A medical image segmentation method based on MECAU-Net network

Abstract

The present invention discloses a medical image segmentation method based on a MECAU‑Net network, comprising: using a 2×2 even convolution module at the encoding end of the MECAU‑Net network to extract features from an imported medical image; constructing a 4×4 even convolution path parallel to the 2×2 even convolution module, using a 4×4 even convolution kernel to extract image information, and directly passing the extracted information to the backbone of the encoding network to fuse feature information in different receptive fields; symmetrically filling the feature map to be segmented at each layer, and then passing the acquired information to the corresponding main network by splicing for the next step of pooling, while eliminating the pixel offset caused by the even convolution and expanding the receptive field of the even convolution kernel. The present invention can extract richer information without adding additional overhead, and obtain better segmentation results with as little computational overhead as possible.

Description

Technical Field

[0001] The present invention belongs to the technical field of image segmentation, and in particular relates to a medical image segmentation method and system based on a MECAU-Net network. Background Art

[0002] Deep learning has been successfully applied in many scenarios such as image classification, semantic segmentation, and target detection. Medical image segmentation has attracted widespread attention from researchers due to its unique application scenarios. Its data requires professional annotation, and it is relatively difficult to obtain data sets, which increases the difficulty of medical image segmentation. With the rapid development of deep learning, researchers have designed various networks to improve the performance of medical image segmentation. Ronneberger et al. proposed the U-Net network, which used the U-shaped network structure to segment medical images for the first time and achieved good performance. Its unique U-shaped structure can well transmit the extracted feature information, which is very suitable for medical image segmentation with small data sets.

[0003] In order to further improve the segmentation performance, researchers have proposed some improved networks of U-Net network. These improved networks mainly enhance the feature extraction ability by increasing parameters and computing power to obtain deeper and wider networks. However, the main feature of U-Net network is lightweight network, and most of the existing improved U-Net obtains better segmentation performance by widening and deepening the network. The high memory requirements and high computing power of these networks seriously limit the application of deep learning in the field of medical image segmentation. How to reduce the required computing resources while maintaining satisfactory segmentation performance has become a new focus of researchers.

[0004] The invention with patent number CN113920108A discloses a training method for training a U-Net model for processing cell images. The training method trains the U-Net model based on a pre-acquired labeled cell image data set. The trained U-Net model can accurately segment cell images, and the contour edges after segmentation are clear, which effectively improves the accuracy of segmenting and / or counting cell images, especially mixed cell images, and effectively solves the problems existing in the prior art such as high image annotation cost, difficult cell image segmentation, and complex cell counting operations. However, the invention does not involve the technical problem of reducing computing resources.

[0005] The invention with patent number CN113902630A discloses a generative adversarial network image restoration method based on multi-scale texture feature branches, constructing a dataset of images to be restored and real images containing missing areas; constructing a generative adversarial network model based on multi-scale texture feature branches; training the generative adversarial network model based on multi-scale texture feature branches to obtain the optimal generative adversarial network model; inputting the image to be restored containing missing areas into the optimal generator network model, and outputting the final restoration result image. The invention can achieve good restoration effects for irregular random missing images and large-area missing images, and the restored images have good structural and texture features. Similarly, the invention does not involve the technical problem of reducing computing resources. Summary of the invention

[0006] Technical problem to be solved: How to obtain better segmentation results with as little computational overhead as possible.

[0007] Technical solution:

[0008] A medical image segmentation method based on a MECAU-Net network, the medical image segmentation method comprising the following steps:

[0009] A 2×2 even convolution module is used at the encoding end of the MECAU-Net network to extract features from the imported medical images. At the same time, based on the multi-scale fusion technology, a 4×4 even convolution path parallel to the 2×2 even convolution module is constructed at the encoding end, and a 4×4 even convolution kernel is used to extract image information. The extracted information is directly passed to the backbone of the encoding network to fuse feature information in different receptive fields.

[0010] Among them, the feature map to be segmented is symmetrically filled in each layer, and then the acquired information is passed to the corresponding main network through splicing for the next pooling, which eliminates the pixel offset caused by the even convolution and expands the receptive field of the even convolution kernel.

[0011] Furthermore, the MECAU-Net network is provided with 5 layers of convolution from top to bottom.

[0012] Furthermore, the medical image segmentation method further comprises:

[0013] A CBAM attention module is embedded after the 2×2 even-numbered convolution module at the encoding end of the MECAU-Net network. The attention map is inferred along two different dimensions, channel and space, and the final result is obtained by fusing the attention map and the feature map.

[0014] Based on the above method, the present invention also mentions a medical image segmentation system based on a MECAU-Net network, wherein the medical image segmentation system comprises an input module, an encoding end, a multi-scale fusion module, a backbone part and an output module;

[0015] The input module is used to import medical images;

[0016] The encoding end includes 5 layers of 2×2 even-numbered convolution modules connected in sequence, which are used to extract features from the imported medical images;

[0017] The multi-scale fusion module includes 5 layers of 4×4 even convolution paths parallel to the 2×2 even convolution module; each layer of 4×4 even convolution kernels extracts image information, and after fusing the extracted information with the feature information extracted by the 2×2 even convolution module of the corresponding layer, it is directly passed to the backbone of the encoding network to fuse the feature information in different receptive fields; wherein, at each layer, the feature map to be segmented is symmetrically filled, and then the acquired information is passed to the corresponding main network by splicing for the next step of pooling, thereby eliminating the pixel offset caused by the even convolution and expanding the receptive field of the even convolution kernel;

[0018] The output module is used to output a multi-scale fused feature map.

[0019] Furthermore, the medical image segmentation system also includes a 5-layer CBAM attention module; the 5-layer CBAM attention module is embedded one-to-one after the 5-layer 2×2 even-numbered convolution module, and the attention map is inferred along two different dimensions of channel and space in sequence, and the final result is obtained by fusing the attention map and the feature map.

[0020] The present invention also relates to an electronic device, comprising:

[0021] one or more processors;

[0022] a storage device for storing one or more programs,

[0023] When the one or more programs are executed by the one or more processors, the one or more processors implement the medical image segmentation method based on the MECAU-Net network as described above.

[0024] The present invention also relates to a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the medical image segmentation method based on the MECAU-Net network as described above.

[0025] Beneficial effects:

[0026] On the basis of continuing to use the U-shaped structure of the U-Net network, the present invention makes certain improvements to the encoding end, including: first, 2×2 even convolution is used at the encoding end to replace the 3×3 convolution of the original U-Net network for feature extraction; secondly, drawing on the multi-scale idea, 4×4 even convolution is used at the encoding end to directly pass the obtained information to the trunk part to obtain more comprehensive image information and reduce additional computational overhead; furthermore, symmetric padding is used at the encoding end to solve the offset problem generated during the information extraction process of even convolution; finally, the encoding end of the U-Net network is further optimized, that is, an attention module in the convolution block is added after the 2×2 even convolution module, combining spatial and channel attention, while extracting richer information, almost no additional overhead is added. The effect of the present invention is that better segmentation results can be obtained with as little computational overhead as possible. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 This is a flow chart of a medical image segmentation method based on a MECAU-Net network according to an embodiment of the present invention.

[0028] Figure 2 Schematic diagram of the structure of a medical image segmentation system based on a MECAU-Net network according to an embodiment of the present invention.

[0029] Figure 3 The figure is a diagram showing the visual comparison of related networks on the DRIVE, ISBI2012, and CHAOS datasets.

[0030] Figure 4 This is a schematic diagram of the data comparison results of related networks on the DRIVE, ISBI2012, and CHAOS datasets. DETAILED DESCRIPTION

[0031] The following examples will enable those skilled in the art to more fully understand the present invention, but are not intended to limit the present invention in any way.

[0032] Figure 2 Schematic diagram of the structure of a medical image segmentation system based on a MECAU-Net network according to an embodiment of the present invention.

[0033] like Figure 1 and Figure 2As shown, an embodiment of the present invention proposes a new multiscale even convolution attention U-Net (Multiscale Even Convolution Attention U-Net, MECAU-Net) network, which is proposed on the basis of the U-shaped structure of the U-Net network. The network not only aims to meet the accuracy requirements of medical image segmentation, but also pays special attention to how to build a high-precision lightweight segmentation network with the lowest possible computational overhead. The MECAU-Net network has five layers from top to bottom. The main body of the network continues to use the U-shaped structure of the U-Net network, and mainly makes certain improvements to the encoding end. The decoding end still uses the original decoding structure of the U-Net structure, thereby proposing a new medical image segmentation method based on the MECAU-Net network. Among them, the improvements on the decoding end mainly include: first, 2×2 even convolution is used to replace the 3×3 convolution of the original U-Net network for feature extraction at the encoding end; secondly, drawing on the multi-scale idea, 4×4 even convolution is used to directly pass the obtained information to the trunk part to obtain more comprehensive image information and reduce additional computing overhead; furthermore, symmetric padding is used to solve the offset problem generated in the process of even convolution to extract information; finally, the encoding part of the U-Net network is further optimized, that is, the attention module in the convolution block is added after the 2×2 even convolution module, combining spatial and channel attention, while extracting richer information, almost no additional overhead is added. The effect of the present invention is that better segmentation results can be obtained with as little computing overhead as possible.

[0034] Among them, the introduction of the multi-scale even convolution mentioned above is as follows:

[0035] Using a smaller convolution kernel is very helpful in reducing the amount of calculation, but a 1×1 convolution kernel can hardly extract the pixel location information, so a 2×2 even-numbered convolution kernel is used to extract image information, reducing the number of parameters while maintaining the receptive field. 2×2 convolution is used instead of 3×3 convolution to extract information from the image, and the convolution kernel becomes smaller, thereby reducing the large amount of computational overhead required for the convolution part, thereby reducing the number of parameters in the convolution part. In addition, in order to avoid the pixel offset problem, the proposed network uses a symmetrical padding method to solve the offset stacking problem in order to solve the pixel offset caused by the even-numbered convolution, while expanding the receptive field while eliminating the pixel offset caused by the even-numbered convolution; multi-scale feature fusion can usually bring better performance to the network. The common multi-scale feature fusion is to introduce 5×5 convolution into the 3×3 convolution network to obtain image information within a larger receptive field, but multi-scale feature fusion will introduce too many parameters and complexity into the network. Considering the problem of parameter growth, a new layer of 4×4 even convolution coding network is added to the encoding part outside the main part of the segmentation network. The additional branches are used to extract image information separately, and the obtained information is directly passed to the main part of the encoding network. Feature information is extracted in a larger receptive field, and feature information in different receptive fields is integrated to ensure that more comprehensive image information is obtained while reducing additional computational overhead. A larger scale 4×4 even convolution kernel is used to extract image information. At the same time, in order to avoid the pixel offset problem caused by the even convolution kernel, the feature map to be segmented is symmetrically filled at each layer, and then the obtained information is passed to the corresponding main network by splicing for the next pooling. The encoding end of the MECAU-Net network uses symmetrical padding to solve the offset problem generated in the process of extracting information from the even convolution. That is, in order to solve the pixel offset problem caused by the even convolution kernel, the idea of ​​symmetrical padding of the even convolution is borrowed to symmetrically pad the image features before the convolution, which eliminates the offset and expands the receptive field of the even convolution kernel, thereby achieving the purpose of reducing parameters while achieving equivalent or even better performance.

[0036] Among them, the introduction of the above-mentioned CBAM attention module is as follows:

[0037] In order to more effectively segment the target area in the medical image, a CBAM attention module is embedded after the 2×2 even convolution module. The CBAM attention module will infer the attention map along the two different dimensions of channel and space, and then fuse the attention map and feature map to obtain the final result. By adding the CBAM attention module to focus on useful feature information, it can extract richer information while adding almost no additional overhead.

[0038] Finally, in order to evaluate the medical image segmentation method proposed in this embodiment, DRIVE, ISBI2012, and CHAOS datasets are selected to demonstrate its experimental results. For these three datasets, two images are selected from the segmentation results of each network and displayed as experimental results for comparison. Figure 3 As shown in the figure, the comparison between MECAU-Net and other networks shows that the improvement of the encoding part is very effective from a visual perspective. The segmentation performance has been significantly improved, and the extracted detail information is more complete. Figure 4 As shown in the figure, the specific measurement indicators for comparing MECAU-Net with other networks, in addition to the common performance indicators, in order to prove the unique advantage of MECAU-Net in terms of network parameter quantity, the parameter quantity is also added as a measurement indicator. From a quantitative perspective, it is proved that the MECAU-Net medical image segmentation method based on the improved U-Net medical image segmentation method, which adds multi-scale even convolution and CBAM attention modules, not only has obvious advantages in the accuracy of medical image segmentation, but also has more outstanding performance in terms of parameter quantity.

[0039] In view of the shortcomings of the existing U-Net technology, the present invention proposes a new multiscale even convolution attention U-Net (Multiscale Even Convolution Attention U-Net, MECAU-Net) network based on the U-shaped structure of the U-Net network. The network not only aims to meet the accuracy requirements of medical image segmentation, but also pays special attention to how to build a high-precision lightweight segmentation network with the lowest possible computational overhead. The MECAU-Net network has five layers from top to bottom. The main body of the network continues to use the U-shaped structure of the U-Net network, and mainly makes certain improvements to the encoding end. The decoding end still uses the original decoding structure of the U-Net structure, thereby proposing a new medical image segmentation method based on the MECAU-Net network. Among them, the improvements on the decoding end mainly include: first, 2×2 even convolution is used to replace the 3×3 convolution of the original U-Net network for feature extraction at the encoding end; secondly, drawing on the multi-scale idea, 4×4 even convolution is used to directly pass the obtained information to the trunk part to obtain more comprehensive image information and reduce additional computing overhead; furthermore, symmetric padding is used to solve the offset problem generated in the process of even convolution to extract information; finally, the encoding part of the U-Net network is further optimized, that is, the attention module in the convolution block is added after the 2×2 even convolution module, combining spatial and channel attention, while extracting richer information, almost no additional overhead is added. The effect of the present invention is that better segmentation results can be obtained with as little computing overhead as possible.

Claims

1. A medical image segmentation method based on MECAU-Net network, characterized in that: The medical image segmentation method comprises the following steps: A 2×2 even convolution module is used at the encoding end of the MECAU-Net network to extract features from the imported medical images. At the same time, based on the multi-scale fusion technology, a 4×4 even convolution path parallel to the 2×2 even convolution module is constructed at the encoding end, and a 4×4 even convolution kernel is used to extract image information. The extracted information is directly passed to the backbone of the encoding network to fuse feature information in different receptive fields. Among them, the feature map to be segmented is symmetrically filled in each layer, and then the acquired information is passed to the corresponding main network through splicing for the next step of pooling, which eliminates the pixel offset caused by the even convolution and expands the receptive field of the even convolution kernel; The MECAU-Net network is configured with 5 layers of convolution from top to bottom; The medical image segmentation method further comprises: A CBAM attention module is embedded after the 2×2 even-numbered convolution module at the encoding end of the MECAU-Net network. The attention map is inferred along two different dimensions, channel and space, and the final result is obtained by fusing the attention map and the feature map.

2. A medical image segmentation system based on a MECAU-Net network based on the method of claim 1, characterized in that: The medical image segmentation system comprises an input module, an encoding end, a multi-scale fusion module, a backbone part and an output module; The input module is used to import medical images; The encoding end includes 5 layers of 2×2 even-numbered convolution modules connected in sequence, which are used to extract features from the imported medical images; The multi-scale fusion module includes 5 layers of 4×4 even convolution paths parallel to the 2×2 even convolution module; each layer of 4×4 even convolution kernels extracts image information, and after fusing the extracted information with the feature information extracted by the 2×2 even convolution module of the corresponding layer, it is directly passed to the backbone of the encoding network to fuse the feature information in different receptive fields; wherein, at each layer, the feature map to be segmented is symmetrically filled, and then the acquired information is passed to the corresponding main network by splicing for the next step of pooling, thereby eliminating the pixel offset caused by the even convolution and expanding the receptive field of the even convolution kernel; The output module is used to output a multi-scale fused feature map.

3. The medical image segmentation system based on MECAU-Net network according to claim 2, characterized in that: The medical image segmentation system also includes a 5-layer CBAM attention module; the 5-layer CBAM attention module is embedded one-to-one after the 5-layer 2×2 even-numbered convolution module, and the attention map is inferred along two different dimensions, channel and space, and the final result is obtained by fusing the attention map and the feature map.

4. An electronic device, characterized in that: include: one or more processors; a storage device for storing one or more programs, When the one or more programs are executed by the one or more processors, the one or more processors implement the medical image segmentation method based on the MECAU-Net network as described in claim 1.

5. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the medical image segmentation method based on the MECAU-Net network as described in claim 1 is implemented.

Images