A chest X-ray multi-object segmentation method, electronic equipment and storage medium

Abstract

The application relates to a chest X-ray multi-target segmentation method, an electronic device and a storage medium, the method comprising the following steps: acquiring an annotated chest X-ray image; inputting the annotated chest X-ray image into a trained single-encoder-double-decoder multi-target segmentation network to obtain different target segmentation results; wherein the training steps of the multi-target segmentation network comprise the following steps: acquiring an annotated chest X-ray image dataset; inputting the image dataset into the single-encoder-double-decoder multi-target segmentation network for training to obtain the trained single-encoder-double-decoder multi-target segmentation network. The collaborative attention jump connection module in the network performs one screening on rib bone information and important information of lung fields, hearts and other targets, thereby effectively improving the segmentation performance of the rib bones; and the attention guide multi-scale feature selection module in the network further improves the segmentation performance of clavicles, posterior rib bones and lung fields.

Description

Technical Field

[0001] This application relates to the field of target segmentation technology for X-ray images, and in particular to a method for multi-target segmentation of chest X-ray images, electronic devices, and storage media. Background Technology

[0002] Chest X-rays are images that use X-ray imaging technology to provide radiologists with information such as the size, shape, and location of the heart, lungs, and bones. They play a crucial role in the diagnosis of lung diseases such as pulmonary nodules, tuberculosis, and COVID-19. With the continuous development of medical imaging technology, the output of image data has increased dramatically. More than 35 million chest X-rays are taken annually for the early diagnosis and treatment of lung diseases, placing a heavy burden on radiologists in terms of reading and analysis. Automated segmentation of medical images plays a vital role in scientific research and healthcare. Automated segmentation technology not only improves clinical efficiency but also reduces the burden on radiologists. In the field of chest X-ray analysis, segmenting targets such as the lung fields, heart, clavicle, and ribs helps doctors quickly and accurately analyze and diagnose various lung diseases. It also provides input for the automated measurement of medical indicators in chest computer-aided diagnostic systems, improving the performance of these systems.

[0003] With the rapid development of deep learning technology, deep learning-based algorithms have become one of the most popular research topics in image segmentation. Single-target or multi-target segmentation methods for lung fields, heart, ribs, and clavicles based on deep learning have achieved rapid progress. Analysis of multiple studies reveals that existing research achieves high segmentation accuracy for large, simple targets like lung fields with high contrast to the background, but performs poorly on smaller targets like the clavicle. In the field of medical image analysis, data annotation is costly, and publicly available fully annotated datasets are rare. Even medical images using the same imaging techniques are often labeled with different regions of interest depending on the research objective.

[0004] In recent years, the application of multiple partially labeled datasets for learning in tasks such as medical image classification, segmentation, and lesion detection has received increasing attention. These researchers have formulated the partial labeling problem as a multi-class segmentation task, treating unlabeled organs as background. This is applicable to organ segmentation tasks where pixels are classified as single-class. However, for chest X-rays, due to the projection properties of X-rays and the overlapping of anatomical structures, a single pixel can simultaneously be the foreground of multiple targets, thus it cannot be simply classified as foreground or background. Furthermore, the lung field and rib areas have a high degree of overlap, which negatively impacts pixel classification if a single decoder is used.

[0005] This application addresses the problem that multi-object segmentation tasks are difficult to complete under partial annotation conditions. Summary of the Invention

[0006] To address the aforementioned issues, this application provides a method for multi-target segmentation of chest X-ray images, an electronic device, and a storage medium.

[0007] In a first aspect, embodiments of this application provide a method for multi-target segmentation of chest X-ray images, including:

[0008] Obtain labeled chest X-ray images; the labeled areas include at least: lung fields, ribs, clavicle, and heart;

[0009] Annotated chest X-ray images are input into a trained single encoder-dual decoder multi-object segmentation network to obtain different object segmentation results;

[0010] The training steps for the trained single-encoder-dual-decoder multi-object segmentation network include:

[0011] Obtain a dataset of labeled chest X-ray images, which includes at least: RCS-CXR, VinDr-RibCXR, JSRT, and Montgomery;

[0012] The image dataset is input into a single encoder-dual decoder multi-object segmentation network for training to obtain a trained single encoder-dual decoder multi-object segmentation network;

[0013] Among them, RCS-CXR labeled the anterior ribs, posterior ribs, and clavicle; VinDr-RibCXR labeled the anterior and posterior ribs as a whole; JSRT labeled the clavicle, lung fields, and heart; and Montgomery only labeled the lung fields.

[0014] Furthermore, in the aforementioned method for multi-object segmentation of chest X-ray images, the labeled chest X-ray image is input into a trained single-encoder-dual-decoder multi-object segmentation network to obtain different object segmentation results, including:

[0015] The labeled chest X-ray image is input into a single encoder to extract the image features;

[0016] The extracted image features are input into a dual decoder to obtain different target segmentation results.

[0017] Furthermore, in the above-mentioned method for multi-target segmentation of chest X-ray images, the single encoder includes at least: a 3*3 convolution, a densely connected block, and a collaborative attention skip connection module;

[0018] Among them, 3*3 convolutions and densely connected blocks are used to extract features from multi-object labeled datasets;

[0019] The collaborative attention skip connection module is used to extract spatial details related to the region to be segmented from the encoder, reduce background noise, and mitigate the impact of the semantic gap in the encoder-decoder feature mapping on segmentation.

[0020] Furthermore, in the above-mentioned method for multi-target segmentation of chest X-ray images, both the first decoder and the second decoder in the dual decoder include: an attention-guided multi-scale feature selection module and a single-layer pre-activated residual block;

[0021] The attention-guided multi-scale feature selection module contains three branches. Each branch consists of a compression-activation module and 3*3 convolutions with different dilation rates. The scaling factor of the compression-activation module in each branch is 16. The attention-guided multi-scale feature selection module reduces the number of channels in each branch to half the number of input channels.

[0022] Single-layer pre-activated residual blocks are used to build decoders related to the segmentation task.

[0023] Furthermore, in the above-mentioned method for multi-target segmentation of chest X-ray images, the first decoder in the decoder is used to process the segmentation of irregular targets, including: ribs and clavicles; the second decoder in the decoder is used to process the segmentation of large-area targets and near-circular targets, including at least: lung fields and heart.

[0024] Furthermore, in the aforementioned method for multi-target segmentation of chest X-ray images, the overall training of the single encoder-dual decoder multi-target segmentation network adopts an alternating training strategy.

[0025] Furthermore, the aforementioned method for multi-target segmentation of chest X-ray images also includes: training the encoder and dual decoder parameters in a unified manner to obtain the optimal multi-target segmentation network parameters of a single encoder-dual decoder.

[0026] Furthermore, the aforementioned method for multi-target segmentation of chest X-ray images also includes:

[0027] During training, the Dice loss function is used in all multi-object segmentation networks with a single encoder and dual decoders.

[0028] When both the RCS-CXR and JSRT datasets are used for training, the weights of the Dice loss function for each target are set to 1, and the average of the branches of the first decoder and the second decoder is calculated.

[0029] When RCS-CXR, JSRT, VinDr-RibCXR, and the Montgomery dataset are all used for training, the clavicle weight of the first decoder is set to 1.5, the heart weight of the second decoder is set to 2, and the weights of the Dice loss function for the other targets are set to 1.

[0030] In a second aspect, embodiments of the present invention also provide an electronic device, including: a processor and a memory;

[0031] The processor executes any of the above-described methods for multi-target segmentation of chest X-ray images by calling programs or instructions stored in the memory.

[0032] Thirdly, embodiments of the present invention also provide a computer-readable storage medium storing a program or instructions that cause a computer to perform any of the above-described methods for multi-target segmentation of chest X-ray images.

[0033] The advantages of this application's embodiments are as follows: The single encoder-dual decoder multi-target segmentation network in this application constructs a shared encoder with densely connected blocks, and uses pre-activated residual blocks with weighted normalization to construct a decoder related to the segmentation task. During the decoding stage, it utilizes an attention-guided multi-scale feature selection module to capture spatial detail information and rich contextual information related to the segmentation task, effectively improving the segmentation accuracy. During the training stage, an alternating training strategy is adopted to enable four partially labeled datasets to participate in the training process of the same network, achieving the goal of segmenting multiple targets in chest X-rays using the same network structure, greatly promoting data utilization efficiency. Attached Figure Description

[0034] To more clearly illustrate the technical solutions in the embodiments of this application or the conventional technology, the drawings used in the description of the embodiments or the conventional technology will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0035] Figure 1 This is a schematic diagram of a multi-target segmentation method for chest X-ray images provided in an embodiment of this application;

[0036] Figure 2 A schematic diagram of the multi-object segmentation network method for training a single encoder-dual decoder provided in the embodiments of this application;

[0037] Figure 3 This is a schematic diagram of the densely connected block of the encoder provided in an embodiment of this application;

[0038] Figure 4This is a schematic diagram of the attention-guided multi-scale feature selection module in the decoder provided in the embodiments of this application;

[0039] Figure 5 The results of all bone ablation experiments in the RCS-CXR test set provided in this embodiment of the invention;

[0040] Figure 6 The JSRT test set provides experimental results of clavicle and cardiac ablation in this embodiment of the invention;

[0041] Figure 7 This is a schematic block diagram of an electronic device provided in an embodiment of this disclosure. Detailed Implementation

[0042] To make the above-mentioned objectives, features, and advantages of this application more apparent and understandable, the specific embodiments of this application are described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of this application. However, this application can be implemented in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of this application. Therefore, this application is not limited to the specific embodiments disclosed below.

[0043] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein in the specification of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0044] Figure 1 This is a schematic diagram of a multi-target segmentation method for chest X-ray images provided in an embodiment of this application.

[0045] In a first aspect, embodiments of this application provide a method for multi-target segmentation of chest X-ray images, combined with Figure 1 It includes two steps, S101 and S102:

[0046] S101: Obtain an annotated chest X-ray image; the annotated areas include at least: lung fields, ribs, clavicle, and heart.

[0047] S102: Input the labeled chest X-ray image into the trained single encoder-dual decoder multi-target segmentation network to obtain different target segmentation results.

[0048] Figure 2 This is a schematic diagram of a multi-target segmentation network method for training a single encoder-dual decoder provided in an embodiment of this application.

[0049] The training steps for the single encoder-dual decoder multi-object segmentation network trained in S102 include two steps, S201 to S202:

[0050] S201: Obtain a dataset of labeled chest X-ray images, wherein the image dataset includes at least: RCS-CXR, VinDr-RibCXR, JSRT, and Montgomery;

[0051] S202: Input the image dataset into the single encoder-dual decoder multi-object segmentation network for training to obtain the trained single encoder-dual decoder multi-object segmentation network;

[0052] Among them, RCS-CXR labeled the anterior ribs, posterior ribs, and clavicle; VinDr-RibCXR labeled the anterior and posterior ribs as a whole; JSRT labeled the clavicle, lung fields, and heart; and Montgomery only labeled the lung fields.

[0053] Furthermore, in the aforementioned method for multi-target segmentation of chest X-ray images, step S102 involves inputting the labeled chest X-ray image into a trained single-encoder-dual-decoder multi-target segmentation network to obtain different target segmentation results, including:

[0054] The labeled chest X-ray image is input into a single encoder to extract the image features;

[0055] The extracted image features are input into a dual decoder to obtain different target segmentation results.

[0056] Furthermore, in the above-mentioned method for multi-target segmentation of chest X-ray images, the single encoder includes at least: a 3*3 convolution, a densely connected block, and a collaborative attention skip connection module;

[0057] Among them, 3*3 convolutions and densely connected blocks are used to extract features from multi-object labeled datasets;

[0058] The collaborative attention skip connection module is used to extract spatial details related to the region to be segmented from the encoder, reduce background noise, and mitigate the impact of the semantic gap in the encoder-decoder feature mapping on segmentation.

[0059] Figure 3 This is a schematic diagram of a densely connected block of an encoder provided in an embodiment of this application.

[0060] Specifically, in this embodiment, a 3x3 convolution is first used in the encoder. Then, densely connected blocks with stronger feature extraction capabilities are used as the basic blocks of the encoder to extract more robust feature information from multiple partially labeled datasets and improve the reusability of features. A schematic diagram of dense connections is shown below. Figure 3As shown, to effectively preserve image boundary and spatial detail information related to the segmented region, a collaborative attention skip connection module is proposed to mine spatial detail information related to the region to be segmented in the encoder, while reducing the impact of background noise and semantic gap in the encoder-decoder feature mapping on segmentation.

[0061] Figure 4 This is a schematic diagram of the attention-guided multi-scale feature selection module in the decoder provided in this embodiment of the application.

[0062] Furthermore, in the above-mentioned method for multi-target segmentation of chest X-ray images, both the first decoder and the second decoder in the dual decoder include: an attention-guided multi-scale feature selection module and a single-layer pre-activated residual block;

[0063] like Figure 4 As shown, the attention-guided multi-scale feature selection module contains three branches. Each branch consists of a compression-activation module and 3*3 convolutions with different dilation rates. The scaling factor of the compression-activation module in each branch is 16. The attention-guided multi-scale feature selection module reduces the number of channels in each branch to half the number of input channels, which is represented by C / 2 in the figure. The single-layer pre-activated residual block is used to construct the decoder related to the segmentation task.

[0064] Specifically, in this embodiment, a decoder related to the segmentation task is constructed using pre-activated residual blocks with weighted normalization.

[0065] Furthermore, in the above-mentioned method for multi-target segmentation of chest X-ray images, the first decoder in the decoder is used to process the segmentation of irregular targets, including: ribs and clavicles; the second decoder in the decoder is used to process the segmentation of large-area targets and near-circular targets, including at least: lung fields and heart.

[0066] In some embodiments, the feature separation module uses 1*1 convolution to separate the common features extracted by the encoder-decoder module and learns the private features of different targets.

[0067] Furthermore, in the aforementioned method for multi-target segmentation of chest X-ray images, the overall training of the single encoder-dual decoder multi-target segmentation network adopts an alternating training strategy.

[0068] Specifically, in the embodiments of this application, an alternating training strategy is adopted. For example, when training with the RCS-CXR and JSRT datasets, the RCS-CXR dataset is used as input for training when the training period is odd, and the JSRT dataset is used as input for training when the period is even.

[0069] It should be understood that the strategy of alternating training during the training phase allows the four partially labeled datasets to participate in the training process of the same network, achieving the goal of segmenting multiple targets in chest X-rays using the same network structure, which greatly improves the efficiency of data utilization.

[0070] Furthermore, the aforementioned method for multi-target segmentation of chest X-ray images also includes: training the encoder and dual decoder parameters in a unified manner to obtain the optimal multi-target segmentation network parameters of a single encoder-dual decoder.

[0071] In some embodiments, the optimal parameters for a single encoder-dual decoder multi-object segmentation network are obtained by controlling the proportion of different datasets involved in training.

[0072] Furthermore, the aforementioned method for multi-target segmentation of chest X-ray images also includes:

[0073] During training, the Dice loss function is used in all multi-object segmentation networks with a single encoder and dual decoders.

[0074] When both the RCS-CXR and JSRT datasets are used for training, the weights of the Dice loss function for each target are set to 1, and the average of the branches of the first decoder and the second decoder is calculated.

[0075] When RCS-CXR, JSRT, VinDr-RibCXR, and the Montgomery dataset are all used for training, the clavicle weight of the first decoder is set to 1.5, the heart weight of the second decoder is set to 2, and the weights of the Dice loss function for the other targets are set to 1.

[0076] For example, the Dice loss function is used for overall network training in both branches (as shown below), where y is the gold standard and x is the network prediction.

[0077]

[0078] Specifically, in this application embodiment, different weight coefficients are set for the loss of different decoders to obtain the optimal multi-target segmentation network parameters of single encoder-dual decoder.

[0079] To verify the effectiveness of the proposed method, experiments were conducted using four publicly available datasets. All methods were implemented on an Ubuntu 20.04 operating system and an NVIDIA GeForce RTX 3090. An alternating training strategy was employed during training. For example, when training with the RCS-CXR and JSRT datasets, the RCS-CXR dataset was used as input for odd-numbered training epochs, and the JSRT dataset was used for even-numbered epochs. The maximum number of iterations was set to 100. After 40 training epochs, the learning rate was reduced to 0.1 times its original value every 20 epochs. When training with four datasets—RCS-CXR, VinDr-RibCXR, JSRT, and Montgomery—the maximum number of iterations was set to 150. After 60 training epochs, the learning rate was reduced to 0.1 times its original value every 40 epochs. During training, the initial learning rate was set to 0.0003, and the Adam optimizer was selected. The batch size for training is set to 2, and to 1 during validation and testing.

[0080] As shown in Table 1, the addition of the Collaborative Attention Skip Connector (CAS) module significantly improved the segmentation performance of ribs on the RCS-CXR test set, increasing the DSC scores for anterior ribs, posterior ribs, and all bones by 4.66%, 3.22%, and 2.19%, respectively. It also improved the segmentation performance of the heart and clavicle on the JSRT dataset. The CAS module effectively filtered important information from skeletal features such as ribs and targets like the lung field and heart, thus improving rib segmentation performance. The addition of the Attention-Guided Multi-Scale Feature Selection (AMFS) module improved the segmentation performance of the clavicle, posterior ribs, and lung field. Optimal segmentation performance was achieved when data augmented using an adversarial network-based method was added to the CAS and AMFS modules.

[0081] Table 1 Ablation experimental results for each target on the RCS-CXR test set

[0082]

[0083] To better observe the role of each module in target segmentation, Figure 5 and Figure 6 The segmentation results for all targets, including bones, clavicle, and heart, were visualized. Figure 5Black boxes were used to label areas with significant differences in skeletal segmentation. The original image shows that poorly segmented areas are mainly concentrated in low-contrast and uneven grayscale regions. The addition of the Collaborative Attention Skip Connection (CAS) module enhances the network's perception ability in low-contrast areas. Furthermore, the Attention-Guided Multi-Scale Feature Selection (AMFS) module demonstrates excellent performance in rib segmentation continuity, and the addition of data augmentation further improves segmentation accuracy and effectively reduces false positive regions. Figure 6 The original image shows the gold standard outline of the clavicle and heart. It reveals that the contrast between the heart and surrounding tissues is extremely low, with blurred boundaries, making it difficult to distinguish the heart from other tissues. The basic network exhibits either severe oversegmentation or undersegmentation in heart segmentation. The addition of the Collaborative Attention Skip Connection (CAS) module and the Attention-Guided Multi-Scale Feature Selection (AMFS) module improves the segmentation results, making them closer to the heart's edge region and more complete.

[0084] To verify the inventiveness of this application in multi-object segmentation of chest X-rays, the experimental results of this application were compared with several classic medical image segmentation networks or segmentation networks associated with single-part labeled datasets. The comparative experimental results on the JSRT, VinDr-RibCXR, and Montgomery datasets are summarized in Tables 2, 3, and 4, respectively.

[0085] As shown in Table 2, this application outperforms other algorithms in lung field and clavicle segmentation on the JSRT test set. Specifically, its clavicle segmentation performance is 1.40 percentage points better than the best clavicle segmentation algorithm, TMI, and second only to the DED-CNN algorithm by 0.25 percentage points in heart segmentation. This further demonstrates that using multiple partially labeled datasets for multi-object segmentation of chest X-rays is feasible and beneficial.

[0086] Table 2 Comparative experiments on the JSRT test set

[0087]

[0088] As shown in Table 3, the rib segmentation performance of this application outperforms other algorithms on the VinDr-RibCXR test set. U-Net w. EfficientNet-B0 represents a network using EfficientNet-B0 as the encoder and the original UNet structure as the decoder; UNet++ w. EfficientNet-B0 is similar. Table 3 shows that when using EfficientNet-B0 as the encoding structure, the network performance is improved compared to the original UNet and UNet++, but still lower than the algorithm in this application. The algorithm in this application outperforms the best algorithm by 1.07 percentage points in rib segmentation performance.

[0089] Table 3 Comparative experiments on the VinDr-RibCXR test set

[0090]

[0091] As shown in Table 4, the lung field segmentation performance of this application is superior to other algorithms on the Montgomery test set, with a DSC index 0.33 percentage points higher than DED-CNN and a Jaccard index 0.61 percentage points higher.

[0092] Table 4 Comparative experiments on the Montgomery test set

[0093]

[0094] In a second aspect, embodiments of the present invention also provide an electronic device, including: a processor and a memory;

[0095] The processor executes any of the above-described methods for multi-target segmentation of chest X-ray images by calling programs or instructions stored in the memory.

[0096] Thirdly, embodiments of the present invention also provide a computer-readable storage medium storing a program or instructions that cause a computer to perform any of the above-described methods for multi-target segmentation of chest X-ray images.

[0097] Figure 7 This is a schematic block diagram of an electronic device provided in an embodiment of this disclosure.

[0098] like Figure 7As shown, the electronic device includes at least one processor 701, at least one memory 702, and at least one communication interface 703. The various components of the electronic device are coupled together via a bus system 704. The communication interface 703 is used for information transmission with external devices. It is understood that the bus system 704 is used to implement communication between these components. In addition to a data bus, the bus system 704 also includes a power bus, a control bus, and a status signal bus. However, for clarity, ... Figure 7 The general designated all buses as Bus System 704.

[0099] It is understood that the memory 702 in this embodiment may be volatile memory or non-volatile memory, or may include both volatile and non-volatile memory.

[0100] In some implementations, memory 702 stores elements such as executable units or data structures, or subsets thereof, or extended sets thereof: operating systems and applications.

[0101] The operating system includes various system programs, such as the framework layer, core library layer, and driver layer, used to implement various basic business functions and handle hardware-based tasks. The application programs include various applications, such as media players and browsers, used to implement various application functions. A program implementing any method in the multi-target segmentation method for chest X-ray images provided in this application embodiment can be included in the application programs.

[0102] In this embodiment, the processor 701 executes the steps of various embodiments of the multi-target segmentation method for chest X-ray provided in this application by calling the program or instructions stored in the memory 702, specifically, the program or instructions stored in the application.

[0103] Obtain labeled chest X-ray images; the labeled areas include at least: lung fields, ribs, clavicle, and heart;

[0104] Annotated chest X-ray images are input into a trained single encoder-dual decoder multi-object segmentation network to obtain different object segmentation results.

[0105] The training steps for the trained single-encoder-dual-decoder multi-object segmentation network include:

[0106] Obtain a dataset of labeled chest X-ray images, which includes at least: RCS-CXR, VinDr-RibCXR, JSRT, and Montgomery;

[0107] The image dataset is input into a single encoder-dual decoder multi-object segmentation network for training to obtain a trained single encoder-dual decoder multi-object segmentation network;

[0108] Among them, RCS-CXR labeled the anterior ribs, posterior ribs, and clavicle; VinDr-RibCXR labeled the anterior and posterior ribs as a whole; JSRT labeled the clavicle, lung fields, and heart; and Montgomery only labeled the lung fields.

[0109] Any of the methods in the multi-target segmentation method for chest X-ray images provided in this application embodiment can be applied to, or implemented by, the processor 701. The processor 701 can be an integrated circuit chip with signal capabilities. During implementation, each step of the above method can be completed by the integrated logic circuitry in the hardware of the processor 701 or by instructions in software form. The processor 701 can be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. The general-purpose processor can be a microprocessor or any conventional device.

[0110] The steps of any method in the multi-target segmentation method for chest X-ray images provided in this application embodiment can be directly implemented by a hardware decoding processor, or implemented by a combination of hardware and software units in the decoding processor. The software units can reside in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. This storage medium is located in memory 702. The processor 701 reads the information in memory 702 and, in conjunction with its hardware, completes the steps of the multi-target segmentation method for chest X-ray images.

[0111] Those skilled in the art will understand that although some embodiments described herein include certain features included in other embodiments but not others, combinations of features from different embodiments are meant to be within the scope of this application and form different embodiments.

[0112] Those skilled in the art will understand that the descriptions of the various embodiments have different focuses, and for parts not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0113] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for multi-object segmentation of chest X-ray, characterized in that, The method comprises the following steps: obtaining a labeled chest X-ray image; the labeled parts at least include lung field, rib, clavicle and heart; inputting the labeled chest X-ray image into a trained single-encoder-double-decoder multi-target segmentation network to obtain different target segmentation results; wherein the training steps of the trained single-encoder-double-decoder multi-target segmentation network comprise: obtaining a labeled chest X-ray image dataset, the image dataset at least including RCS-CXR, VinDr-RibCXR, JSRT and Montgomery; inputting the image dataset into a single-encoder-double-decoder multi-target segmentation network for training to obtain a trained single-encoder-double-decoder multi-target segmentation network; wherein RCS-CXR labels the anterior rib, posterior rib and clavicle; VinDr-RibCXR labels the anterior rib and posterior rib as a whole, JSRT labels the clavicle, lung field and heart; and Montgomery only labels the lung field; inputting the labeled chest X-ray image into the trained single-encoder-double-decoder multi-target segmentation network to obtain different target segmentation results, comprising: inputting the labeled chest X-ray image into a single encoder to extract image features; inputting the extracted image features into a double decoder to obtain different target segmentation results; the single encoder at least includes 3*3 convolution, dense connection block and collaborative attention jump connection module; wherein the 3*3 convolution and dense connection block are used to extract features from the multi-target labeled dataset; the collaborative attention jump connection module is used to mine spatial detail information related to the region to be segmented in the encoder, reduce background noise, and reduce the influence of semantic gap in the feature mapping between the encoder and the decoder on segmentation; the first decoder and the second decoder in the double decoder both include an attention-guided multi-scale feature selection module and a single-layer pre-activation residual block; the attention-guided multi-scale feature selection module includes three branches, each branch is composed of a compression-excitation module and a 3*3 convolution with different hole rates, the scaling factor of the compression-excitation module in each branch is 16, and the attention-guided multi-scale feature selection module reduces the channel number of each branch to half of the input channel number; the single-layer pre-activation residual block is used to build a decoder related to the segmentation task; the first decoder in the decoder is used to process the segmentation of irregular targets, the irregular targets including rib, clavicle; the second decoder in the decoder is used to process the segmentation of large-area targets and circular targets, the large-area targets and circular targets at least including lung field and heart; the overall training of the single-encoder-double-decoder multi-target segmentation network adopts an alternating training strategy; the method further comprises: during the training process, the single-encoder-double-decoder multi-target segmentation network uses a Dice loss function; when RCS-CXR and JSRT datasets are involved in the training, the Dice loss function weight of each target is set to 1, and the branches of the first decoder and the second decoder are averaged. When the RCS-CXR, JSRT, VinDr-RibCXR and Montgomery datasets are all involved in training, the clavicle weight of the first decoder is set to 1.5, the heart weight of the second decoder is set to 2, and the weights of the Dice loss functions of the remaining targets are set to 1.

2. The method of claim 1, wherein, The method further comprises: uniformly training parameters of the encoder and the double decoder to obtain optimal single-encoder-double-decoder multi-target segmentation network parameters.

3. An electronic device, comprising: Comprise: A processor and a memory; The processor is configured to execute the chest X-ray multi-target segmentation method according to any one of claims 1-2 by invoking programs or instructions stored in the memory.

4. A computer-readable storage medium, characterized in that, The computer readable storage medium stores programs or instructions, which enable the computer to execute the chest X-ray multi-target segmentation method according to any one of claims 1-2.

Images