A hyperspectral image-based tissue segmentation method and system

Abstract

The application discloses a kind of based on hyperspectral image tissue segmentation method and system, comprising: by constructing feature distillation network and applying pathological priori constraint etc., fine segmentation is carried out to hyperspectral tissue image, accurate analysis to cell tissue is realized.Specific steps include dataset creation, training data expansion, feature distillation network construction, model training and optimization, test set forward calculation, pathological priori constraint calculation, tissue fine segmentation.The application has the advantages of high precision, self-adaptability, simple structure, strong practicability and optimizability, and can be widely applied in medical image diagnosis, biomedical research and other fields.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer vision, in particular to a hyperspectral image-based tissue segmentation method and system. BACKGROUND

[0002] Hyperspectral imaging technology is a technology for obtaining spectral information of an object or scene in a continuous wavelength range. Compared with traditional color imaging, hyperspectral imaging can provide more detailed and comprehensive spectral information. The image generated by hyperspectral imaging is called a spectral image, which contains spectral information of each pixel point in a continuous wavelength range, which can provide detailed information on the composition, chemical properties, optical properties, etc. of the target object or scene. Through data processing and analysis of the spectral image of the target object or scene, corresponding problems can be solved according to different needs.

[0003] Hyperspectral imaging technology is a non-contact, non-invasive imaging method. Hyperspectral images have hundreds of narrow and detailed bands from visible light to near-infrared light. The hyperspectral image corresponding to biomedical tissue can obtain a series of images in a continuous spectral range, with rich spatial and spectral information, and has the characteristics of "space-spectrum integration". Different categories of substances have different absorption and transmission of light intensity in the same band, and the spectral curves they present are also different, thereby providing a basis for detection and discrimination of different categories of substances.

[0004] At present, although there are some segmentation strategies based on hyperspectral images, such as using 3D convolution to enhance the correlation between each band, using pyramid features to propose multi-scale information of data, etc. However, due to the lack of sufficient feature expression and adaptive ability of the network constructed by these methods, it cannot well mine the effective information of high-dimensional images, which limits the performance of the model. SUMMARY

[0005] The present application provides a hyperspectral image-based tissue segmentation method and system to overcome the defects of the prior art. The network performance is better improved, the representation ability of high-dimensional information is improved, and an efficient feature distillation module is designed. The module extracts deep features of data through the lower branch, refines the features through the upper branch, and enriches the feature information in the module by using the parallel mode of the upper and lower branches. Finally, through the channel attention unit, useless features are effectively suppressed and useful features are enhanced, so as to achieve the purpose of feature distillation, enhance the expression ability of the network, and improve the performance of the network.

[0006] In order to achieve the above application purposes, the technical solutions adopted by the present application are as follows:

[0007] A hyperspectral image-based tissue segmentation method, comprising the following steps:

[0008] Step 1, obtaining a training data set;

[0009] Obtaining hyperspectral tissue data to establish a data set library, and labeling the data set library to obtain label images corresponding to the hyperspectral images, and combining the hyperspectral images and the label images Figure One corresponding to each other to form a set as a training data set;

[0010] Step 2, expanding the training data set;

[0011] Randomly cropping the training data set obtained in step 1, and expanding the training data set by using a random rotation and flipping data augmentation method;

[0012] Step 3, constructing a feature distillation network;

[0013] The feature distillation network is composed of five feature distillation modules and four decoding modules; the input image of the feature distillation network is obtained after the feature distillation module A to obtain the intermediate feature map and , passing through the feature distillation module B, the feature distillation module C, the feature distillation module D and the feature distillation module E in turn, respectively obtaining the intermediate feature , , , , , , ; taking and as the input of the decoding module A to obtain the intermediate feature , and as the input of the decoding module B to obtain the intermediate feature , and as the input of the decoding module C to obtain the intermediate feature map , and after passing through the decoding module D, the output coarse segmentation result image is obtained, and the construction of the feature distillation network is completed;

[0014] The feature distillation module comprises eight convolution units, a segmentation layer and a down-sampling layer; the input of the feature distillation module is , wherein is an integer, and , after passing through the convolution unit A to obtain the intermediate feature map ; after continuously passing through the convolution unit B, the convolution unit C and the convolution unit D, the intermediate feature map is obtained; taking Input into the segmentation layer for segmentation, to obtain two sub-feature maps and ; After continuously passing through the convolution unit E, the convolution unit F and the convolution unit G, an intermediate feature map is obtained ; and After dimension splicing, an intermediate feature map is obtained , and and After pixel-by-pixel addition, an intermediate feature map is obtained ; After passing through the channel attention unit and the convolution unit H, an intermediate feature is obtained ; After passing through the down-sampling layer, an intermediate feature is obtained .

[0015] Step 4, training the model;

[0016] The augmented training set data obtained in step 2 is input into the feature distillation network constructed in step 3, and a cross-entropy loss function is used as the optimization objective for training. The model parameters are updated using an SGD optimizer until the loss no longer decreases, and a trained tissue segmentation model is obtained;

[0017] Step 5, tissue coarse segmentation;

[0018] The tissue segmentation model trained in step 4 is used to segment the test images in the test set, and a tissue coarse segmentation result is obtained;

[0019] Step 6, pathological prior constraint;

[0020] All pathological label maps in the tissue segmentation training set data are counted, and all labeled areas of cancerous cell nuclei are traversed. Let the area of each labeled area of cancerous cell nuclei be , where is a positive integer and , representing the number of labeled areas; let the minimum labeled area be , then the minimum labeled area is calculated; in each training set label, the distance between the two nearest cancerous cell nuclei is calculated, denoted as , is a positive integer and , representing the number of training labels; after traversing all training set labels, a set representing the distance between the two nearest cancerous cell nuclei in each training set is obtained , so is obtained; the obtained and are used to construct a pathological prior constraint;

[0021] Step 7, fine segmentation is performed by using pathological prior constraints;

[0022] The coarse segmentation obtained in step 5 is subjected to secondary region fine division; first, the coarse segmentation result is denoised, and all regions smaller than in the coarse segmentation image are removed to obtain a segmentation result ; then, the region of a cancerous cell nucleus in is taken as a center, and a single cancerous cell region is approximated by a circle with a radius of to perform secondary division of the tissue region, and a tissue fine segmentation result is obtained.

[0023] Further, the decoding module in step 3 includes an upsampling layer and two convolution units; the decoding module inputs are and After the upsampling layer, the obtained intermediate features are dimensionally spliced with , and then sequentially pass through the convolution unit I and the convolution unit J to obtain the output .

[0024] The application further discloses a tissue segmentation system based on hyperspectral images, which can be used to implement the above-mentioned tissue segmentation method based on hyperspectral images, and specifically comprises a dataset creation module, a feature distillation network construction module, a model training optimization module, a tissue coarse segmentation module, a pathological prior constraint module and a tissue fine segmentation module.

[0025] The dataset creation module is used to create a hyperspectral tissue dataset library, label the corresponding hyperspectral images, correspond the hyperspectral images and labels Figure One , and expand the training dataset;

[0026] The feature distillation network construction module is used to construct a feature distillation network, including the construction of five feature distillation modules and four decoding modules, to realize feature extraction and segmentation of the hyperspectral tissue images;

[0027] The model training optimization module is used to input the expanded training dataset into the feature distillation network, use a cross-entropy loss function to train the model, and update the model parameters through an optimizer to realize optimization of the model;

[0028] The tissue coarse segmentation module is used to segment the test images in the test set to obtain a tissue coarse segmentation result;

[0029] The pathological prior constraint module is used to calculate pathological prior constraints according to the cancerous cell nucleus label area and the distance between two nearest cancerous cell nuclei in the training set, to be used for subsequent secondary region fine division;

[0030] The tissue fine segmentation module is used for denoising and secondary region fine division of the tissue coarse segmentation result, so as to obtain a tissue fine segmentation result.

[0031] The application further discloses a tissue segmentation system based on a hyperspectral image.

[0032] The data set creation module is used for creating a hyperspectral tissue data set library, and labeling corresponding hyperspectral images. Figure One The hyperspectral images and labels are matched and the training data set is expanded.

[0033] The feature distillation network construction module is used for constructing a feature distillation network, including construction of five feature distillation modules and four decoding modules, so as to realize feature extraction and segmentation of the hyperspectral tissue image.

[0034] The model training optimization module is used for inputting the expanded training data set into the feature distillation network, training the model by using a cross-entropy loss function, and updating the model parameters by using an optimizer, so as to realize optimization of the model.

[0035] The tissue coarse segmentation module is used for segmenting the to-be-tested image in the test set, so as to obtain a tissue coarse segmentation result.

[0036] The pathological prior constraint module is used for calculating a pathological prior constraint according to a cancer cell nucleus label area and a distance between two nearest cancer cell nuclei in the training set, so as to realize secondary region fine division.

[0037] The tissue fine segmentation module is used for denoising and secondary region fine division of the tissue coarse segmentation result, so as to obtain a tissue fine segmentation result.

[0038] The application further discloses a computer device, which comprises a memory, a processor and a computer program stored in the memory and capable of running on the processor.

[0039] The application further discloses a computer readable storage medium, which stores a computer program.

[0040] Compared with the prior art, the application has the following advantages:

[0041] Firstly, the present application designs an efficient feature distillation module to address the lack of feature expression capability of existing hyperspectral image-based segmentation models for high-dimensional data. The module extracts deep features of the data through the lower branch and refines the features through the upper branch, enriching the feature information within the module through the parallel upper and lower branches. Finally, through the channel attention unit, useless features are effectively suppressed and useful features are enhanced, thereby achieving the purpose of feature distillation, enhancing the expression capability of the network, and improving the performance of the network.

[0042] Secondly, the present application constructs two stages of learning and decision-making. In the learning stage, the feature distillation network is used to obtain the coarse segmentation result of the tissue. In the decision-making stage, the prior information constraint of pathology is constructed according to the existing data information and characteristics, and the segmentation region is further refined to obtain the fine segmentation result. From coarse segmentation to fine segmentation, the present application fully excavates the deep information of high-dimensional data and constructs constraint conditions according to the characteristics of the data, and the two stages promote each other to improve the segmentation performance and refine the segmentation region.

[0043] Thirdly, the present application has strong practicability and can be widely applied in medical image diagnosis, biomedical research and other fields, and can provide reliable cell tissue analysis tools for doctors and researchers. BRIEF DESCRIPTION OF DRAWINGS

[0044] Figure 1 The tissue segmentation method based on hyperspectral images of the embodiment of the present application is shown in the flowchart.

[0045] Figure 2 The feature distillation network structure diagram of the embodiment of the present application is shown in the flowchart.

[0046] Figure 3 The feature distillation module structure diagram of the embodiment of the present application is shown in the flowchart.

[0047] Figure 4 The decoding module structure diagram of the embodiment of the present application is shown in the flowchart.

[0048] Figure 5 The convolution unit structure diagram of the embodiment of the present application is shown in the flowchart.

[0049] Figure 6 The channel attention unit structure diagram of the embodiment of the present application is shown in the flowchart.

[0050] Figure 7 The decision-making stage structure diagram of the embodiment of the present application is shown in the flowchart.

[0051] Figure 8 The coarse segmentation result comparison diagram of the embodiment of the present application and other methods under the same conditions is shown in the flowchart.

[0052] Figure 9 The fine segmentation result comparison diagram of the embodiment of the present application and other methods under the same conditions is shown in the flowchart. DETAILED DESCRIPTION

[0053] In order to make the objectives, technical solutions and advantages of the present application clearer, the present application will be further described in detail below according to the accompanying drawings and examples.

[0054] Example 1

[0055] As shown in the drawings, the present application provides a tissue segmentation method based on hyperspectral images, comprising the following steps: Figure 1

[0056] Step 1, obtaining a training data set;

[0057] Obtaining hyperspectral tissue data to establish a data set library, and labeling it to obtain a label image corresponding to the hyperspectral image, and the hyperspectral image and the label image are matched to form a set as the training data set. Figure One Step 2, expanding the training data set;

[0058] Step 2, expanding the training data set;

[0059] Randomly cropping the training data set obtained in step 1, and expanding the training data set by using random rotation and flip data augmentation.

[0060] Step 3, constructing a feature distillation network;

[0061] Figure 2 As shown in the drawings, the feature distillation network of the embodiment of the present application is a structure diagram, which is composed of five feature distillation modules and four decoding modules; the size of the input of the network is a hyperspectral image, and after passing through the feature distillation module A1, an intermediate feature map with a size of , ;​​​​​​​​​​​​​​​​​​​ and As the input of the decoding module A1, the intermediate feature with the size of is obtained , and As the input of the decoding gate module B2, the intermediate feature with the size of is obtained , and As the input of the decoding module C3, the intermediate feature with the size of is obtained , and After the decoding module D4, the coarse segmentation result image with the output size of is obtained, and the construction of the feature distillation network is completed.

[0062] Figure 3 As shown in the structural diagram of the feature distillation module of the embodiment of the present application, the function of the feature distillation module is to extract and enhance the deep features of the information, so as to enhance the useful features and inhibit the useless features, and achieve the purpose of enhancing the expression ability of the network. The input of each stage feature distillation module is the input image in Figure 2 , , , , .

[0063] Figure 4 As shown in the structural diagram of the decoder module of the embodiment of the present application, the function is to combine the features of different information and improve the nonlinear expression ability of the network. The decoder module has the same structure, which is composed of an up-sampling layer, a dimension splicing operation and two convolution units; the input of each decoding module corresponds to Figure 2 , and , and , and , and .

[0064] Figure 5 As shown in the structural diagram of the convolution unit of the embodiment of the present application, the convolution unit has the same structure, which is composed of a convolution layer, a BN layer and a ReLU layer; the function of the convolution unit is to extract the feature information of the target input into the convolution unit.

[0065] Figure 6The figure shows a channel attention unit structure of an embodiment of the application. The channel attention unit is composed of a pooling layer, two convolution layers, a ReLU layer and a Sigmiod layer; the role of the channel attention unit is to adaptively adjust the features input into the channel attention unit, and to calculate the weights of different information amount feature maps using the global and local correlation between channels.

[0066] Figure 7 The figure shows a decision stage structure of an embodiment of the application. In the decision stage, a pathological constraint is constructed according to pathological prior information, and the constraint is used to perform secondary fine region segmentation on the coarse segmentation result.

[0067] Step 4, training the model;

[0068] The extended training set data obtained in step 2 is input into the feature distillation network constructed in step 3, and a cross-entropy loss function is used as the optimization objective for training. The model parameters are updated using an SGD optimizer until the loss no longer decreases, and a trained tissue segmentation model is obtained.

[0069] Step 5, tissue coarse segmentation;

[0070] The tissue segmentation model trained in step 4 is used to segment the test images in the test set, and the tissue coarse segmentation result is obtained.

[0071] Step 6, tissue coarse segmentation;

[0072] The pathological label images in the tissue segmentation training set data 1 are counted, and all the labeled regions of cancerous nuclei are traversed. The area of each labeled region of cancerous nuclei is set to , where is a positive integer and represents the number of labeled regions; the minimum labeled area is set to , and the minimum labeled area can be calculated; the distance between the two nearest cancerous nuclei in each training set label is calculated, denoted as , is a positive integer and represents the number of training labels; after traversing all the training set labels, a set of distances between the two nearest cancerous nuclei in each training set is obtained , and can be obtained; the and obtained above are used to construct a pathological prior constraint.

[0073] Step 7, fine segmentation using the pathological prior constraint;

[0074] The coarse segmentation obtained in step 5 is subjected to secondary regional fine division; first, the coarse segmentation result is denoised, and all cells smaller than in the coarse segmentation image are removed to obtain a segmentation result ; then, the cancer cell nucleus region in is taken as the center of a circle, and is taken as the radius to approximate the region of a single cancer cell with a circle to perform secondary division of the tissue region, and a tissue fine segmentation result is obtained.

[0075] Example 2

[0076] The method in example 1 (except steps 6 and 7) is used to perform image coarse segmentation experiments on a hyperspectral test data set. The computer environment in this experiment is: the operating system is Linux Ubuntu 18.04 version, an NVIDIA 309024G GPU, and the software platform is PyTorch.

[0077] Figure 8 The results obtained by the embodiment of the present application and other methods under the same experimental conditions are shown in the comparison chart. Figure 8 In Figure 8 , from top to bottom and from left to right are, respectively, a histological section (H&E) stained with H&E, SQNet, DSI-Net, CGNet, HyperNet, BiSeNet V2, SegNet, UNet, ResUNet, Deeplab V3, RefineNet, and the coarse segmentation image obtained in the learning stage of the present application; the first row to the second row are example 1, and the third row to the fourth row are example 2. From Figure 8 it can be seen that, compared with other methods, the present application can achieve better segmentation effect and the noise interference is the smallest.In Example 1, other segmentation methods all have noise, and the loss degree of the segmented region is different, among which Hyper-Net and BiSeNet V2 are the most obvious, while the application can obtain a more accurate segmented region with minimal noise interference. This is because these methods do not reasonably use high-dimensional information in the feature extraction stage, resulting in redundant information interfering with the network expression, while the application can effectively mine deep information of high-dimensional data. In Example 2, DSI-Net, SQNet and SegNet cannot identify effective segmentation regions. In the coarse segmentation results of RefineNet, CGNet, BiSeNet V2 and Deep LabV3, the shape of the cancer cell nucleus is distorted in the segmentation results, and a more complete segmentation region cannot be obtained. The method of the application obtains a more complete segmentation region, and the clear edge of the cancer cell nucleus is visible. The main reason is that the proposed feature distillation block module provides a more suitable and richer combination of receptive fields to capture more accurate local structures and details.

[0078] Example 3

[0079] The steps 6 and 7 in Example 1 are used to perform a secondary fine segmentation experiment on the coarse segmentation result obtained in Example 2. The computer environment of this experiment is: the operating system is Linux Ubuntu 18.04 version, one NVIDIA 309024G GPU, and the software platform is PyTorch.

[0080] Figure 9 The results obtained by the application and other methods under the same experimental conditions are shown in the comparison chart. Figure 9 In Example 1, the results obtained by the application and other methods under the same experimental conditions are shown in the comparison chart. Figure 9 As can be seen from Example 1, compared with other methods, the application can obtain the most accurate segmentation result closest to the immunohistochemical result image.

[0081] In this embodiment, mean intersection over union (MIoU), overall classification accuracy (OA), per-class classification accuracy (AA), sensitivity (SE) and DICE coefficient are used to evaluate the fine segmentation results obtained by different methods.

[0082] In another embodiment of the application, a tissue segmentation system based on hyperspectral images is provided, which can be used to implement the above-mentioned tissue segmentation method based on hyperspectral images. Specifically, it comprises: a module

[0083] Dataset creation module: used to create a library of hyperspectral tissue datasets and label corresponding hyperspectral images, and expand the training dataset. Figure One

[0084] Feature distillation network construction module: used to construct a feature distillation network, including the construction of five feature distillation modules and four decoding modules, to achieve feature extraction and segmentation of hyperspectral tissue images.

[0085] Model training and optimization module: used to input the expanded training dataset into the feature distillation network, use the cross-entropy loss function for model training, and update the model parameters through the optimizer to achieve model optimization.

[0086] Tissue coarse segmentation module: used to segment the test images in the test set to obtain tissue coarse segmentation results.

[0087] Pathological prior constraint module: used to calculate the pathological prior constraint based on the area of cancerous cell nucleus label and the distance between the two nearest cancerous cell nuclei in the training set, for subsequent secondary region fine division.

[0088] Tissue fine segmentation module: used to denoise and perform secondary region fine division on the tissue coarse segmentation results to obtain tissue fine segmentation results.

[0089] Before applying a tissue segmentation system based on hyperspectral images to actual scenarios, the model needs to be deployed, integrating the trained model into software systems or mobile applications so that it can be directly used by users.

[0090] Users can upload hyperspectral tissue images to be tested and use the deployed model to perform fine segmentation and analysis of cell tissues. For medical imaging diagnosis scenarios, doctors can use this module to analyze medical images and diagnose whether cell tissues have abnormalities. For the field of biomedical research, scientists can use this module to analyze cell tissues and help researchers understand the function and structure of cell tissues.

[0091] When using the system, users need to provide hyperspectral tissue images to be tested, and the module will automatically segment and analyze the images, outputting fine segmentation results of cell tissues and related statistical information. Users can make judgments and analyses based on the output results of the module to complete the fine segmentation and analysis of cell tissues.

[0092] ​In still another embodiment of the present application, a terminal device is provided, which comprises a processor and a memory, the memory being configured to store a computer program, the computer program comprising program instructions, and the processor being configured to execute the program instructions stored in the computer storage medium. The processor can be a central processing unit (CPU), and can also be other general-purpose processors, a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic device, discrete gate or transistor logic, discrete hardware components, etc., which are the computing core and control core of the terminal, and are suitable for implementing one or more instructions, and are particularly suitable for loading and executing one or more instructions to implement a corresponding method flow or a corresponding function; the processor in the embodiments of the present application can be used for the operation of a tissue segmentation method based on hyperspectral images, comprising the following steps:

[0093] Step 1, obtaining a training data set;

[0094] Step 2, expanding the training data set;

[0095] Step 3, constructing a feature distillation network;

[0096] Step 4, training a model;

[0097] Step 5, coarse segmentation of tissues;

[0098] Step 6, coarse segmentation of tissues;

[0099] Step 7, fine segmentation using pathological prior constraints.

[0100] In another embodiment of the present application, the present application further provides a storage medium, specifically a computer readable storage medium (Memory), which is a memory device in the terminal device, used for storing programs and data. It can be understood that the computer readable storage medium herein can include the built-in storage medium in the terminal device, and of course can also include the expansion storage medium supported by the terminal device. The computer readable storage medium provides a storage space, which stores the operating system of the terminal. In addition, one or more instructions suitable for being loaded and executed by the processor are also stored in the storage space, and the instructions can be one or more computer programs (including program codes). It should be noted that the computer readable storage medium herein can be a high-speed RAM memory, or a non-volatile memory such as at least one disk memory.

[0101] The one or more instructions stored in the computer readable storage medium can be loaded and executed by the processor to implement the corresponding steps of the above-mentioned embodiment of the tissue segmentation method based on hyperspectral images. The one or more instructions stored in the computer readable storage medium are loaded and executed by the processor to implement the following steps:

[0102] Step 1, obtaining a training data set;

[0103] Step 2, expanding the training data set;

[0104] Step 3, constructing a feature distillation network;

[0105] Step 4, training the model;

[0106] Step 5, coarse segmentation of tissues;

[0107] Step 6, coarse segmentation of tissues;

[0108] Step 7, fine segmentation using pathological prior constraints.

[0109] It should be understood by those skilled in the art that the embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can be in the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can be in the form of a computer program product implemented on one or more computer usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer usable program codes.

[0110] The embodiments of methods, hardware systems, software systems, or computer program products described herein can be implemented using computer- executable instructions, such as program modules being executed by a computer. Generally, program modules include routines, programs, objects, components, data structures, etc. that perform particular tasks or implement particular abstract data types. Computer-executable instructions, associated data structures, and program modules represent examples of the program code means for implementing the steps described herein. The particular sequence of instructions can not be required to achieve the functions taught in Figure One one or more functions specified in the flow or flows and / or blocks Figure One one or more functions specified in the flow or flows and / or blocks

[0111] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instructions which implement the flow Figure One one or more functions specified in the flow or flows and / or blocks Figure One one or more functions specified in the flow or flows and / or blocks

[0112] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the flow Figure One one or more functions specified in the flow or flows and / or blocks Figure One Figure One one or more functions specified in the flow or flows and / or blocks

[0113] Those of skill in the art would further appreciate that the various illustrative logical blocks, modules, circuits, and algorithm steps described in connection with the embodiments disclosed herein can be implemented as electronic hardware, computer software, or combinations of both. The various illustrative components, blocks, modules, circuits, and steps have been described above generally in terms of their functionality, without limitation to the particular sequences of functions executed on a computer, a general purpose computer, a microprocessor or a microprocessor array or other programmable data processing apparatus, a general purpose computer configured by a computer program, a dedicated computer, or a computerized device programmed to perform particular tasks. As should be readily appreciated, the functionalities of some of the various components, blocks, modules, circuits, and steps have been described herein as being performed by passage of signals through the relevant hardware, software, firmware, microcode or microinstructions on a computer, a processor or other programmable data processing apparatus, whether such computer, processor or other programmable apparatus is dedicated to the relevant function or employed for other functions as well. It should further be appreciated that the various logical blocks, modules, circuits, and steps described herein are implemented or performed by one or more electronic components configured to perform the described functions or steps, by instructions stored in computer- readable medium, or by a combination of hardware, software, and / or instructions stored in computer-readable medium. Those of skill would further appreciate that the various illustrative logical blocks, modules, circuits, and algorithm steps described herein are implemented as electronic hardware, computer software, or combinations of both, and that the various hardware and software components so found in the computer program product or other computer- readable medium are given with the sole purpose to illustrate preferred embodiments of the application and are not intended to limit the scope of the present application in any way.

Claims

1. A method for tissue segmentation based on hyperspectral images, characterized in that, The method comprises the following steps: Step 1, obtaining a training data set; Step 2, expanding the training data set; Step 3, constructing a feature distillation network; Step 4, training a model; Step 5, performing coarse segmentation of tissues; The feature distillation network is composed of five feature distillation modules and four decoding modules; the input image of the feature distillation network is obtained after the feature distillation module A to obtain the intermediate feature map and , After sequentially passing through the feature distillation module B, the feature distillation module C, the feature distillation module D and the feature distillation module E, the intermediate feature , , , , , , ; take and as the input of the decoding module A to obtain the intermediate feature , and as the input of the decoding module B to obtain the intermediate feature , and as the input of the decoding module C to obtain the intermediate feature map , and After passing through the decoding module D, the output coarse segmentation result image is obtained, and the construction of the feature distillation network is completed; The feature distillation module comprises eight convolution units, a split layer, and a down-sampling layer; the feature distillation module input is wherein is an integer, and , After passing through the convolution unit A, an intermediate feature map is obtained. After sequentially passing through the convolution unit B, the convolution unit C, and the convolution unit D, an intermediate feature map is obtained. is input into the split layer for splitting to obtain two sub-feature maps and . After sequentially passing through the convolution unit E, the convolution unit F, and the convolution unit G, an intermediate feature map is obtained. and are obtained by dimension splicing to obtain an intermediate feature map . and are pixel-by-pixel added to obtain an intermediate feature map . After passing through the channel attention unit and the convolution unit H, an intermediate feature is obtained. After passing through the down-sampling layer, an output of the feature distillation module is obtained . Step 6, performing pathological prior constraint; Step 7, performing fine segmentation of tissues by using the pathological prior constraint; The system can be used to implement the tissue segmentation method based on hyperspectral images according to claim 1 or 2, and specifically comprises a data set creation module, a feature distillation network construction module, a model training optimization module, a tissue coarse segmentation module, a pathological prior constraint module, and a tissue fine segmentation module. The data set creation module is used to create a hyperspectral tissue data set library, label corresponding hyperspectral images, and expand the training data set. The feature distillation network construction module is used to construct a feature distillation network, including the construction of five feature distillation modules and four decoding modules, to realize feature extraction and segmentation of hyperspectral tissue images. The statistics of all pathological label maps in the tissue segmentation training set data 1 are performed, all labeled areas of cancerous nuclei are traversed, the area of each labeled area of cancerous nuclei is set as , wherein is a positive integer and , represents the number of labeled areas; the minimum labeled area is set as , the minimum labeled area is calculated; the distance between the two closest cancerous nuclei in each training set label is calculated as , is a positive integer and , represents the number of training labels; after traversing all training set labels, a set representing the distance between the two closest cancerous nuclei in each training set is obtained, and is obtained; the obtained and are used to construct a pathological prior constraint; The model training optimization module is used to input the expanded training data set into the feature distillation network, use a cross-entropy loss function to train the model, and update the model parameters through an optimizer to optimize the model. The coarse segmentation obtained in step 5 is subjected to secondary region fine division; first, the coarse segmentation result is denoised, and all regions smaller than in the coarse segmentation image are removed to obtain a segmentation result ; then, the cancer cell nucleus region in is taken as the center of a circle, is taken as the radius, and the region of a single cancer cell is approximated by a circle to perform secondary division of the tissue region, obtaining a tissue fine segmentation result.

2. The method of claim 1, wherein: The decoding module in step 3 comprises an up-sampling layer and two convolution units; the decoding module input is and ; After the up-sampling layer, the obtained intermediate feature is dimensionally spliced with , and then sequentially passed through the convolution unit I and the convolution unit J to obtain the output .

3. A hyperspectral image-based tissue segmentation system, characterized by: The tissue coarse segmentation module is used to segment the test images in the test set to obtain coarse segmentation results of tissues. The pathological prior constraint module is used to calculate the pathological prior constraint according to the labeled area of the cancer cell nucleus and the distance between the two nearest cancer cell nuclei in the training set, to be used for subsequent secondary region fine division. The tissue fine segmentation module is used to denoise and perform secondary region fine division on the coarse segmentation results of tissues to obtain fine segmentation results of tissues. The computer program is stored in the memory and can be executed on the processor, and the processor executes the program to implement the tissue segmentation method based on hyperspectral images according to claim 1 or 2. The computer program is stored in the memory and can be executed on the processor, and the processor executes the program to implement the tissue segmentation method based on hyperspectral images according to claim 1 or 2. ​ ​ 4. A computer device, characterized by: ​ 5. A computer-readable storage medium, characterized in that: ​

Images