A Multi-Organ Cell Nucleus Segmentation Method Based on Cue Learning

Abstract

A multi-organ cell nucleus segmentation method based on cue-based learning, belonging to the field of medical image processing technology, fully mines image information by utilizing text and image multimodal information, learns the correlation between semantic information and segmentation targets, and comprehensively learns the segmentation of target regions. Based on the clip model, it learns a large amount of text and image pairing knowledge from six publicly available cell nucleus datasets to obtain semantic understanding prior knowledge of cell nuclei, making the model fully suitable for cell nucleus segmentation tasks. The model is constructed by inputting images and text cues, utilizing text and image multimodal information to complete the identification and accurate segmentation of cell nuclei of six different organs, with higher computational efficiency. This model can also complete accurate segmentation tasks on datasets with insufficient annotations using sufficient text cues, making it more practical and scalable.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, specifically to a method for multi-organ cell nucleus segmentation based on cue learning. Background Technology

[0002] Most current medical image segmentation models focus solely on improving and exploring methods for image data, with limited research on segmentation guided by multiple modalities. They often only consider visual factors and lack comprehensive learning for target region segmentation. In pathological slide analysis, multi-organ cell segmentation is a recognized challenge, with most datasets specifically designed for particular cell nucleus segmentation. Therefore, segmentation models are typically built on partially labeled datasets or datasets specific to a particular organ, thus only segmenting the nuclei of a single organ. While an increasing number of segmentation models demonstrate excellent performance in segmenting single-organ cell nuclei, research on segmenting multiple organ cell nuclei using a single model remains limited. To enable segmentation models to recognize more categories, they need to be retrained to achieve segmentation capabilities. Therefore, utilizing multimodal information from text and images to guide segmentation models for segmenting cell nuclei of multiple organ categories is a significant challenge. Summary of the Invention

[0003] To overcome the shortcomings of the above technologies, this invention provides a method for accurately completing unsupervised segmentation of cell nuclei of a specified organ on a dataset lacking annotations, with sufficient textual prompts.

[0004] The technical solution adopted by this invention to overcome its technical problems is:

[0005] A multi-organ cell nucleus segmentation method based on cue learning includes the following steps:

[0006] a) Collect N images of brain cell nuclei, N images of kidney cell nuclei, N images of liver cell nuclei, N images of breast cell nuclei, N images of colon cell nuclei, and N images of stomach cell nuclei to obtain a cell nucleus image-text dataset T.

[0007]

[0008] T i brain The image represents the i-th brain cell nucleus. "a nuclei photo of [brain]" is a medical text template for the brain. T i kidney The image shows the i-th kidney cell nucleus. "a nuclei photo of [kidney]" is a medical text template for kidney information. T i liverThe image shows the i-th liver cell nucleus. "a nuclei photo of [liver]" is a medical text template for the liver. T i breast This is the image of the i-th breast cell nucleus, where "a nuclei photo of [breast]" is a medical text template for breast information, and "T" represents the image. i colon This is the image of the i-th colon cell nucleus, where "a nuclei photo of [colon]" is a medical text template for the colon, and T... i stomach Let i be the image of the i-th gastric cell nucleus, and let "a nuclei photo of [stomach]" be the medical information template for the stomach.

[0009] b) Extract the i-th brain cell nucleus image T from the cell nucleus image-text dataset T. i brain Medical prompts for the brain: a nuclei photo of [brain], image of the i-th kidney cell nucleus T i kidney Medical cues for the kidneys: a nuclei photo of [kidney], image of the i-th liver cell nucleus (T). i liver Medical cues related to the liver: a nuclei photo of [liver], image of the i-th breast cell nucleus (T). i breast Medical prompt text template related to breast: a nuclei photo of [breast], image of the i-th colon cell nucleus T i colon Medical cues related to the kidneys: anuclei photo of [colon], image of the i-th gastric cell nucleus T i stomach The medical information template for the stomach, a nucleiphoto of [stomach], is input into the clip model to obtain the optimized clip model;

[0010] c) Obtain a brain cell nucleus images, b kidney cell nucleus images, c liver cell nucleus images, d breast cell nucleus images, e colon cell nucleus images, and f stomach cell nucleus images from the NuInsSeg multi-organ category dataset. a+b+c+d+e+f=n, resulting in the cell nucleus image-text dataset Y.

[0011] , where Y ibrain For the i-th brain cell nucleus image, Y i kidney For the i-th kidney cell nucleus image, Y i liver For the i-th liver cell nucleus image, Y i breast For the i-th image of a breast cell nucleus, Y i colon For the i-th colon cell nucleus image, Y i stomach This is the image of the i-th gastric cell nucleus;

[0012] d) Divide the cell nucleus image-text dataset Y into a training set and a test set, and scale the cell nucleus images in the training set to 572×572;

[0013] e) Construct a segmentation network model, which consists of a text module, an image module, and an MLP module;

[0014] f) Input the medical prompt text templates from the training set into the text module and output the text vector;

[0015] g) Input the cell nucleus images from the training set into the image module of the segmentation network model, and output the cell nucleus segmentation result image and feature vector;

[0016] h) Input the feature vector into the MLP module of the segmentation network model and output the parameters;

[0017] i) Update the segmentation network model by updating the parameters to obtain the updated segmentation network model;

[0018] j) Train the updated segmentation network model to obtain the optimized segmentation network model;

[0019] k) Input the cell nucleus images from the test set into the optimized segmentation network model, and output the final segmentation result image.

[0020] Preferably, N is 500.

[0021] Further, in step a), N brain cell nucleus images are collected from the MoNuSeg dataset and / or the CMP-15 dataset and / or the CMP-17 dataset and / or the CryonNuSeg dataset and / or the NuInsSeg dataset; N kidney cell nucleus images are collected from the MoNuSeg dataset and / or the kumor dataset and / or the Irshad dataset and / or the CryonNuSeg dataset and / or the MoNuSAC dataset and / or the NuInsSeg dataset; N liver cell nucleus images are collected from the MoNuSeg dataset and / or the CryonNuSeg dataset and / or the Crowedsource dataset and / or the kumor dataset and / or the NuInsSeg dataset; and N liver cell nucleus images are collected from the MoNuSeg dataset and / or the BCNup dataset and / or the MoNuSAC dataset and / or the Nucls dataset and / or the TNBC dataset and / or the Janowczyk dataset. Collect N mammary cell nucleus images from the mCryonNuSeg dataset and / or Gelasca dataset and / or Naylor dataset and / or MoNuSAC dataset and / or NulunsSeg dataset and / or kumor dataset; collect N colon cell nucleus images from the CoNsep dataset and / or CRCHisto dataset and / or CryonNuSeg dataset and / or NuInsSeg dataset and / or kumor dataset; collect N gastric cell nucleus images from the MoNuSeg CryonNuSeg dataset and / or Wienert dataset and / or NuInsSeg dataset and / or kumor dataset.

[0022] Preferably, in step d), the cell nucleus image-text dataset Y is divided into a training set and a test set in a 7:3 ratio.

[0023] Furthermore, step f) includes the following steps:

[0024] The text module of the f-1) segmentation network model is composed of an optimized clip model;

[0025] f-2) Input the medical prompt text template of the brain from the training set (a nuclei photo of [brain]) into the text module, and output the text vector N. brain N brain ∈R L×N Let R be the real number space, L be the text length, and N be the length of the last word in the text. Then, let the text vector N be... brain The text vector N′ is obtained by expanding the channel dimension using the torch.unsqueeze function in Python's PyTorch library. brain ;

[0026] f-3) Input the medical prompt text template for the kidney from the training set, such as "a nuclei photo of [kidney]", into the text module, and output the text vector N. kidney N kidney ∈R L×N , the text vector N kidney The text vector N′ is obtained by expanding the channel dimension using the torch.unsqueeze function in Python's PyTorch library. kidney ;

[0027] f-4) Input the medical prompt text template of liver (a nuclei photo of [liver]) from the training set into the text module, and output the text vector N. liver N liver ∈R L×N , the text vector N liver The text vector N′ is obtained by expanding the channel dimension using the torch.unsqueeze function in Python's PyTorch library. liver ;

[0028] f-5) Input the medical prompt text template for breast tissue from the training set, such as "a nuclei photo of [breast]", into the text module, and output the text vector N. breast N breast ∈R L×N , the text vector N breast The text vector N′ is obtained by expanding the channel dimension using the torch.unsqueeze function in Python's PyTorch library. breast ;

[0029] f-6) Input the medical prompt text template for the kidney from the training set, "a nuclei photo of [colon]", into the text module, and output the text vector N. colon N colon ∈R L×N , the text vector N colon The text vector N′ is obtained by expanding the channel dimension using the torch.unsqueeze function in Python's PyTorch library. colon ;

[0030] f-7) Input the medical prompt text template for the stomach from the training set, "a nuclei photo of [stomach]", into the text module, and output the text vector N. stomach N stomach ∈R L×N, the text vector N stomach The text vector N′ is obtained by expanding the channel dimension using the torch.unsqueeze function in Python's PyTorch library. stomach .

[0031] Furthermore, step g) includes the following steps:

[0032] The image module of the g-1 segmentation network model consists of an image encoder, an image decoder, and a GAP module;

[0033] The image encoder of the g-2) image module consists of a first CRM module, a second CRM module, a third CRM module, a fourth CRM module, and a fifth CRM module. The first, second, third, and fourth CRM modules each consist of a first convolutional layer, a first ReLU activation function, a second convolutional layer, a second ReLU activation function, and a max-pooling layer, respectively. The fifth CRM module consists of a first convolutional layer, a first ReLU activation function, a second convolutional layer, and a second ReLU activation function, respectively. The image encoder of the i-th brain cell nucleus image Y from the training set is used as the input. i brain The input is given to the first CRM module, and the output is the feature. Features The input is fed into the second CRM module, and the output is the feature. Features The input is fed into the third CRM module, and the output is the feature. Features The input is fed into the fourth CRM module, and the output is the feature. Features The input is fed into the fifth CRM module, and the output is the feature. The i-th kidney cell nucleus image Y in the training set i kidney The input is given to the first CRM module, and the output is the feature. Features The input is fed into the second CRM module, and the output is the feature. Features The input is fed into the third CRM module, and the output is the feature. Features The input is fed into the fourth CRM module, and the output is the feature. Features The input is fed into the fifth CRM module, and the output is the feature. The i-th liver cell nucleus image Y in the training set i liver The input is given to the first CRM module, and the output is the feature. Features The input is fed into the second CRM module, and the output is the feature. Features The input is fed into the third CRM module, and the output is the feature. Features The input is fed into the fourth CRM module, and the output is the feature. Features The input is fed into the fifth CRM module, and the output is the feature. The i-th breast cell nucleus image Y in the training set i breast The input is given to the first CRM module, and the output is the feature. Features The input is fed into the second CRM module, and the output is the feature. Features The input is fed into the third CRM module, and the output is the feature. Features The input is fed into the fourth CRM module, and the output is the feature. Features The input is fed into the fifth CRM module, and the output is the feature. The i-th colon cell nucleus image Y in the training set i colon The input is given to the first CRM module, and the output is the feature. Features The input is fed into the second CRM module, and the output is the feature. Features The input is fed into the third CRM module, and the output is the feature. Features The input is fed into the fourth CRM module, and the output is the feature. Features The input is fed into the fifth CRM module, and the output is the feature. The i-th gastric cell nucleus image Y in the training set i stomach The input is given to the first CRM module, and the output is the feature. Features The input is fed into the second CRM module, and the output is the feature. Features The input is fed into the third CRM module, and the output is the feature. Features The input is fed into the fourth CRM module, and the output is the feature. Features The input is fed into the fifth CRM module, and the output is the feature. The image decoder of the g-3 image module consists of a first GRU module, a second GRU module, a third GRU module, and a fourth GRU module. Each of these modules comprises a first convolutional layer, a first ReLU activation function, a second convolutional layer, a second ReLU activation function, and an upsampling layer, sequentially processing the features. The input is given to the first CRM module, and the output is the feature. Features The input is fed into the second CRM module, and the output is the feature. Features The input is fed into the third CRM module, and the output is the feature. Features The input is fed into the fourth CRM module, and the output is a brain cell nucleus segmentation result image. Features The input is given to the first CRM module, and the output is the feature. Features The input is fed into the second CRM module, and the output is the feature. Features The input is fed into the third CRM module, and the output is the feature. Features The input is fed into the fourth CRM module, and the output is a segmentation result image of kidney cell nuclei. Features The input is given to the first CRM module, and the output is the feature. Features The input is fed into the second CRM module, and the output is the feature. Features The input is fed into the third CRM module, and the output is the feature. Features The input is fed into the fourth CRM module, which outputs a liver cell nucleus segmentation result image. Features The input is given to the first CRM module, and the output is the feature. Features The input is fed into the second CRM module, and the output is the feature. Features The input is fed into the third CRM module, and the output is the feature. Features The input is fed into the fourth CRM module, and the output is a segmentation result image of breast cell nuclei. Features The input is given to the first CRM module, and the output is the feature. Features The input is fed into the second CRM module, and the output is the feature. Features The input is fed into the third CRM module, and the output is the feature. Features The data is input into the fourth CRM module, and the output is a segmentation result image of colon cell nuclei. Features The input is given to the first CRM module, and the output is the feature. Features The input is fed into the second CRM module, and the output is the feature. Features The input is fed into the third CRM module, and the output is the feature. Features The input is fed into the fourth CRM module, and the output is a segmentation result image of the gastric cell nuclei.

[0034] The GAP module of the g-4 image module consists of a Batch Normalization (BN) layer, a ReLU activation function, and an adaptive average pooling layer, which integrates features. The input is fed into the GAP module and the output is the feature PG. brain , will feature The input is fed into the GAP module and the output is the feature PG. kidney , will feature The input is fed into the GAP module and the output is the feature PG. liver , will feature The input is fed into the GAP module and the output is the feature PG. breast , will feature The input is fed into the GAP module and the output is the feature PG. colon , will feature The input is fed into the GAP module and the output is the feature PG. stomach , will feature PG brain With text vector N′ brain Perform a concatenation operation to obtain the feature vector. Feature PG kidney With text vector N′ kidney Perform a concatenation operation to obtain the feature vector. PG liver With text vector N′ liver Perform a concatenation operation to obtain the feature vector. Feature PG breast With text vector N′ breast Perform a concatenation operation to obtain the feature vector. Feature PG colon With text vector N′ colon Perform a concatenation operation to obtain the feature vector. Feature PG stomach With text vector N′ stomachPerform a concatenation operation to obtain the feature vector. Furthermore, step h) includes the following steps:

[0035] The MLP module of the segmentation network model (h-1) consists of a first convolutional layer, a second convolutional layer, and a third convolutional layer, with the kernel size of the first, second, and third convolutional layers all being 1*1.

[0036] h-2) eigenvectors The input is fed into the MLP module, and the output is the feature. eigenvectors The input is fed into the MLP module, and the output is the feature. eigenvectors The input is fed into the MLP module, and the output is the feature. eigenvectors The input is fed into the MLP module, and the output is the feature. eigenvectors The input is fed into the MLP module, and the output is the feature. eigenvectors The input is fed into the MLP module, and the output is the feature. h-3) features The input is given to the Sigmoid function, and the output is the parameter. Features The input is given to the Sigmoid function, and the output is the parameter. Features The input is given to the Sigmoid function, and the output is the parameter. Features The input is given to the Sigmoid function, and the output is the parameter. Features The input is given to the Sigmoid function, and the output is the parameter. Features The input is given to the Sigmoid function, and the output is the parameter. Furthermore, step i) includes the following steps:

[0037] i-1) via parameters Using the `reshape` function from the PyTorch library in Python, the first and second convolutional layers of the first, second, third, fourth, and fifth CRM modules of the image encoder are reshaped, followed by sequential convolution and ReLU activation function operations to complete the first update of the image encoder. The first update of the image decoder is completed by using the reshape function in the PyTorch library of Python to reshape the first convolutional layer and the second convolutional layer of the first GRU module, the second GRU module, the third GRU module and the fourth GRU module. Then, convolution and ReLU activation function operations are performed in sequence.

[0038] i-2) via parameters Using the `reshape` function from the PyTorch library in Python, the first and second convolutional layers of the first, second, third, fourth, and fifth CRM modules of the image encoder are reshaped, followed by sequential convolution and ReLU activation function operations to complete the second update of the image encoder. The first convolutional layer and the second convolutional layer of the first GRU module, the second GRU module, the third GRU module, and the fourth GRU module of the image decoder are reshaped using the reshape function in the PyTorch library of Python. Then, convolution and ReLU activation function operations are performed in sequence to complete the second update of the image decoder.

[0039] i-3) via parameters Using the `reshape` function from the PyTorch library in Python, the first and second convolutional layers of the first, second, third, fourth, and fifth CRM modules of the image encoder are reshaped, followed by convolution and ReLU activation function operations, completing the third update of the image encoder. The first convolutional layer and the second convolutional layer of the first GRU module, the second GRU module, the third GRU module, and the fourth GRU module of the image decoder are reshaped using the reshape function in the PyTorch library of Python. Then, convolution and ReLU activation function operations are performed in sequence to complete the third update of the image decoder.

[0040] i-4) via parameters Using the `reshape` function from the PyTorch library in Python, the first and second convolutional layers of the first, second, third, fourth, and fifth CRM modules of the image encoder are reshaped, followed by sequential convolution and ReLU activation function operations, completing the fourth update of the image encoder. The first convolutional layer and the second convolutional layer of the first GRU module, the second GRU module, the third GRU module and the fourth GRU module of the image decoder are reshaped using the reshape function in the PyTorch library of Python. Then, convolution and ReLU activation function operations are performed in sequence to complete the fourth update of the image decoder.

[0041] i-5) via parameters Using the `reshape` function from the PyTorch library in Python, the first and second convolutional layers of the first, second, third, fourth, and fifth CRM modules of the image encoder are reshaped, followed by sequential convolution and ReLU activation function operations to complete the fifth update of the image encoder. Using the reshape function in the PyTorch library of Python, the first convolutional layer and the second convolutional layer of the first GRU module, the second GRU module, the third GRU module and the fourth GRU module of the image decoder are reshaped and then convolution and ReLU activation function operations are performed in sequence to complete the fifth update of the image decoder.

[0042] i-6) via parameters Using the `reshape` function from the PyTorch library in Python, the first and second convolutional layers of the first, second, third, fourth, and fifth CRM modules of the image encoder are reshaped, followed by sequential convolution and ReLU activation function operations, completing the sixth update of the image encoder. The reshape function in the PyTorch library of Python is used to reshape the first and second convolutional layers of the first, second, third, and fourth GRU modules of the image decoder. Then, convolution and ReLU activation functions are performed sequentially to complete the sixth update of the image decoder and obtain the updated segmentation network model.

[0043] Furthermore, in step j), the Adam optimizer is used to train the updated segmentation network model using the DSC loss function to obtain the optimized segmentation network model.

[0044] Furthermore, step k) includes the following steps:

[0045] k-1) The i-th brain cell nucleus image Y in the training set i brainThe images are sequentially input into the image encoder and image decoder of the optimized segmentation network model, and the final segmentation result image of the brain cell nucleus is output.

[0046] k-2) The i-th kidney cell nucleus image Y in the training set i kidney The images are sequentially input into the image encoder and image decoder of the optimized segmentation network model, and the final segmentation result image of the kidney cell nucleus is output. k-3) The i-th liver cell nucleus image Y in the training set i liver The images are sequentially input into the image encoder and image decoder of the optimized segmentation network model, and the final segmentation result of the liver cell nucleus is output.

[0047] k-4) The i-th breast cell nucleus image Y in the training set i breast The images are sequentially input into the image encoder and image decoder of the optimized segmentation network model, and the final segmentation result image of the breast cell nucleus is output. k-5) The i-th colon cell nucleus image Y in the training set i colon The images are sequentially input into the image encoder and image decoder of the optimized segmentation network model, and the final segmentation result of the colon cell nucleus is output.

[0048] k-6) The i-th gastric cell nucleus image Y in the training set i stomach The images are sequentially input into the image encoder and image decoder of the optimized segmentation network model, and the final segmentation result of the gastric cell nucleus is output.

[0049] The beneficial effects of this invention are as follows: While an increasing number of segmentation models have demonstrated significant impact on the segmentation of cell nuclei in single organs, research on segmentation models for multiple organ cell nuclei remains limited and cannot be extended to new domains. To address these issues, we propose a text-hint-based segmentation network model. This model fully leverages multimodal information from text and images to learn the correlation between semantic information and the segmentation target, performing comprehensive learning for target region segmentation. Based on the clip model, this model learns a large amount of text and image pairing knowledge from six publicly available cell nucleus datasets to acquire prior knowledge of cell nucleus semantic understanding, making the model perfectly suited for cell nucleus segmentation tasks. By inputting images and text hints, the model utilizes multimodal information from text and images to accurately identify and segment cell nuclei in six different organs, achieving higher computational efficiency. Furthermore, this model can also achieve accurate segmentation tasks on datasets lacking annotations using sufficient text hints, making it more practical and scalable. Attached Figure Description

[0050] Figure 1 This is a network structure diagram of the present invention. Detailed Implementation

[0051] The following is in conjunction with the appendix Figure 1 The present invention will be further described below.

[0052] A multi-organ cell nucleus segmentation method based on cue learning includes the following steps:

[0053] a) Collect N images of brain cell nuclei, N images of kidney cell nuclei, N images of liver cell nuclei, N images of breast cell nuclei, N images of colon cell nuclei, and N images of stomach cell nuclei to obtain a cell nucleus image-text dataset T.

[0054]

[0055] T i brain The image represents the i-th brain cell nucleus. "a nuclei photo of [brain]" is a medical text template for the brain. T i kidney The image shows the i-th kidney cell nucleus. "a nuclei photo of [kidney]" is a medical text template for kidney information. T i liver The image shows the i-th liver cell nucleus. "a nuclei photo of [liver]" is a medical text template for the liver. T i breastThis is the image of the i-th breast cell nucleus, where "a nuclei photo of [breast]" is a medical text template for breast information, and "T" represents the image. i colon This is the image of the i-th colon cell nucleus, where "a nuclei photo of [colon]" is a medical text template for the colon, and T... i stomach Let i be the image of the i-th gastric cell nucleus, and let "a nuclei photo of [stomach]" be the medical information template for the stomach.

[0056] b) Extract the i-th brain cell nucleus image T from the cell nucleus image-text dataset T. i brain Medical prompts for the brain: a nuclei photo of [brain], image of the i-th kidney cell nucleus T i kidney Medical cues for the kidneys: a nuclei photo of [kidney], image of the i-th liver cell nucleus (T). i liver Medical cues related to the liver: a nuclei photo of [liver], image of the i-th breast cell nucleus (T). i breast Medical prompt text template related to breast: a nuclei photo of [breast], image of the i-th colon cell nucleus T i colon Medical cues related to the kidneys: anuclei photo of [colon], image of the i-th gastric cell nucleus T i stomach The medical information template for the stomach, a nucleiphoto of [stomach], is input into the clip model to obtain the optimized clip model;

[0057] c) Obtain a brain cell nucleus images, b kidney cell nucleus images, c liver cell nucleus images, d breast cell nucleus images, e colon cell nucleus images, and f stomach cell nucleus images from the NuInsSeg multi-organ category dataset. a+b+c+d+e+f=n, resulting in the cell nucleus image-text dataset Y.

[0058] , where Y i brain For the i-th brain cell nucleus image, Y i kidney For the i-th kidney cell nucleus image, Y iliver For the i-th liver cell nucleus image, Y i breast For the i-th image of a breast cell nucleus, Y i colon For the i-th colon cell nucleus image, Y i stomach This is the image of the i-th gastric cell nucleus;

[0059] d) Divide the cell nucleus image-text dataset Y into a training set and a test set, and scale the cell nucleus images in the training set to 572×572;

[0060] e) Construct a segmentation network model, which consists of a text module, an image module, and an MLP module;

[0061] f) Input the medical prompt text templates from the training set into the text module and output the text vector;

[0062] g) Input the cell nucleus images from the training set into the image module of the segmentation network model, and output the cell nucleus segmentation result image and feature vector;

[0063] h) Input the feature vector into the MLP module of the segmentation network model and output the parameters;

[0064] i) Update the segmentation network model by updating the parameters to obtain the updated segmentation network model;

[0065] j) Train the updated segmentation network model to obtain the optimized segmentation network model;

[0066] k) Input the cell nucleus images from the test set into the optimized segmentation network model, and output the final segmentation result image.

[0067] This paper presents a multi-organ cell nucleus segmentation method based on cue learning, which can segment cell nuclei of multiple different organ categories using multimodal information from text and images. Furthermore, on unlabeled datasets, it can accurately complete unsupervised segmentation tasks of specified organ cell nuclei with sufficient textual cues.

[0068] In one embodiment of the present invention, N is set to 500. That is, there are 500 images of cell nuclei for each organ.

[0069] In one embodiment of the present invention, step a) involves collecting N brain cell nucleus images from the MoNuSeg dataset and / or the CMP-15 dataset and / or the CMP-17 dataset and / or the CryonNuSeg dataset and / or the NuInsSeg dataset; collecting N kidney cell nucleus images from the MoNuSeg dataset and / or the kumor dataset and / or the Irshad dataset and / or the CryonNuSeg dataset and / or the MoNuSAC dataset and / or the NuInsSeg dataset; collecting N liver cell nucleus images from the MoNuSeg dataset and / or the CryonNuSeg dataset and / or the Crowedsource dataset and / or the kumor dataset and / or the NuInsSeg dataset; and collecting N liver cell nucleus images from the MoNuSeg dataset and / or the BCNup dataset and / or the MoNuSAC dataset and / or the Nucls dataset and / or the TNBC dataset and / or the Janowczyk dataset. Collect N mammary cell nucleus images from the mCryonNuSeg dataset and / or Gelasca dataset and / or Naylor dataset and / or MoNuSAC dataset and / or NulunsSeg dataset and / or kumor dataset; collect N colon cell nucleus images from the CoNsep dataset and / or CRCHisto dataset and / or CryonNuSeg dataset and / or NuInsSeg dataset and / or kumor dataset; collect N gastric cell nucleus images from the MoNuSeg CryonNuSeg dataset and / or Wienert dataset and / or NuInsSeg dataset and / or kumor dataset.

[0070] In one embodiment of the present invention, in step d), the cell nucleus image-text dataset Y is divided into a training set and a test set in a 7:3 ratio.

[0071] In one embodiment of the present invention, step f) includes the following steps:

[0072] The text module of the f-1) segmentation network model is composed of an optimized clip model;

[0073] f-2) Input the medical prompt text template of the brain from the training set (a nuclei photo of [brain]) into the text module, and output the text vector N. brain N brain ∈R L×N Let R be the real number space, L be the text length, and N be the length of the last word in the text. Then, let the text vector N be... brain The text vector N′ is obtained by expanding the channel dimension using the torch.unsqueeze function in Python's PyTorch library. brain ;

[0074] f-3) Input the medical prompt text template for the kidney from the training set, such as "a nuclei photo of [kidney]", into the text module, and output the text vector N. kidney N kidney ∈R L×N , the text vector N kidney The text vector N′ is obtained by expanding the channel dimension using the torch.unsqueeze function in Python's PyTorch library. kidney ;

[0075] f-4) Input the medical prompt text template of liver (a nuclei photo of [liver]) from the training set into the text module, and output the text vector N. liver N liver ∈R L×N , the text vector N liver The text vector N′ is obtained by expanding the channel dimension using the torch.unsqueeze function in Python's PyTorch library. liver ;

[0076] f-5) Input the medical prompt text template for breast tissue from the training set, such as "a nuclei photo of [breast]", into the text module, and output the text vector N. breast N breast ∈R L×N , the text vector N breast The text vector N′ is obtained by expanding the channel dimension using the torch.unsqueeze function in Python's PyTorch library. breast ;

[0077] f-6) Input the medical prompt text template for the kidney from the training set, "a nuclei photo of [colon]", into the text module, and output the text vector N. colon N colon ∈R L×N , the text vector N colon The text vector N′ is obtained by expanding the channel dimension using the torch.unsqueeze function in Python's PyTorch library. colon ;

[0078] f-7) Input the medical prompt text template for the stomach from the training set, "a nuclei photo of [stomach]", into the text module, and output the text vector N. stomach N stomach ∈RL×N , the text vector N stomach The text vector N′ is obtained by expanding the channel dimension using the torch.unsqueeze function in Python's PyTorch library. stomach .

[0079] In one embodiment of the present invention, step g) includes the following steps:

[0080] The image module of the g-1 segmentation network model consists of an image encoder, an image decoder, and a GAP module;

[0081] The image encoder of the g-2) image module consists of a first CRM module, a second CRM module, a third CRM module, a fourth CRM module, and a fifth CRM module. The first, second, third, and fourth CRM modules each consist of a first convolutional layer, a first ReLU activation function, a second convolutional layer, a second ReLU activation function, and a max-pooling layer, respectively. The fifth CRM module consists of a first convolutional layer, a first ReLU activation function, a second convolutional layer, and a second ReLU activation function, respectively. The image encoder of the i-th brain cell nucleus image Y from the training set is used as the input. i brain The input is given to the first CRM module, and the output is the feature. Y i brain ∈R C×H×W C is the number of image channels, H is the image height, and W is the image width. The feature PE1... brain The input is fed into the second CRM module, and the output is the feature. Features The input is fed into the third CRM module, and the output is the feature. Features The input is fed into the fourth CRM module, and the output is the feature. Features The input is fed into the fifth CRM module, and the output is the feature. The i-th kidney cell nucleus image Y in the training set i kidney The input is given to the first CRM module, and the output is the feature. Y i kidney ∈R C×H×W , will feature The input is fed into the second CRM module, and the output is the feature. Features The input is fed into the third CRM module, and the output is the feature. Features The input is fed into the fourth CRM module, and the output is the feature. Features The input is fed into the fifth CRM module, and the output is the feature. The i-th liver cell nucleus image Y in the training set i liver The input is given to the first CRM module, and the output is the feature. Y i liver ∈R C×H×W , will feature The input is fed into the second CRM module, and the output is the feature. Features The input is fed into the third CRM module, and the output is the feature. Features The input is fed into the fourth CRM module, and the output is the feature. Features The input is fed into the fifth CRM module, and the output is the feature. The i-th breast cell nucleus image Y in the training set i breast The input is given to the first CRM module, and the output is the feature. Y i breast ∈R C×H×W , will feature The input is fed into the second CRM module, and the output is the feature. Features The input is fed into the third CRM module, and the output is the feature. Features The input is fed into the fourth CRM module, and the output is the feature. Features The input is fed into the fifth CRM module, and the output is the feature. The i-th colon cell nucleus image Y in the training set i colon The input is given to the first CRM module, and the output is the feature. Y i colon ∈R C×H×W , will feature The input is fed into the second CRM module, and the output is the feature. Features The input is fed into the third CRM module, and the output is the feature. Features The input is fed into the fourth CRM module, and the output is the feature. Features The input is fed into the fifth CRM module, and the output is the feature. The i-th gastric cell nucleus image Y in the training set istomach The input is given to the first CRM module, and the output is the feature. Y i stomach ∈R C×H×W , will feature The input is fed into the second CRM module, and the output is the feature. Features The input is fed into the third CRM module, and the output is the feature. Features The input is fed into the fourth CRM module, and the output is the feature. Features The input is fed into the fifth CRM module, and the output is the feature. The image decoder of the g-3 image module consists of a first GRU module, a second GRU module, a third GRU module, and a fourth GRU module. Each of these modules comprises a first convolutional layer, a first ReLU activation function, a second convolutional layer, a second ReLU activation function, and an upsampling layer, sequentially processing the features. The input is given to the first CRM module, and the output is the feature. Features The input is fed into the second CRM module, and the output is the feature. Features The input is fed into the third CRM module, and the output is the feature. Features The input is fed into the fourth CRM module, and the output is a brain cell nucleus segmentation result image. Features The input is given to the first CRM module, and the output is the feature. Features The input is fed into the second CRM module, and the output is the feature. Features The input is fed into the third CRM module, and the output is the feature. Features The input is fed into the fourth CRM module, and the output is a segmentation result image of kidney cell nuclei. Features The input is given to the first CRM module, and the output is the feature. Features The input is fed into the second CRM module, and the output is the feature. Features The input is fed into the third CRM module, and the output is the feature. Features The input is fed into the fourth CRM module, which outputs a liver cell nucleus segmentation result image. Features The input is given to the first CRM module, and the output is the feature. Features The input is fed into the second CRM module, and the output is the feature. Features The input is fed into the third CRM module, and the output is the feature. Features The input is fed into the fourth CRM module, and the output is a segmentation result image of breast cell nuclei. Features The input is given to the first CRM module, and the output is the feature. Features The input is fed into the second CRM module, and the output is the feature. Features The input is fed into the third CRM module, and the output is the feature. Features The data is input into the fourth CRM module, and the output is a segmentation result image of colon cell nuclei. Features The input is given to the first CRM module, and the output is the feature. Features The input is fed into the second CRM module, and the output is the feature. Features The input is fed into the third CRM module, and the output is the feature. Features The input is fed into the fourth CRM module, and the output is a segmentation result image of the gastric cell nuclei. The GAP module of the g-4 image module consists of a Batch Normalization (BN) layer, a ReLU activation function, and an adaptive average pooling layer, which integrates features. The input is fed into the GAP module and the output is the feature PG. brain , will feature The input is fed into the GAP module and the output is the feature PG. kidney , will feature The input is fed into the GAP module and the output is the feature PG. liver , will feature The input is fed into the GAP module and the output is the feature PG. breast , will feature The input is fed into the GAP module and the output is the feature PG. colon , will feature The input is fed into the GAP module and the output is the feature PG. stomach , will feature PG brain With text vector N b ′ rain Perform a concatenation operation to obtain the feature vector. Feature PG kidney With text vector N′kidney Perform a concatenation operation to obtain the feature vector. PG liver With text vector N′ liver Perform a concatenation operation to obtain the feature vector. Feature PG breast With text vector N′ breast Perform a concatenation operation to obtain the feature vector. Feature PG colon With text vector N′ colon Perform a concatenation operation to obtain the feature vector. Feature PG stomach With text vector N′ stomach Perform a concatenation operation to obtain the feature vector. In one embodiment of the present invention, step h) includes the following steps:

[0082] The MLP module of the segmentation network model (h-1) consists of a first convolutional layer, a second convolutional layer, and a third convolutional layer, with the kernel size of the first, second, and third convolutional layers all being 1*1.

[0083] h-2) eigenvectors The input is fed into the MLP module, and the output is the feature. eigenvectors The input is fed into the MLP module, and the output is the feature. eigenvectors The input is fed into the MLP module, and the output is the feature. eigenvectors The input is fed into the MLP module, and the output is the feature. eigenvectors The input is fed into the MLP module, and the output is the feature. eigenvectors The input is fed into the MLP module, and the output is the feature.

[0084] h-3) features The input is given to the Sigmoid function, and the output is the parameter. Features The input is given to the Sigmoid function, and the output is the parameter. Features The input is given to the Sigmoid function, and the output is the parameter. Features The input is given to the Sigmoid function, and the output is the parameter. Features The input is given to the Sigmoid function, and the output is the parameter. Features The input is given to the Sigmoid function, and the output is the parameter. In one embodiment of the present invention, step i) includes the following steps:

[0085] i-1) via parameters Using the `reshape` function from the PyTorch library in Python, the first and second convolutional layers of the first, second, third, fourth, and fifth CRM modules of the image encoder are reshaped, followed by sequential convolution and ReLU activation function operations to complete the first update of the image encoder. The first update of the image decoder is completed by using the reshape function in the PyTorch library of Python to reshape the first convolutional layer and the second convolutional layer of the first GRU module, the second GRU module, the third GRU module and the fourth GRU module. Then, convolution and ReLU activation function operations are performed in sequence.

[0086] i-2) via parameters Using the `reshape` function from the PyTorch library in Python, the first and second convolutional layers of the first, second, third, fourth, and fifth CRM modules of the image encoder are reshaped, followed by sequential convolution and ReLU activation function operations to complete the second update of the image encoder. The first convolutional layer and the second convolutional layer of the first GRU module, the second GRU module, the third GRU module, and the fourth GRU module of the image decoder are reshaped using the reshape function in the PyTorch library of Python. Then, convolution and ReLU activation function operations are performed in sequence to complete the second update of the image decoder.

[0087] i-3) via parameters Using the `reshape` function from the PyTorch library in Python, the first and second convolutional layers of the first, second, third, fourth, and fifth CRM modules of the image encoder are reshaped, followed by convolution and ReLU activation function operations, completing the third update of the image encoder. The first convolutional layer and the second convolutional layer of the first GRU module, the second GRU module, the third GRU module, and the fourth GRU module of the image decoder are reshaped using the reshape function in the PyTorch library of Python. Then, convolution and ReLU activation function operations are performed in sequence to complete the third update of the image decoder.

[0088] i-4) via parameters Using the `reshape` function from the PyTorch library in Python, the first and second convolutional layers of the first, second, third, fourth, and fifth CRM modules of the image encoder are reshaped, followed by sequential convolution and ReLU activation function operations, completing the fourth update of the image encoder. The first convolutional layer and the second convolutional layer of the first GRU module, the second GRU module, the third GRU module and the fourth GRU module of the image decoder are reshaped using the reshape function in the PyTorch library of Python. Then, convolution and ReLU activation function operations are performed in sequence to complete the fourth update of the image decoder.

[0089] i-5) via parameters Using the `reshape` function from the PyTorch library in Python, the first and second convolutional layers of the first, second, third, fourth, and fifth CRM modules of the image encoder are reshaped, followed by sequential convolution and ReLU activation function operations to complete the fifth update of the image encoder. Using the reshape function in the PyTorch library of Python, the first convolutional layer and the second convolutional layer of the first GRU module, the second GRU module, the third GRU module and the fourth GRU module of the image decoder are reshaped and then convolution and ReLU activation function operations are performed in sequence to complete the fifth update of the image decoder.

[0090] i-6) via parameters Using the `reshape` function from the PyTorch library in Python, the first and second convolutional layers of the first, second, third, fourth, and fifth CRM modules of the image encoder are reshaped, followed by sequential convolution and ReLU activation function operations, completing the sixth update of the image encoder. The reshape function in the PyTorch library of Python is used to reshape the first and second convolutional layers of the first, second, third, and fourth GRU modules of the image decoder. Then, convolution and ReLU activation functions are performed sequentially to complete the sixth update of the image decoder and obtain the updated segmentation network model.

[0091] In one embodiment of the present invention, in step j), the Adam optimizer is used to train the updated segmentation network model using the DSC (Dice Similarity Coefficient) loss function to obtain the optimized segmentation network model.

[0092] In one embodiment of the present invention, step k) includes the following steps:

[0093] k-1) The i-th brain cell nucleus image Y in the training set i brain The images are sequentially input into the image encoder and image decoder of the optimized segmentation network model, and the final segmentation result image of the brain cell nucleus is output.

[0094] k-2) The i-th kidney cell nucleus image Y in the training set i kidney The images are sequentially input into the image encoder and image decoder of the optimized segmentation network model, and the final segmentation result image of the kidney cell nucleus is output.

[0095] k-3) The i-th liver cell nucleus image Y in the training set i liver The images are sequentially input into the image encoder and image decoder of the optimized segmentation network model, and the final segmentation result of the liver cell nucleus is output.

[0096] k-4) The i-th breast cell nucleus image Y in the training set i breast The images are sequentially input into the image encoder and image decoder of the optimized segmentation network model, and the final segmentation result image of the breast cell nucleus is output.

[0097] k-5) The i-th colon cell nucleus image Y in the training set i colon The images are sequentially input into the image encoder and image decoder of the optimized segmentation network model, and the final segmentation result of the colon cell nucleus is output.

[0098] k-6) The i-th gastric cell nucleus image Y in the training set i stomach The images are sequentially input into the image encoder and image decoder of the optimized segmentation network model, and the final segmentation result of the gastric cell nucleus is output.

[0099] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A multi-organ cell nucleus segmentation method based on cue-based learning, characterized in that, Includes the following steps: a) Collect N images of brain cell nuclei, N images of kidney cell nuclei, N images of liver cell nuclei, N images of breast cell nuclei, N images of colon cell nuclei, and N images of stomach cell nuclei to obtain a cell nucleus image-text dataset T. Where T i brain The image represents the i-th brain cell nucleus. "a nuclei photo of [brain]" is a medical text template for the brain. T i kidney The image shows the i-th kidney cell nucleus. "a nuclei photo of [kidney]" is a medical text template for kidney information. T i liver The image shows the i-th liver cell nucleus. "a nuclei photo of [liver]" is a medical text template for the liver. T i breast This is the image of the i-th breast cell nucleus, where "a nuclei photo of [breast]" is a medical text template for breast information, and "T" represents the image. i colon This is the image of the i-th colon cell nucleus, where "a nuclei photo of [colon]" is a medical text template for the colon, and T... i stomach Let i be the image of the i-th gastric cell nucleus, and let "a nuclei photo of [stomach]" be the medical information template for the stomach. b) Extract the i-th brain cell nucleus image T from the cell nucleus image-text dataset T. i brain Medical prompts for the brain: a nuclei photo of [brain], image of the i-th kidney cell nucleus T i kidney Medical cues for the kidneys: a nuclei photo of [kidney], image of the i-th liver cell nucleus (T). i liver Medical cues related to the liver: photo of [liver], image of the i-th breast cell nucleus (T). i breast Medical prompts related to breast tissue, including a photo of the i-th colon cell nucleus (T). i colon Medical cues related to the kidneys: anuclei photo of [colon], image of the i-th gastric cell nucleus T i stomach The medical information template for the stomach, a nucleiphoto of [stomach], is input into the clip model to obtain the optimized clip model; c) Obtain a brain cell nucleus images, b kidney cell nucleus images, c liver cell nucleus images, d breast cell nucleus images, e colon cell nucleus images, and f stomach cell nucleus images from the NuInsSeg multi-organ category dataset. a+b+c+d+e+f=n, resulting in the cell nucleus image-text dataset Y. Where Y i brain For the i-th brain cell nucleus image, Y i kidney For the i-th kidney cell nucleus image, Y i liver For the i-th liver cell nucleus image, Y i breast For the i-th image of a breast cell nucleus, Y i colon For the i-th colon cell nucleus image, Y i stomach This is the image of the i-th gastric cell nucleus; d) Divide the cell nucleus image-text dataset Y into a training set and a test set, and scale the cell nucleus images in the training set to 572×572; e) Construct a segmentation network model, which consists of a text module, an image module, and an MLP module; f) Input the medical prompt text templates from the training set into the text module and output the text vector; g) Input the cell nucleus images from the training set into the image module of the segmentation network model, and output the cell nucleus segmentation result image and feature vector; h) Input the feature vector into the MLP module of the segmentation network model and output the parameters; i) Update the segmentation network model by updating the parameters to obtain the updated segmentation network model; j) Train the updated segmentation network model to obtain the optimized segmentation network model; k) Input the cell nucleus images from the test set into the optimized segmentation network model, and output the final segmentation result image.

2. The multi-organ cell nucleus segmentation method based on cue learning according to claim 1, characterized in that: N takes the value 500.

3. The multi-organ cell nucleus segmentation method based on cue learning according to claim 1, characterized in that: In step a), collect N brain cell nucleus images from the MoNuSeg dataset and / or the CMP-15 dataset and / or the CMP-17 dataset and / or the CryonNuSeg dataset and / or the NuInsSeg dataset; collect N kidney cell nucleus images from the MoNuSeg dataset and / or the kumor dataset and / or the Irshad dataset and / or the CryonNuSeg dataset and / or the MoNuSAC dataset and / or the NuInsSeg dataset; and collect N liver images from the MoNuSeg dataset and / or the CryonNuSeg dataset and / or the Crowedsource dataset and / or the kumor dataset and / or the NuInsSeg dataset. Collect N mammary cell nucleus images from the MoNuSeg dataset and / or BCNup dataset and / or MoNuSAC dataset and / or Nucls dataset and / or TNBC dataset and / or JanowczykmCryonNuSeg dataset and / or Gelasca dataset and / or Naylor dataset and / or MoNuSAC dataset and / or NulunsSeg dataset and / or kumor dataset; collect N colon cell nucleus images from the CoNsep dataset and / or CRCHisto dataset and / or CryonNuSeg dataset and / or NuInsSeg dataset and / or kumor dataset; collect N gastric cell nucleus images from the MoNuSeg CryonNuSeg dataset and / or Wienert dataset and / or NuInsSeg dataset and / or kumor dataset.

4. The multi-organ cell nucleus segmentation method based on cue learning according to claim 1, characterized in that: In step d), the cell nucleus image-text dataset Y is divided into a training set and a test set in a 7:3 ratio.

5. The multi-organ cell nucleus segmentation method based on cue learning according to claim 1, characterized in that, Step f) includes the following steps: The text module of the f-1) segmentation network model is composed of an optimized clip model; f-2) Input the medical prompt text template of the brain from the training set (a nuclei photo of [brain]) into the text module, and output the text vector N. brain N brain ∈R L×N Let R be the real number space, L be the text length, and N be the length of the last word in the text. Then, let the text vector N be... brain The text vector N′ is obtained by expanding the channel dimension using the torch.unsqueeze function in Python's PyTorch library. brain ; f-3) Input the medical prompt text template for the kidney from the training set, such as "a nuclei photo of [kidney]", into the text module, and output the text vector N. kidney N kidney ∈R L×N , the text vector N kidney The text vector N′ is obtained by expanding the channel dimension using the torch.unsqueeze function in Python's PyTorch library. kidney ; f-4) Input the medical prompt text template of liver (a nuclei photo of [liver]) from the training set into the text module, and output the text vector N. liver N liver ∈R L×N , the text vector N liver The text vector N′ is obtained by expanding the channel dimension using the torch.unsqueeze function in Python's PyTorch library. liver ; f-5) Input the medical prompt text template for breast tissue from the training set, such as "a nuclei photo of [breast]", into the text module, and output the text vector N. breast N breast ∈R L×N , the text vector N breast The text vector N′ is obtained by expanding the channel dimension using the torch.unsqueeze function in Python's PyTorch library. breast ; f-6) Input the medical prompt text template for the kidney from the training set, "a nuclei photo of [colon]", into the text module, and output the text vector N. colon N colon ∈R L×N , the text vector N colon The text vector N′ is obtained by expanding the channel dimension using the torch.unsqueeze function in Python's PyTorch library. colon ; f-7) Input the medical prompt text template for the stomach from the training set, "a nuclei photo of [stomach]", into the text module, and output the text vector N. stomach N stomach ∈R L×N , the text vector N stomach The text vector N′ is obtained by expanding the channel dimension using the torch.unsqueeze function in Python's PyTorch library. stomach .

6. The multi-organ cell nucleus segmentation method based on cue learning according to claim 5, characterized in that, Step g) includes the following steps: The image module of the g-1 segmentation network model consists of an image encoder, an image decoder, and a GAP module; The image encoder of the g-2) image module consists of a first CRM module, a second CRM module, a third CRM module, a fourth CRM module, and a fifth CRM module. The first, second, third, and fourth CRM modules each consist of a first convolutional layer, a first ReLU activation function, a second convolutional layer, a second ReLU activation function, and a max-pooling layer, respectively. The fifth CRM module consists of a first convolutional layer, a first ReLU activation function, a second convolutional layer, and a second ReLU activation function, respectively. The image encoder of the i-th brain cell nucleus image Y from the training set is used as the input. i brain The input is given to the first CRM module, and the output is the feature. Features The input is fed into the second CRM module, and the output is the feature. Features The input is fed into the third CRM module, and the output is the feature. Features The input is fed into the fourth CRM module, and the output is the feature. Features The input is fed into the fifth CRM module, and the output is the feature. The i-th kidney cell nucleus image Y in the training set i kidney The input is given to the first CRM module, and the output is the feature. Features The input is fed into the second CRM module, and the output is the feature. Features The input is fed into the third CRM module, and the output is the feature. Features The input is fed into the fourth CRM module, and the output is the feature. Features The input is fed into the fifth CRM module, and the output is the feature. The i-th liver cell nucleus image Y in the training set i liver The input is given to the first CRM module, and the output is the feature. Features The input is fed into the second CRM module, and the output is the feature. Features The input is fed into the third CRM module, and the output is the feature. Features The input is fed into the fourth CRM module, and the output is the feature. Features The input is fed into the fifth CRM module, and the output is the feature. The i-th breast cell nucleus image Y in the training set i breast The input is given to the first CRM module, and the output is the feature. Features The input is fed into the second CRM module, and the output is the feature. Features The input is fed into the third CRM module, and the output is the feature. Features The input is fed into the fourth CRM module, and the output is the feature. Features The input is fed into the fifth CRM module, and the output is the feature. The i-th colon cell nucleus image Y in the training set i colon The input is given to the first CRM module, and the output is the feature. Features The input is fed into the second CRM module, and the output is the feature. Features The input is fed into the third CRM module, and the output is the feature. Features The input is fed into the fourth CRM module, and the output is the feature. Features The input is fed into the fifth CRM module, and the output is the feature. The i-th gastric cell nucleus image Y in the training set i stomach The input is given to the first CRM module, and the output is the feature. Features The input is fed into the second CRM module, and the output is the feature. Features The input is fed into the third CRM module, and the output is the feature. Features The input is fed into the fourth CRM module, and the output is the feature. Features The input is fed into the fifth CRM module, and the output is the feature. The image decoder of the g-3 image module consists of a first GRU module, a second GRU module, a third GRU module, and a fourth GRU module. Each of these modules comprises a first convolutional layer, a first ReLU activation function, a second convolutional layer, a second ReLU activation function, and an upsampling layer, sequentially processing the features. The input is given to the first CRM module, and the output is the feature. Features The input is fed into the second CRM module, and the output is the feature. Features The input is fed into the third CRM module, and the output is the feature. Features The input is fed into the fourth CRM module, and the output is a brain cell nucleus segmentation result image. Features The input is given to the first CRM module, and the output is the feature. Features The input is fed into the second CRM module, and the output is the feature. Features The input is fed into the third CRM module, and the output is the feature. Features The input is fed into the fourth CRM module, and the output is a segmentation result image of kidney cell nuclei. Features The input is given to the first CRM module, and the output is the feature. Features The input is fed into the second CRM module, and the output is the feature. Features The input is fed into the third CRM module, and the output is the feature. Features The input is fed into the fourth CRM module, which outputs a liver cell nucleus segmentation result image. Features The input is given to the first CRM module, and the output is the feature. Features The input is fed into the second CRM module, and the output is the feature. Features The input is fed into the third CRM module, and the output is the feature. Features The input is fed into the fourth CRM module, and the output is a segmentation result image of breast cell nuclei. Features The input is given to the first CRM module, and the output is the feature. Features The input is fed into the second CRM module, and the output is the feature. Features The input is fed into the third CRM module, and the output is the feature. Features The data is input into the fourth CRM module, and the output is a segmentation result image of colon cell nuclei. Features The input is given to the first CRM module, and the output is the feature. Features The input is fed into the second CRM module, and the output is the feature. Features The input is fed into the third CRM module, and the output is the feature. Features The input is fed into the fourth CRM module, and the output is a segmentation result image of the gastric cell nuclei. The GAP module of the g-4 image module consists of a Batch Normalization (BN) layer, a ReLU activation function, and an adaptive average pooling layer, which integrates features. The input is fed into the GAP module and the output is the feature PG. brain , will feature The input is fed into the GAP module and the output is the feature PG. kidney , will feature The input is fed into the GAP module and the output is the feature PG. liver , will feature The input is fed into the GAP module and the output is the feature PG. breast , will feature The input is fed into the GAP module and the output is the feature PG. colon , will feature The input is fed into the GAP module and the output is the feature PG. stomach , will feature PG brain With text vector N′ brain Perform a concatenation operation to obtain the feature vector. Feature PG kidney With text vector N′ kidney Perform a concatenation operation to obtain the feature vector. PG liver With text vector N′ liver Perform a concatenation operation to obtain the feature vector. Feature PG breast With text vector N′ breast Perform a concatenation operation to obtain the feature vector. Feature PG colon With text vector N′ colon Perform a concatenation operation to obtain the feature vector. Feature PG stomach With text vector N′ stomach Perform a concatenation operation to obtain the feature vector.

7. The multi-organ cell nucleus segmentation method based on cue learning according to claim 6, characterized in that, Step h) includes the following steps: The MLP module of the segmentation network model (h-1) consists of a first convolutional layer, a second convolutional layer, and a third convolutional layer, with the kernel size of the first, second, and third convolutional layers all being 1*1. h-2) eigenvectors The input is fed into the MLP module, and the output is the feature. eigenvectors The input is fed into the MLP module, and the output is the feature. eigenvectors The input is fed into the MLP module, and the output is the feature. eigenvectors The input is fed into the MLP module, and the output is the feature. eigenvectors The input is fed into the MLP module, and the output is the feature. eigenvectors The input is fed into the MLP module, and the output is the feature. h-3) features The input is given to the Sigmoid function, and the output is the parameter. Features The input is given to the Sigmoid function, and the output is the parameter. Features The input is given to the Sigmoid function, and the output is the parameter. Features The input is given to the Sigmoid function, and the output is the parameter. Features The input is given to the Sigmoid function, and the output is the parameter. Features The input is given to the Sigmoid function, and the output is the parameter.

8. The multi-organ cell nucleus segmentation method based on cue learning according to claim 7, characterized in that, Step i) includes the following steps: i-1) via parameters Using the `reshape` function from the PyTorch library in Python, the first and second convolutional layers of the first, second, third, fourth, and fifth CRM modules of the image encoder are reshaped, followed by sequential convolution and ReLU activation function operations to complete the first update of the image encoder. The first update of the image decoder is completed by using the reshape function in the PyTorch library of Python to reshape the first convolutional layer and the second convolutional layer of the first GRU module, the second GRU module, the third GRU module and the fourth GRU module. Then, convolution and ReLU activation function operations are performed in sequence. i-2) via parameters Using the `reshape` function from the PyTorch library in Python, the first and second convolutional layers of the first, second, third, fourth, and fifth CRM modules of the image encoder are reshaped, followed by sequential convolution and ReLU activation function operations to complete the second update of the image encoder. The first convolutional layer and the second convolutional layer of the first GRU module, the second GRU module, the third GRU module, and the fourth GRU module of the image decoder are reshaped using the reshape function in the PyTorch library of Python. Then, convolution and ReLU activation function operations are performed in sequence to complete the second update of the image decoder. i-3) via parameters Using the `reshape` function from the PyTorch library in Python, the first and second convolutional layers of the first, second, third, fourth, and fifth CRM modules of the image encoder are reshaped, followed by convolution and ReLU activation function operations, completing the third update of the image encoder. The first convolutional layer and the second convolutional layer of the first GRU module, the second GRU module, the third GRU module, and the fourth GRU module of the image decoder are reshaped using the reshape function in the PyTorch library of Python. Then, convolution and ReLU activation function operations are performed in sequence to complete the third update of the image decoder. i-4) via parameters Using the `reshape` function from the PyTorch library in Python, the first and second convolutional layers of the first, second, third, fourth, and fifth CRM modules of the image encoder are reshaped, followed by sequential convolution and ReLU activation function operations, completing the fourth update of the image encoder. The first convolutional layer and the second convolutional layer of the first GRU module, the second GRU module, the third GRU module and the fourth GRU module of the image decoder are reshaped using the reshape function in the PyTorch library of Python. Then, convolution and ReLU activation function operations are performed in sequence to complete the fourth update of the image decoder. i-5) via parameters Using the `reshape` function from the PyTorch library in Python, the first and second convolutional layers of the first, second, third, fourth, and fifth CRM modules of the image encoder are reshaped, followed by sequential convolution and ReLU activation function operations to complete the fifth update of the image encoder. Using the reshape function in the PyTorch library of Python, the first convolutional layer and the second convolutional layer of the first GRU module, the second GRU module, the third GRU module and the fourth GRU module of the image decoder are reshaped and then convolution and ReLU activation function operations are performed in sequence to complete the fifth update of the image decoder. i-6) via parameters Using the `reshape` function from the PyTorch library in Python, the first and second convolutional layers of the first, second, third, fourth, and fifth CRM modules of the image encoder are reshaped, followed by sequential convolution and ReLU activation function operations, completing the sixth update of the image encoder. The reshape function in the PyTorch library of Python is used to reshape the first and second convolutional layers of the first, second, third, and fourth GRU modules of the image decoder. Then, convolution and ReLU activation functions are performed sequentially to complete the sixth update of the image decoder and obtain the updated segmentation network model.

9. The multi-organ cell nucleus segmentation method based on cue learning according to claim 1, characterized in that: In step j), the Adam optimizer is used to train the updated segmentation network model using the DSC loss function to obtain the optimized segmentation network model.

10. The multi-organ cell nucleus segmentation method based on cue learning according to claim 6, characterized in that, Step k) includes the following steps: k-1) The i-th brain cell nucleus image Y in the training set i brain The images are sequentially input into the image encoder and image decoder of the optimized segmentation network model, and the final segmentation result image of the brain cell nucleus is output. k-2) The i-th kidney cell nucleus image Y in the training set i kidney The images are sequentially input into the image encoder and image decoder of the optimized segmentation network model, and the final segmentation result image of the kidney cell nucleus is output. k-3) The i-th liver cell nucleus image Y in the training set i liver The images are sequentially input into the image encoder and image decoder of the optimized segmentation network model, and the final segmentation result of the liver cell nucleus is output. k-4) The i-th breast cell nucleus image Y in the training set i breast The images are sequentially input into the image encoder and image decoder of the optimized segmentation network model, and the final segmentation result image of the breast cell nucleus is output. k-5) The i-th colon cell nucleus image Y in the training set i colon The images are sequentially input into the image encoder and image decoder of the optimized segmentation network model, and the final segmentation result of the colon cell nucleus is output. k-6) The i-th gastric cell nucleus image Y in the training set i stomach The images are sequentially input into the image encoder and image decoder of the optimized segmentation network model, and the final segmentation result of the gastric cell nucleus is output.

Images