0028]

[0029] In formula (2), M i (q) is the original segmentation mask of object i, local(x) is the local reference feature; q is the element within the neighborhood; N(x) is the small neighborhood of pixel x;

[0030] Specifically, by comparing the similarity between local(x) and y i (x), a mask z i (x) is generated for each pixel x;

[0031] z i (x) = s i (x)Res(local(x), yi (x)) (3)

[0032] In formula (3), the residual network module local(x) is used to learn and predict local similarity; s i (x) = max q∈N(x) M i (x), which is used to measure the confidence score of the refinement mechanism. The final segmentation function is:

[0033] F i (x) = M i (x) + U(x)z i (x) (4)

[0034] In formula (4), M i (x) is the original segmentation mask of object i, U(x) represents the uncertainty, and z i (x) represents the mask generated according to the similarity.

[0035] It should be noted that the dataset used in the present invention is from the spleen and liver segmentation datasets in Medical Segmentation Decathlon (MSD).

[0036] Although the embodiments of the present invention have been shown and described, for those of ordinary skill in the art, it can be understood that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A medical 2D image segmentation method, characterized in that, It includes the following steps: Step 1: Obtain multiple segmentation images to be processed, extract the hierarchical slices of each of the images to be processed, further remove the samples without labels in the slices, then summarize the slices, and randomly sample and construct a data set according to the ratio of training set:validation set of 3:1; Step 2: Input the data set into an encoder-decoder network with a large convolution kernel, and sequentially perform feature extraction through four pyramid spatial attention modules, and simultaneously save the extraction result of the second time as low-level features; Step 3: Upsample the features extracted for the fourth time, fuse the features after each upsampling with the features extracted by the pyramid spatial attention module, and perform the next upsampling on the fused features. After a total of four upsamplings, a primary fuzzy segmentation feature is obtained; Step 4: Input the primary fuzzy segmentation feature in Step 3 and the low-level features in Step 2 into a fuzzy region refinement module to obtain the finally optimized high-precision segmentation image; The pyramid spatial attention module consists of four parts: The first part divides each data in the data set according to channels, obtains feature maps on different channels, respectively obtains feature maps of different scales through three convolution kernels of different sizes, and then obtains the original feature map through feature fusion; The second part uses a channel attention module to obtain feature weights of different scales; The third part is to use Softmax to recalibrate the obtained weights to obtain recalibrated feature weights; The fourth part multiplies the obtained original feature map and feature weights by corresponding elements to obtain features with refined multi-scale information; The fuzzy region refinement module generates the uncertainty U(x) of each pixel x according to the confidence level, and quantitatively represents the uncertainty of the pixel; a residual structure Res is designed to generate the reference feature y i (x), y(x) = {y i (x)|i ∈ [1, L]}, where L represents the number of segmentation objects; in addition, the pixel x is processed by global average and global max pooling to generate the local reference feature local(x), and finally, by judging the proximity between the local reference feature local(x) and the reference feature y i (x), if the local feature is very close to the reference feature y i (x), the pixel is classified as the target i; the reference feature y i (x) is obtained by weighted averaging within the small neighborhood N(p). M in formula (2) i (q) is the original segmentation mask of object i, local(x) is the local reference feature; q is an element in the neighborhood; N(x) is the small neighborhood of pixel x; Specifically, it is by comparing the similarity between local(x) and y i to generate a mask z for each pixel i (x); z i (x) = s i (x) Res(local(x), y i (x)) (3) In formula (3), the residual network module local(x) is used to learn and predict local similarity; s i (x) = max q∈N(x) M i (x), which is used to measure the confidence score of the refinement mechanism. The final segmentation function is: F i F(x) = M i F(x)+U(x)z i F(x) (4) M in formula (4) i (x) is the original segmentation mask of object i, U(x) represents the uncertainty, and z i (x) represents the mask generated according to the similarity.

2. The medical 2D image segmentation method according to claim 1, wherein, In Step 2, the pyramid spatial attention module first weights the importance of each channel selectively through a channel attention module to obtain more hierarchical semantic feature information; The channel attention module includes bidirectional squeezing and excitation; In the squeezing stage, global average pooling and global max pooling are performed on the original input features to obtain two intermediate feature weights Gc and Mc respectively The attention weight calculation of the C-th channel in the channel attention module is as follows w c = σ(W1δ(W0(G c )) + W1δ(W0(M c ))) (1) Among them W0 and W1 represent two different fully connected layers, δ represents the ReLU function symbol, and σ represents the activation function; Secondly, the result w of the above channel attention module is c input into the tower-shaped spatial attention module described above to obtain four feature maps of different scales.

Images