A microsatellite instability detection system based on colorectal pathological images

Abstract

The invention discloses a microsatellite instability detection system based on colorectal pathological images, wherein the image acquisition module is used to acquire the colorectal pathological images of the detected user; the tumor tissue detection module is used to cut the image into a plurality of image blocks and detect whether each image block is a tumor image; the microsatellite instability detection module uses a microsatellite instability classification model to identify whether each tumor image belongs to high microsatellite instability; when training the tumor tissue detection model, the pre-training weight is used for initialization, the top classifier and the auxiliary classifier are trained at a first training rate, and the parameters of other layers of the tumor tissue detection model are frozen at the same time until the tumor tissue detection model converges, all the parameters of the tumor tissue detection model are unfrozen, and the tumor tissue detection model is fitted at a second training rate. The invention is reliable, widely usable and cost-effective.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image processing, and in particular to a microsatellite instability detection system based on colorectal pathological images. Background Art

[0002] The statements in this section merely mention background art related to the present invention and do not necessarily constitute prior art.

[0003] Studies have shown that colorectal cancer is the third most common cancer and the second most deadly cancer in the world today, with approximately 900,000 people dying from colorectal cancer each year. Microsatellite instability-high (MSI-H) refers to a highly variable phenotype in DNA mismatch repair deficiency (dMMR), which makes the tumor sensitive to immune checkpoint blockade therapy. MSI-H is also a hallmark of hereditary non-polyposis colorectal cancer, namely Lynch syndrome (LS), and is a favorable prognostic factor and a negative predictor for stage II colorectal cancer patients receiving fluorouracil-based adjuvant chemotherapy. Therefore, it is recommended that all colorectal cancer patients undergo polymerase chain reaction (PCR) or immunohistochemistry (IHC) testing of MSI-H and dMMR.

[0004] However, due to the high cost, long time consumption and the lack of testing conditions in small hospitals, a large number of colorectal cancer patients do not have the opportunity to undergo microsatellite instability (MSI) testing. In addition, molecular testing and other existing MSI detection methods, such as the MsPath (Microsatellite instability by Pathology) method and the PREDICT (Pathological RolE in the Determination of Instability in Colorectal Tumors) method, have differences in sensitivity and specificity, and may therefore produce inconsistent results. Summary of the invention

[0005] In order to address the deficiencies of the prior art, the present invention provides a microsatellite instability detection system based on colorectal pathology images, which is reliable, widely applicable and cost-effective.

[0006] In a first aspect, the present invention provides a microsatellite instability detection system based on colorectal pathological images;

[0007] A microsatellite instability detection system based on colorectal pathological images, comprising:

[0008] An image acquisition module, used to acquire colorectal pathological images of the user being tested;

[0009] The tumor tissue detection module is used to cut the colorectal pathology image into several image blocks and use the tumor tissue detection model to detect whether each image block is a tumor image;

[0010] A microsatellite instability detection module is used to identify whether each tumor image belongs to high microsatellite instability by using a microsatellite instability classification model, and to use the proportion of tumor images belonging to high microsatellite instability to all tumor images as the probability that the tested user belongs to high microsatellite instability;

[0011] Among them, when training the tumor tissue detection model, first, the pre-trained weights are used for initialization, and the top classifier and the auxiliary classifier are trained at the first training rate, and the parameters of other layers of the tumor tissue detection model are frozen until the tumor tissue detection model converges; then, all parameters of the tumor tissue detection model are unfrozen, and the tumor tissue detection model is fitted at the second training rate.

[0012] Furthermore, the backbone network of the tumor tissue detection model adopts a lightweight neural network.

[0013] Furthermore, the loss function of the tumor tissue detection model is a weighted sum of the loss function of the top classifier and the loss function of the auxiliary classifier.

[0014] Furthermore, the microsatellite instability classification model includes two deep convolutional neural networks and a collaborative network;

[0015] When training the microsatellite instability classification model, a pair of randomly matched tumor images and their category labels are respectively input into two deep convolutional neural networks, and the deep image features output by the penultimate fully connected layers of the two deep convolutional neural networks are spliced ​​and input into the collaborative network together with the collaborative labels.

[0016] Furthermore, the loss function of the microsatellite instability classification model is the sum of the loss functions of the two deep convolutional neural networks and the loss function of the collaborative network.

[0017] Furthermore, the loss functions of the two deep convolutional neural networks both adopt a binary cross entropy function.

[0018] Furthermore, the result of whether each tumor image output by the microsatellite instability detection module belongs to high microsatellite instability is obtained according to the probability that each tumor image output by the two deep convolutional neural networks belongs to high microsatellite instability.

[0019] Furthermore, the first training rate is higher than the second training rate.

[0020] In a second aspect, the present invention further provides an electronic device, comprising:

[0021] a memory for non-transitory storage of computer-readable instructions; and

[0022] a processor for executing the computer readable instructions,

[0023] When the computer readable instructions are executed by the processor, the following steps are performed:

[0024] Acquire colorectal pathological images of the tested user;

[0025] The colorectal pathology image is cut into several image blocks, and the tumor tissue detection model is used to detect whether each image block is a tumor image;

[0026] The microsatellite instability detection module is used to use the microsatellite instability classification model to identify whether each tumor image belongs to high microsatellite instability, and use the proportion of tumor images belonging to high microsatellite instability to all tumor images as the probability that the tested user belongs to high microsatellite instability;

[0027] Among them, when training the tumor tissue detection model, first, the pre-trained weights are used for initialization, and the top classifier and the auxiliary classifier are trained at the first training rate, and the parameters of other layers of the tumor tissue detection model are frozen until the tumor tissue detection model converges; then, all parameters of the tumor tissue detection model are unfrozen, and the tumor tissue detection model is fitted at the second training rate.

[0028] In a third aspect, the present invention further provides a storage medium that non-temporarily stores computer-readable instructions, wherein when the non-temporary computer-readable instructions are executed by a computer, the following steps are performed:

[0029] Acquire colorectal pathological images of the tested user;

[0030] The colorectal pathology image is cut into several image blocks, and the tumor tissue detection model is used to detect whether each image block is a tumor image;

[0031] The microsatellite instability detection module is used to use the microsatellite instability classification model to identify whether each tumor image belongs to high microsatellite instability, and use the proportion of tumor images belonging to high microsatellite instability to all tumor images as the probability that the tested user belongs to high microsatellite instability;

[0032] Among them, when training the tumor tissue detection model, first, the pre-trained weights are used for initialization, and the top classifier and the auxiliary classifier are trained at the first training rate, and the parameters of other layers of the tumor tissue detection model are frozen until the tumor tissue detection model converges; then, all parameters of the tumor tissue detection model are unfrozen, and the tumor tissue detection model is fitted at the second training rate.

[0033] Compared with the prior art, the present invention has the following beneficial effects:

[0034] The system of the present invention trains the top classifier and the auxiliary classifier at a first training rate, and freezes the parameters of other layers of the tumor tissue detection model at the same time. After the tumor tissue detection model converges, all parameters of the tumor tissue detection model are unfrozen, and the tumor tissue detection model is fitted at a second training rate, thereby solving the gradient vanishing problem encountered when the tumor tissue detection model uses a pre-trained model.

[0035] The system of the present invention can reduce the problems of large parameter quantity, heavy computing memory load, difficulty in training, etc. caused by non-sharing of parameters in multiple parallel networks in terms of its microsatellite instability classification model; on the other hand, the ability of the model to extract low-dimensional features can be enhanced by sharing the bottom-layer parameters of the two networks, while retaining the ability of higher-level network branches to extract different high-dimensional features. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0037] Figure 1 This is a structural diagram of a microsatellite instability detection system based on colorectal pathological images in Example 1;

[0038] Figure 2 is a structural diagram of the tumor tissue detection model of Example 1;

[0039] Figure 3 It is a structural diagram of the microsatellite instability classification model of Example 1;

[0040] Figure 4 Schematic diagram of training and verification accuracy of three different methods based on the MobileNet network structure of Example 1;

[0041] Figure 5 A schematic diagram of training and verification AUC of three different methods based on the MobileNet network structure in Example 1;

[0042] Figure 6 Schematic diagram of training and verification accuracy of three different methods based on the Inception network structure in Example 1;

[0043] Figure 7 Schematic diagram of training and verification AUC of three different methods based on the Inception network structure in Example 1. DETAILED DESCRIPTION

[0044] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0045] In the absence of conflict, the embodiments of the present invention and the features of the embodiments may be combined with each other.

[0046] In this embodiment, all data is obtained in compliance with laws and regulations and based on the user's consent, and is used legally.

[0047] Embodiment 1

[0048] This embodiment provides a microsatellite instability detection system based on colorectal pathological images;

[0049] like Figure 1 As shown, a microsatellite instability detection system based on colorectal pathological images includes: an image acquisition module, a tumor tissue detection module and a microsatellite instability detection module connected in sequence.

[0050] The image acquisition module is used to acquire the colorectal pathological image (colorectal cancer tissue pathological image) of the user being tested.

[0051] Specifically, colorectal pathology images are full-field digital slices.

[0052] The tumor tissue detection module is used to cut the user's colorectal pathology image into several image blocks, and use the tumor tissue detection model to detect whether each image block is a tumor image.

[0053] In medical image processing, due to the difficulties in data collection and annotation, medical image data sets are usually small, and overfitting often occurs when using machine learning and deep learning techniques. In addition to increasing the amount of data through data enhancement and other methods, using a pre-trained model trained on a large-scale data set is an effective way to solve overfitting. However, deep convolutional neural networks often use a large number of convolutional layers, so during training, they face challenges such as weakened back propagation and slow model convergence caused by gradient vanishing, which ultimately leads to reduced model classification performance. Therefore, the present invention proposes a fine-tuning model based on deep supervision to solve the gradient vanishing problem encountered when the tumor tissue detection module uses a pre-trained model.

[0054] like Figure 2 As shown in the figure, the proposed tumor tissue detection model is a fine-tuning model based on deep supervision, which mainly consists of the following four parts: input layer, backbone network, a top classifier and two auxiliary supervised classifiers. Taking the lightweight neural network (Efficient Convolutional Neural Networks for Mobile Vision Applications, MobileNet) as an example, the 13 depthwise convolution blocks of MobileNet constitute the backbone network. A fully connected layer containing 1024 units and a top sigmoid binary classifier are added after the 13th depthwise convolution block. In the middle hidden layer, that is, after the 7th and 10th depthwise convolution blocks, a fully connected layer containing 1024 units and an auxiliary sigmoid binary classifier are added.

[0055] Unlike the method of directly using the pre-trained model for fine-tuning, the pre-trained model is first modified. In addition to matching the classifier at the top of the model to the binary classifier for tumor tissue detection, two auxiliary classifiers are added to the middle hidden layer of the pre-trained model to enhance supervision of the model and reduce the gradient vanishing problem.

[0056] First, the pre-trained model is initialized with the pre-trained weights, and the top classifier and two auxiliary classifiers are trained at the first training rate (0.01), while freezing the parameters of other layers until the model converges. Then, all parameters of the network are unfrozen and the model is fitted at the second training rate (0.0001). The first training rate is higher than the second training rate.

[0057] The present invention emphasizes not directly using the pre-trained model trained on a large-scale dataset for fine-tuning, but introducing a deep supervision layer from the hidden layer in the middle of the model during fine-tuning, and reducing the gradient vanishing problem by adding additional classifiers. In the model proposed by the present invention, the loss function of the top classifier of the model is:

[0058]

[0059] Among them, y represents the category label, y = 1 represents tumor tissue, and y = 0 represents non-tumor tissue. represents the predicted probability. The loss function of the i-th auxiliary classifier is defined as:

[0060]

[0061] in, Represents the predicted probability of the i-th auxiliary classifier. The loss function of the tumor tissue detection model is the weighted sum of the loss function of the top classifier and the loss function of the auxiliary classifier. Therefore, the total loss function of the tumor tissue detection model is defined as follows:

[0062]

[0063] Among them, α and β as weight adjustment parameters can be set to any values. In the present invention, their values ​​are set to 0.2 and 0.3 respectively, and decay by 10% after every 10 iterations of training.

[0064] The microsatellite instability detection module is used to input each tumor image into the microsatellite instability classification model (MSI classification model) for classification, to determine whether each tumor image belongs to high microsatellite instability (MSI-H), and to use the proportion of tumor images belonging to MSI-H to all tumor images as the predicted probability that the tested user belongs to MSI-H.

[0065] like Figure 3 As shown, the MSI classification model proposed in the present invention is mainly composed of three parts: an image paired input layer, two completely identical deep convolutional neural networks (DCNN), namely, a first deep convolutional neural network (DCNN1) and a second deep convolutional neural network (DCNN2), and a collaborative network.

[0066] Taking the MobileNet architecture as an example, the 13 depthwise convolution blocks of MobileNet are used as the DCNN in this model. After the 13th depthwise convolution block, a fully connected layer with 1024 units and a top sigmoid binary classifier are added to predict MSI-H. The collaborative network includes an input layer, a concatenation layer, a fully connected learning layer, and a sigmoid binary classification layer. Specifically, the collaborative network uses the features output by the fully connected layers of DCNN1 and DCNN2 as input. The two features are concatenated together after the concatenation operation, input into a fully connected layer with 2048 units for learning, and then output through the sigmoid binary classification layer, thereby predicting whether the two inputs of DCNN1 and DCNN2 belong to the same category.

[0067] When training the microsatellite instability classification model, a pair of randomly matched tumor images and their category labels are respectively input into two deep convolutional neural networks, and the deep image features output by the penultimate fully connected layers of the two deep convolutional neural networks are spliced ​​and input into the collaborative network. Specifically, the MSI classification model takes a pair of randomly matched tumor images detected in the first stage as input, and each tumor image and its corresponding category label are respectively input into DCNN1 and DCNN2 for learning to determine the category to which the image belongs (MSI-H or MSI-L / MSS (MSI-L means low-frequency microsatellite instability, the full name is microsatellite instability-low; MSS means microsatellite stability, the full name is microsatellitestable)). Let (x1,x2) be the input of (DCNN1,DCNN2), and the output of the penultimate fully connected layer of each DCNN is defined as the extracted deep image feature, which can be obtained by the following forward calculation:

[0068]

[0069]

[0070] in, represents the forward calculation, θ s represents the parameters shared between DCNN1 and DCNN2 (taking the MobileNet architecture as an example, the 13 depthwise convolution blocks of MobileNet are used as the DCNN in this model. Here, the first 4 depthwise convolution blocks of MobileNet are used as shared modules, and other modules are not shared). i It represents the independent and unshared parameters of DCNNi, i∈(1,2). f1 and f2 are concatenated to form feature fc , feature f c The collaborative network is fed with its corresponding collaborative label to learn whether the input image pair (x1, x2) belongs to the same category (MSI-H or MSI-L / MSS). The collaborative label and collaborative loss function are defined as follows:

[0071]

[0072]

[0073] When the collaborative label y s = 1, indicating that x1 and x2 belong to the same category. s = 0 means x1 and x2 belong to different categories. The loss function of the microsatellite instability classification model is the sum of the loss functions of the two deep convolutional neural networks and the loss function of the collaborative network:

[0074]

[0075] in, and Represent the binary cross entropy functions of DCNN1 and DCNN2 respectively.

[0076] The result of whether each tumor image output by the microsatellite instability detection module belongs to high microsatellite instability is obtained according to the probability of each tumor image output by the two deep convolutional neural networks belonging to high microsatellite instability. Specifically, during the test, the prediction probabilities given by the two DCNNs are expressed as represents the i-th DCNN, K represents the K-th input image; the probability that the model classifies the image as MSI-H is:

[0077]

[0078] If the probability value of the model classifying the image as MSI-H is greater than the threshold, the tumor image belongs to high microsatellite instability.

[0079] The partial parameter sharing strategy proposed by the MSI classification model can, on the one hand, reduce the problems of large parameter amount, heavy computational memory load, and difficulty in training caused by non-parameter sharing of multiple parallel networks; on the other hand, it can enhance the model's ability to extract low-dimensional features by sharing the underlying parameters of the two networks (DCNN1 and DCNN2), while retaining the ability of higher-level network branches to extract different high-dimensional features.

[0080] Enhancing the ability to extract low-dimensional features can reduce the interference of intra-class differences in tumor images on classification, and the different high-dimensional features extracted can lead to certain differences between the two models. In this way, when one of the networks has a large deviation in prediction, the other network can correct it through the collaborative network, thereby prompting the two models to learn collaboratively, stimulate each other, and improve the model's resolution ability.

[0081] In order to verify the present invention, a total of 144 cases were collected from three hospitals, each with a full-field digital slice and its MSI-H status was known. Finally, images of 124 cases (66 MSI-H and 58 MSI-L / MSS) from the same hospital were divided into training and test sets in a ratio of 7:3, and 20 cases (15 MSI-H and 5 MSI-L / MSS) from the other two hospitals were used as test sets.

[0082] Professional pathologists manually annotated the tumor and non-tumor tissue areas on the full-view digital slides, cropped the annotated full-view digital slides into small images of 224 pixels × 224 pixels, and uniformly performed standardized preprocessing. The specific division of the data is shown in Table 1.

[0083] Table 1. Data division and usage

[0084]

[0085] In Table 1, a represents the number of cases, b Represents the number of tumor images. Data A comes from the same hospital, and data B and data C come from two other hospitals respectively.

[0086] The proposed fine-tuning model based on deep supervision is verified on four different deep convolutional neural network structures, namely, Deep residual network (ResNet), Efficient Convolutional Neural Networks for Mobile Vision Applications (MobileNet), Inception V3 and EfficientNet B0, and is compared with the accuracy (Acc) and area under the ROC curve (AUC) of not using a pre-trained model and using a fine-tuning method. The experimental results are shown in Tables 2, 3 and 4.

[0087] Table 2. Tumor detection results without using pre-trained models

[0088]

[0089]

[0090] Table 3. Tumor detection results using fine-tuning

[0091]

[0092] Table 4 Tumor detection results of the system proposed in the present invention.

[0093]

[0094] The experimental results show that on the validation set, both fine-tuning (Table 3) and fine-tuning based on deep supervision (Table 4) have higher accuracy than the model without pre-training, while fine-tuning based on deep supervision is better than fine-tuning, and the model based on the MobileNet network structure has the best performance in tumor tissue detection. In addition, the AUC of the fine-tuning model based on deep supervision is also higher than the fine-tuning method overall, although slightly lower than the method without pre-training model.

[0095] Taking the MobileNet network structure as an example, Figure 4 and Figure 5 The accuracy and AUC curves of training and validation without using a pre-trained model (From Scratch), using fine-tuning (Fine-tuning), and using a fine-tuned model based on deep supervision (Proposed) are shown.

[0096] The present invention also externally validated the proposed system on data from two other hospitals (data B and data C). Compared with the data used for training and validation, the test data came from other hospitals and were prepared under different conditions, so the staining conditions showed diversity, which posed a great challenge to the stability of the algorithm. Nevertheless, the experimental results show that the fine-tuning based on deep supervision proposed by the present invention is almost the best among the three methods.

[0097] In order to verify the performance of the proposed MSI classification model, it was also verified on four different network structures, ResNet, MobileNet, Inception and EfficientNet, and compared with other methods. The experimental results are shown in Table 5 (taking InceptionV3 as an example, InceptionV3 is the original baseline network architecture, Inception-SDL is a collaborative learning model based on the InceptionV3 structure, and InceptionV3-Proposed is a parameter partial sharing model based on the InceptionV3 structure proposed in this invention, that is, PPsNet based on the InceptionV3 structure).

[0098] As can be seen from Table 5, the accuracy and AUC of the system proposed in the present invention are the highest on the validation set compared with the other two methods, and the improvement based on the InceptionV3 network structure is the best in the performance of MSI classification. Compared with the current advanced synergic deep learning (SDL) method, a large number of parameters are reduced. In addition, on data B and data C, the accuracy and AUC of the method proposed in the present invention are also the highest in most cases.

[0099] Taking the Inception network structure as an example, Figure 6 and Figure 7 The training and validation accuracy and AUC curves of the three methods are shown.

[0100] Table 5. Comparison of the accuracy and AUC results of the proposed system and other methods in MSI classification

[0101]

[0102] After experiments, it was finally decided to use MobileNet, which was fine-tuned based on deep supervision, as the tumor tissue detection model.

[0103] Table 6 shows the comparison results of the proposed system and other methods after continuous tumor tissue detection and MSI classification. The experimental results show that after the same tumor detection model, the proposed system, compared with the benchmark model and the advanced SDL method, has the highest accuracy and AUC of the MSI classification results on the validation set among the three methods regardless of the network structure, and the advantage on the test set is also very obvious.

[0104] Table 7. Comparison of the accuracy and AUC results of the system proposed in the present invention and other methods in MSI classification after tumor tissue detection

[0105]

[0106]

[0107] Embodiment 2

[0108] This embodiment further provides an electronic device, including: one or more processors, one or more memories, and one or more computer programs; wherein the processor is connected to the memory, the one or more computer programs are stored in the memory, and when the electronic device is running, the processor executes the one or more computer programs stored in the memory, so that the electronic device performs the following steps:

[0109] Acquire colorectal pathological images of the tested user;

[0110] The colorectal pathology image is cut into several image blocks, and the tumor tissue detection model is used to detect whether each image block is a tumor image;

[0111] The microsatellite instability detection module is used to use the microsatellite instability classification model to identify whether each tumor image belongs to high microsatellite instability, and use the proportion of tumor images belonging to high microsatellite instability to all tumor images as the probability that the tested user belongs to high microsatellite instability;

[0112] Among them, when training the tumor tissue detection model, first, the pre-trained weights are used for initialization, and the top classifier and the auxiliary classifier are trained at the first training rate, and the parameters of other layers of the tumor tissue detection model are frozen until the tumor tissue detection model converges; then, all parameters of the tumor tissue detection model are unfrozen, and the tumor tissue detection model is fitted at the second training rate.

[0113] It should be understood that in this embodiment, the processor may be a central processing unit CPU, and the processor may also be other general-purpose processors, digital signal processors DSP, application-specific integrated circuits ASIC, off-the-shelf programmable gate arrays FPGA or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc.

[0114] The memory may include a read-only memory and a random access memory, and provide instructions and data to the processor. A portion of the memory may also include a non-volatile random access memory. For example, the memory may also store information about the device type.

[0115] In the implementation process, the above steps can be completed by hardware integrated logic circuits in the processor or instructions in software form.

[0116] Embodiment 3

[0117] This embodiment further provides a computer-readable storage medium for storing computer instructions. When the computer instructions are executed by a processor, the following steps are performed:

[0118] Acquire colorectal pathological images of the tested user;

[0119] The colorectal pathology image is cut into several image blocks, and the tumor tissue detection model is used to detect whether each image block is a tumor image;

[0120] The microsatellite instability detection module is used to use the microsatellite instability classification model to identify whether each tumor image belongs to high microsatellite instability, and use the proportion of tumor images belonging to high microsatellite instability to all tumor images as the probability that the tested user belongs to high microsatellite instability;

[0121] Among them, when training the tumor tissue detection model, first, the pre-trained weights are used for initialization, and the top classifier and the auxiliary classifier are trained at the first training rate, and the parameters of other layers of the tumor tissue detection model are frozen until the tumor tissue detection model converges; then, all parameters of the tumor tissue detection model are unfrozen, and the tumor tissue detection model is fitted at the second training rate.

[0122] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A microsatellite instability detection system based on colorectal pathological images, characterized in that: include: An image acquisition module, used to acquire colorectal pathological images of the user being tested; The tumor tissue detection module is used to cut the colorectal pathology image into several image blocks and use the tumor tissue detection model to detect whether each image block is a tumor image; A microsatellite instability detection module is used to identify whether each tumor image belongs to high microsatellite instability by using a microsatellite instability classification model, and to use the proportion of tumor images belonging to high microsatellite instability to all tumor images as the probability that the tested user belongs to high microsatellite instability; When training the tumor tissue detection model, first, the pre-trained weights are used for initialization, and the top classifier and the auxiliary classifier are trained at a first training rate, while the parameters of other layers of the tumor tissue detection model are frozen until the tumor tissue detection model converges; then, all parameters of the tumor tissue detection model are unfrozen, and the tumor tissue detection model is fitted at a second training rate; The microsatellite instability classification model includes two deep convolutional neural networks and a collaborative network; When training the microsatellite instability classification model, a pair of randomly matched tumor images and their category labels are respectively input into two deep convolutional neural networks, and the deep image features output by the penultimate fully connected layers of the two deep convolutional neural networks are spliced ​​and input into the collaborative network together with the collaborative labels; The loss function of the microsatellite instability classification model is the sum of the loss functions of the two deep convolutional neural networks and the loss function of the collaborative network.

2. A microsatellite instability detection system based on colorectal pathological images as claimed in claim 1, characterized in that: The backbone network of the tumor tissue detection model adopts a lightweight neural network.

3. The microsatellite instability detection system based on colorectal pathological images as claimed in claim 1, characterized in that: The loss function of the tumor tissue detection model is a weighted sum of the loss function of the top classifier and the loss function of the auxiliary classifier.

4. The microsatellite instability detection system based on colorectal pathological images as claimed in claim 1, characterized in that: The loss functions of the two deep convolutional neural networks both adopt a binary cross entropy function.

5. The microsatellite instability detection system based on colorectal pathological images as claimed in claim 1, characterized in that: The result of whether each tumor image output by the microsatellite instability detection module belongs to high microsatellite instability is obtained according to the probability that each tumor image output by the two deep convolutional neural networks belongs to high microsatellite instability.

6. The microsatellite instability detection system based on colorectal pathological images as claimed in claim 1, characterized in that: The first training rate is higher than the second training rate.

7. An electronic device, comprising: a memory for non-transitory storage of computer readable instructions; as well as a processor for executing the computer readable instructions, When the computer readable instructions are executed by the processor, the following steps are performed: Acquire colorectal pathological images of the tested user; The colorectal pathology image is cut into several image blocks, and the tumor tissue detection model is used to detect whether each image block is a tumor image; The microsatellite instability classification model is used to identify whether each tumor image belongs to high microsatellite instability, and the proportion of tumor images with high microsatellite instability to all tumor images is used as the probability that the tested user belongs to high microsatellite instability. When training the tumor tissue detection model, first, the pre-trained weights are used for initialization, and the top classifier and the auxiliary classifier are trained at a first training rate, while the parameters of other layers of the tumor tissue detection model are frozen until the tumor tissue detection model converges; then, all parameters of the tumor tissue detection model are unfrozen, and the tumor tissue detection model is fitted at a second training rate; The microsatellite instability classification model includes two deep convolutional neural networks and a collaborative network; When training the microsatellite instability classification model, a pair of randomly matched tumor images and their category labels are respectively input into two deep convolutional neural networks, and the deep image features output by the penultimate fully connected layers of the two deep convolutional neural networks are spliced ​​and input into the collaborative network together with the collaborative labels; The loss function of the microsatellite instability classification model is the sum of the loss functions of the two deep convolutional neural networks and the loss function of the collaborative network.

8. A storage medium characterized by being non-temporary wherein when the non-transitory computer-readable instructions are executed by a computer, the following steps are performed: Acquire colorectal pathological images of the tested user; The colorectal pathology image is cut into several image blocks, and the tumor tissue detection model is used to detect whether each image block is a tumor image; The microsatellite instability classification model is used to identify whether each tumor image belongs to high microsatellite instability, and the proportion of tumor images with high microsatellite instability to all tumor images is used as the probability that the tested user belongs to high microsatellite instability. When training the tumor tissue detection model, first, the pre-trained weights are used for initialization, and the top classifier and the auxiliary classifier are trained at a first training rate, while the parameters of other layers of the tumor tissue detection model are frozen until the tumor tissue detection model converges; then, all parameters of the tumor tissue detection model are unfrozen, and the tumor tissue detection model is fitted at a second training rate; The microsatellite instability classification model includes two deep convolutional neural networks and a collaborative network; When training the microsatellite instability classification model, a pair of randomly matched tumor images and their category labels are respectively input into two deep convolutional neural networks, and the deep image features output by the penultimate fully connected layers of the two deep convolutional neural networks are spliced ​​and input into the collaborative network together with the collaborative labels; The loss function of the microsatellite instability classification model is the sum of the loss functions of the two deep convolutional neural networks and the loss function of the collaborative network.

Images