Colorectal cancer pathological image semi-supervised semantic segmentation method and system for regulating and controlling pseudo labels

Abstract

The invention discloses a colorectal cancer pathological image semi-supervised semantic segmentation method for regulating and controlling a pseudo label, and relates to the field of medical image visual segmentation. The method comprises the following steps: firstly, performing image acquisition and marking on a target pathological region, and dividing the target pathological region into a training set, a verification set and a test set in proportion; in the consistency regularization process of training the teacher model, sorting and reducing the categories of the low-confidence pseudo labels according to the confidence scores; the confidence coefficient of the pseudo-labels is evaluated through the prediction score difference between the final model and the optimal model of the trained teacher network, and the pseudo-labels with the high confidence coefficient are refined; according to the number of categories in the pseudo tag, setting the triggering probability of each sub-data enhancement method, and generating cascaded pixel random change; and finally, constructing a semi-supervised semantic segmentation model for the colorectal cancer pathological image false label. The problem that the semi-supervised semantic segmentation task of the colorectal cancer pathological image is low in precision is solved.

Description

Technical Field

[0001] The present invention relates to the field of image segmentation computer vision technology, and in particular to a method and system for semi-supervised semantic segmentation of colorectal cancer pathology images with regulated pseudo labels. Background Art

[0002] Histopathological images are an important diagnostic basis for diagnosing malignant tumors and selecting treatment plans. Conventional diagnostic methods rely on pathologists to make a diagnosis through a visual microscope. The entire process is very time-consuming, labor-intensive, and highly subjective, and the diagnostic results vary from person to person. The rapidly emerging computational pathology research is expected to achieve objective diagnosis, prediction of treatment response, and identification of new clinically relevant morphological features. However, using large labeled data sets to train neural networks is the main approach in computational pathology, and the acquisition of detailed labeled data is usually time-consuming, labor-intensive, and expensive. The main task of semi-supervised learning is to make full use of unlabeled data to train models under limited supervised data, thereby reducing the cost burden of data annotation.

[0003] Semi-supervised semantic segmentation methods based on pseudo labels have made some progress in histopathology image analysis in recent years, mainly including the following typical methods: Fixed pseudo-label generation method: Some early methods directly use the model's prediction results on unlabeled data as pseudo-labels, and keep these pseudo-labels unchanged during subsequent training. For example, PseudoSeg and Pseudo-Labeling. This type of method is simple and easy to implement, and can effectively use the prediction results of unlabeled data to improve model performance, but it also has some limitations. Because pseudo-labels are static and cannot dynamically adapt to changes in the model, the model is easily misled by incorrect labels in early predictions, accumulating errors and affecting the final segmentation effect.

[0004] Adaptive pseudo-label update method: In order to overcome the shortcomings of static pseudo-labels, some methods introduce adaptive pseudo-label update strategies so that pseudo-labels can be continuously adjusted as the model is updated, such as Mean Teacher and Temporal Ensembling. These methods usually guide the learning of the student model by setting a teacher model, which is the time average result of the student model. Although such methods have improved the quality of pseudo-labels to a certain extent, they rely on a large number of training iterations and consume a lot of computing resources. In addition, the lag in updating the teacher model may cause fluctuations in model performance.

[0005] Uncertainty-based pseudo-label filtering methods: Some methods filter pseudo-labels by uncertainty estimation, thereby removing low-confidence pseudo-labels to reduce the accumulation of errors. For example, methods such as Entropy Minimization and Class-Balanced Self-Training (CBST) filter out unreliable pseudo-labels by setting confidence thresholds. These methods reduce the interference of pseudo-label errors to a certain extent, but for complex samples with high uncertainty, useful information may be lost, resulting in poor performance of the model on difficult-to-classify samples.

[0006] Pseudo-labeling method based on consistency regularization: This method emphasizes the invariance of the model to input perturbations, and improves the stability and accuracy of pseudo-labels by generating consistent pseudo-labels for different versions of the same input (such as data augmentation or different network branches). For example, U2PL and CPS (Cross Pseudo Supervision). The consistency regularization method performs well in semi-supervised segmentation, but there may be a contradiction between the diversity of its branch networks and the consistency of pseudo-labels. In addition, consistency is difficult to guarantee when processing high-noise scenes such as pathological images.

[0007] In general, these pseudo-label-based semi-supervised segmentation methods have their own advantages and disadvantages. The static pseudo-label method is simple but prone to error accumulation. The adaptive update method can reduce static error accumulation but has a high computational overhead. The uncertainty screening method can reduce the interference of erroneous pseudo-labels but may ignore complex samples. The consistency regularization method improves the quality of pseudo-labels, but it is necessary to resolve the trade-off between consistency and diversity. The limitations of these methods still have room for improvement in practical applications, especially in histopathology images. The high complexity and high noise of pathology images place higher requirements on the quality of pseudo-labels. Summary of the invention

[0008] The present invention provides a method and system for semi-supervised semantic segmentation of colorectal cancer pathology images with regulated pseudo-labels, thereby solving the problems of low utilization rate of low-confidence pseudo-labels, poor accuracy of high-confidence pseudo-labels, and imbalance of pseudo-label categories in the semi-supervised semantic segmentation task of colorectal cancer pathology images, and realizing multi-category precise segmentation of colorectal cancer pathology images.

[0009] The present invention adopts the following technical solution: On one hand, the present invention provides a method for semi-supervised semantic segmentation of colorectal cancer pathology images with controlled pseudo labels, comprising: S1 acquires images of the target pathological area through an image acquisition device to obtain an image data set of the target pathological area; S2 performs semantic segmentation mask annotation on the acquired image dataset and crops it into small-sized images, which are then divided into training set, validation set, and test set in proportion; S3 improves the utilization of low-confidence pseudo-labels in the consistency regularization process of training the teacher model, sorts the categories of low-confidence pseudo-labels according to the confidence scores and reduces the category space, and enhances the certainty of low-confidence pseudo-labels by excluding confusing categories; S4 uses the difference in prediction scores between the final model of the trained teacher network and the optimal model to evaluate the confidence of the pseudo-labels, and further refines the pseudo-labels with high confidence to solve the problems of blurred boundaries and lack of semantic information; S5 sets the probability of triggering each sub-data enhancement method according to the number of categories in the pseudo-label to generate cascaded pixel random changes, thereby alleviating the category imbalance problem in the pseudo-label data; S6 builds a semi-supervised semantic segmentation model for pseudo-labels of colorectal cancer pathology images based on the exclusion of confusing categories of low-confidence pseudo-labels, pseudo-label confidence evaluation and refinement, and cascaded pixel random changes. The model is weighted on the training set, and the trained weights are used for testing and verification on the test set and validation set.

[0010] The S1 includes: Determining a pixel range based on an image acquisition boundary of an image acquisition device; Within the pixel range, marking the target pathological region and obtaining data of several pathological regions; The images corresponding to the plurality of pathological region data are integrated to obtain an image data set of the target pathological region.

[0011] The S2 includes: Preprocess the acquired image data (full-field digital slices), use professional annotation tools to perform semantic segmentation mask annotation on the preprocessed images, and accurately identify the boundaries and category information of the pathological area; The annotated full-view images are cropped into small-size images suitable for model training and divided into training set, validation set and test set in proportion.

[0012] The S3 includes: Set the pseudo-label category confidence threshold; The pixels whose maximum category confidence value in the pseudo-label is greater than the set threshold are identified as credible pixels, and the cross entropy calculation is performed directly; The pixels whose maximum category confidence value in the pseudo-label is lower than the set threshold are identified as untrustworthy pixels, and the category with the second highest confidence value of the pixel (the confused category) is excluded, and the cross entropy is recalculated on the remaining categories after exclusion.

[0013] The S4 includes: Determine whether the pseudo-label is reliable by calculating the difference in prediction scores between the final model of the network and the optimal model of the network; The rough pseudo-label is used to capture the feature information of each pixel independent of its surrounding pixels, and then the spatial relationship and color correlation between pixels are captured according to the original image, and the rough pseudo-label is refined; Morphological operations are used to remove small holes and small objects, and then the image is repaired by filling the holes.

[0014] The S5 includes: Multiple data augmentation operations are cascaded and images are randomly cut out and overlapped to achieve diverse training data augmentation. Design a new data augmentation method: make the new pseudo-label consist of half pseudo-label and half real label to reduce the probability of single-class pseudo-label appearing; The probability of using each sub-data augmentation operation is set according to the number of pseudo-label categories.

[0015] The S6 includes: By excluding the confusing categories of low-confidence pseudo-labels, the certainty of low-confidence pseudo-labels is enhanced and their utilization rate is improved; Through the pseudo-label confidence evaluation and refinement, the problem of fuzzy boundaries and lack of semantic information of high-confidence pseudo-labels is solved; The class imbalance problem in pseudo-label data is alleviated by the cascaded random pixel change.

[0016] The image dataset is segmented based on the exclusion of confusion categories of low-confidence pseudo-labels, pseudo-label confidence evaluation and refinement, and cascaded pixel random changes.

[0017] Another aspect of the present invention provides a pseudo-label confidence-regulated semi-supervised semantic segmentation system for colorectal cancer pathology images, comprising: Image acquisition module: collects images of the target pathological area through an image acquisition device to obtain an image data set of the target pathological area; Dataset division module: semantic segmentation mask annotation is performed on the acquired image dataset and it is cropped into small-size images, and then divided into training set, validation set and test set according to proportion; Category sorting and reduction module: sorts and reduces the categories of low-confidence pseudo-labels according to the confidence scores, and enhances the certainty of low-confidence pseudo-labels by excluding confusing categories, thereby improving the utilization rate of low-confidence pseudo-labels; Confidence evaluation and refinement module: The confidence of the pseudo-label is evaluated by using the difference in prediction scores between the final model of the trained teacher network and the optimal model, and pseudo-labels with high confidence are further refined to address their blurred boundaries and lack of semantic information; Generate cascaded pixel random change module: According to the number of categories in the pseudo-label, set the probability of triggering each sub-data enhancement method to generate cascaded pixel random changes, thereby alleviating the category imbalance problem in the pseudo-label data; Model building and training module: Through the above-mentioned elimination of confusing categories of low-confidence pseudo-labels, pseudo-label confidence evaluation and refinement, and cascaded pixel random changes, a semi-supervised semantic segmentation model for pseudo-labels of colorectal cancer pathology images is constructed. The model is weighted on the training set, and the trained weights are used for testing and verification on the test set and validation set.

[0018] One or more technical solutions provided in the present invention have at least the following technical effects or advantages: First, in the teacher model training stage, the confusion category exclusion strategy is implemented to remove the confusion categories in the low-confidence pseudo-labels to enhance the certainty of the low-confidence pseudo-labels and improve their utilization. Secondly, in the student model training stage, the pseudo-label confidence evaluation and refinement module is added, and the strong data enhancement method of generating cascaded pixel random changes is used to improve the problem of blurred boundaries and lack of semantic information by performing fusion refinement operations on the evaluated high-confidence pseudo-labels, and cascade multiple data enhancement operations to randomly cut out and overlap the images, thereby achieving diverse training data expansion and realizing the technical solution of the semi-supervised semantic segmentation model for pseudo-labels of colorectal cancer pathology images. Then, the confusion category exclusion strategy is implemented to remove the confusion categories in the low-confidence pseudo-labels to enhance the certainty of the low-confidence pseudo-labels and improve their utilization. Secondly, the high-confidence pseudo-labels after evaluation are refined by fusion to improve their blurred boundaries and lack of semantic information, and the strong data enhancement method that can generate cascaded pixel random changes is used to alleviate the category imbalance in the pseudo-label data. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 A schematic diagram of a process flow of a method for semi-supervised semantic segmentation of colorectal cancer pathology images with regulated pseudo labels provided by the present invention; Figure 2 A schematic diagram of the semi-supervised semantic segmentation network structure for pseudo-labeling of colorectal cancer pathology images provided by the present invention; Figure 3 A schematic diagram of the confusion category structure for excluding low-confidence pseudo-labels provided by the present invention; Figure 4 A schematic diagram of the pseudo-label confidence evaluation and refinement structure provided by the present invention; Figure 5 A schematic diagram of a cascaded pixel random change structure provided by the present invention; Figure 6 A schematic diagram of the structure of a semi-supervised semantic segmentation system for colorectal cancer pathology images with regulated pseudo labels provided by the present invention; Explanation of the reference numerals: image acquisition module a, data set division module b, category sorting and reduction module c, confidence assessment and refinement module d, generation of cascaded pixel random change module e, model building and training module f. DETAILED DESCRIPTION

[0020] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments. Embodiment 1

[0021] like Figure 1-2 As shown, the present invention provides a method for semi-supervised semantic segmentation of colorectal cancer pathology images with regulated pseudo labels, comprising: S1 acquires images of the target pathological area through an image acquisition device to obtain an image data set of the target pathological area; S2 performs semantic segmentation mask annotation on the acquired image dataset and crops it into small-sized images, which are then divided into training set, validation set, and test set in proportion; S3 improves the utilization of low-confidence pseudo-labels in the consistency regularization process of training the teacher model, sorts the categories of low-confidence pseudo-labels according to the confidence scores and reduces the category space, and enhances the certainty of low-confidence pseudo-labels by excluding confusing categories; S4 uses the difference in prediction scores between the final model of the trained teacher network and the optimal model to evaluate the confidence of the pseudo-labels, and further refines the pseudo-labels with high confidence to solve the problems of blurred boundaries and lack of semantic information; S5 sets the probability of triggering each sub-data enhancement method according to the number of categories in the pseudo-label to generate cascaded pixel random changes, thereby alleviating the category imbalance problem in the pseudo-label data; S6 builds a semi-supervised semantic segmentation model for pseudo-labels of colorectal cancer pathology images based on the exclusion of confusing categories of low-confidence pseudo-labels, pseudo-label confidence evaluation and refinement, and cascaded pixel random changes. The model is weighted on the training set and the trained weights are used for testing and verification on the test set and validation set.

[0022] Specifically, the image acquisition device used in S1 is an electronic colonoscope. As a special acquisition device, the electronic colonoscope enters the rectum through the anus and displays the patient's rectum and colon images clearly on the computer screen to assist doctors in observing internal lesions. The electronic colonoscope uses its lens to acquire images from multiple perspectives and fields of view. The acquired images are sorted and collected to form an image dataset of the target pathological area. This dataset provides the basis for subsequent mask annotation and data division.

[0023] Specifically, three pathologists with 10 years of experience performed semantic segmentation mask annotation on the image dataset to obtain multiple pathological region data. The images corresponding to these pathological region data were integrated to form an image dataset of the target pathological region. These full-view digital slices and annotated full-view images were cropped and divided into slices of 224×224 size. These images were divided into training sets, validation sets, and test sets in proportion, which were used to learn the characteristics of image data, evaluate model performance, and ensure the generalization ability of the model.

[0024] Specifically, Figure 3 As shown, the specific steps of S3 include: Step S310: Unlabeled data The main decoder Get segmentation prediction ,when The confidence value of the largest category in is lower than the set threshold When The pixel at is an untrustworthy pixel. We start to exclude the category with the second highest confidence value (confusion category) of this pixel, and recalculate the softmax on the removed (C-1) category. If the confidence value is If the confidence value of the largest category in has exceeded the preset threshold, then only the category with the second highest confidence value can be removed, and then the confidence value can be calculated. ; Step S320: If the confidence If the confidence value of the largest category in does not exceed the preset threshold, the category with the third highest confidence value will be removed until the confidence value of the largest category exceeds the preset threshold. ; Step S330: If The maximum class confidence value in is higher than the threshold , then it is determined that The pixel at is a credible pixel, and the .

[0025] By implementing the confusing category exclusion strategy in the consistency regularization framework, setting the category confidence threshold to find untrustworthy pixels in pseudo-labels, reducing the category space of untrustworthy pixels and excluding confusing categories, the model becomes certain about the originally uncertain samples, thereby improving the utilization rate of low-confidence pseudo-labels in training.

[0026] Exemplarily, as shown in Table 1, the confusion category exclusion (CCE) strategy is implemented in the consistency regularization framework of the baseline network to improve the model's learning ability and usage rate of low-confidence pseudo labels, so that the average segmentation accuracy of the model reaches 70.33%.

[0027]

[0028] Specifically, Figure 4 As shown, the specific steps of S4 include: Step S410: Apply the best model in the teacher network to the unlabeled image The predicted value of As the true label, the model produced by the last iteration in the teacher network is calculated on the unlabeled image The predicted value on The mIoU score , thereby ranking the confidence scores of all unlabeled data; Step S420, extracting finer semantic targets of the colorectal cancer pathology image by using the unary potential energy and binary potential energy of DenseCRF; Step S430: Fill small holes and remove small objects by using the hole repair module Hole-Repair.

[0029] A strategy combining conditional random fields and hole filling is adopted to refine the high-confidence pseudo-labels after confidence evaluation to eliminate holes and error masks on high-confidence pseudo-labels, thereby obtaining more accurate pseudo-labels.

[0030] Exemplarily, as shown in Table 2, adding the confidence evaluation and refinement module (CER) in the student network training stage enables the model to obtain more accurate high-quality pseudo labels, making the average segmentation accuracy of the model reach 68.89%.

[0031]

[0032] Specifically, Figure 5 As shown, the specific steps of S5 include: Step S510: adopt the idea of ​​random cascading, cascade multiple data enhancement operations such as CutOut and CutMix, and randomly cut and overlap the image, so as to achieve diverse training data expansion; Step S520: A new data enhancement method CutTwo is designed. This method makes the new pseudo-label consist of half of the pseudo-label and half of the real label, which reduces the probability of the appearance of a single-class pseudo-label and strengthens the data disturbance. Step S530: adaptively expand the data changes by setting the occurrence probability of each sub-data enhancement method to generate cascaded pixel random changes. When the number of categories is less than or equal to 2, the probability of using CutTwo is set to 1, otherwise the probability of using CutTwo is 0.5, while CutOut and CutMix maintain the probability of 0.5.

[0033] By designing a new data augmentation method, multiple data augmentation operations are cascaded and the images are randomly cut out and overlapped to generate cascaded pixel random changes. At the same time, the probability of each sub-data augmentation operation is determined according to the number of pseudo-label categories to alleviate the category imbalance problem.

[0034] For example, as shown in Table 3, a cascaded pixel random change (CPRC) is constructed in the student network training stage to improve the category imbalance problem of pseudo labels during model training, so that the average segmentation accuracy of the model reaches 69.74%.

[0035]

[0036] Specifically, in view of the low utilization rate of low-confidence pseudo-labels in the semi-supervised semantic segmentation task of colorectal cancer pathology images, a strategy is proposed to exclude confusing categories of low-confidence pseudo-labels. The confusing categories in the low-confidence pseudo-labels generated by the untrained teacher model are excluded to enhance certainty, thereby improving the utilization rate of low-confidence pseudo-labels. Secondly, the accuracy of high-confidence pseudo-labels is poor. In this regard, it is proposed to perform confidence evaluation and refinement on the pseudo-labels generated by the trained teacher model. By performing a fused refinement operation on the evaluated high-confidence pseudo-labels, the problems of blurred boundaries and lack of semantic information in high-confidence pseudo-labels are improved. Finally, in view of the imbalance problem of pseudo-label categories, a strong data augmentation method that can generate cascaded pixel random changes is designed.

[0037] By integrating the above three strategies, the quality of pseudo labels and the adaptability of the model to pseudo labels are further improved, thereby significantly improving the overall performance of the model. In Table 4, the model with the comprehensive strategy achieved an mIoU of 74.09%, which is 6.43% higher than the basic network, and showed significant advantages in indicators such as mDice, mAcc, and mPre, reaching 85.04%, 82.86%, and 85.32%.

[0038]

[0039] It can be seen from Table 4 that the combined strategy proposed in the present invention is superior to the existing ST, CCT, ELN, U2PL and other methods in terms of mIoU, mDice, mAc and mPre indicators, especially in terms of mIoU and mDice, which has achieved a significant improvement, verifying the effectiveness of the CCE, CER and CPRC strategies, and proving the excellence of this method in the semi-supervised segmentation task of colorectal cancer pathology images.

[0040] In summary, the semi-supervised semantic segmentation method for colorectal cancer pathology images with regulated pseudo labels provided in this application has the following technical effects: A confusing category exclusion strategy is implemented in the consistency regularization of the teacher model to remove confusing categories in low-confidence pseudo-labels to enhance the certainty of the low-confidence pseudo-labels and improve their contribution rate. In addition, during the student model training phase, a pseudo-label confidence evaluation and refinement module and a strong data enhancement method for generating cascaded pixel random changes are added. The high-confidence pseudo-labels after evaluation are subjected to a fused refinement operation to improve their blurred boundaries and lack of semantic information. At the same time, multiple data enhancement operations are adaptively cascaded to perform random clipping and overlapping of images, thereby generating cascaded pixel random changes and achieving diverse training data expansion. Finally, the segmentation accuracy of the semi-supervised network model for colorectal cancer pathology images is improved by excluding confusing categories of low-confidence pseudo-labels, the pseudo-label confidence evaluation and refinement module, and the strong data enhancement for generating cascaded pixel random changes. Embodiment 2

[0041] Based on the same inventive concept as the method for semi-supervised semantic segmentation of colorectal cancer pathology images with pseudo-label regulation in the aforementioned embodiment, the present invention also provides a semi-supervised semantic segmentation system for colorectal cancer pathology images with pseudo-label confidence regulation, such as Figure 6 As shown, the system comprises: Image acquisition module a: acquires images of the target pathological area through an image acquisition device to obtain an image data set of the target pathological area; Dataset division module b: semantic segmentation mask annotation is performed on the acquired image dataset and it is cropped into small-size images, and then divided into training set, validation set and test set according to the proportion; Category sorting and reduction module c: sorts and reduces the categories of low-confidence pseudo-labels according to the confidence scores, and enhances the certainty of low-confidence pseudo-labels by excluding confusing categories, thereby improving the utilization rate of low-confidence pseudo-labels; Confidence evaluation and refinement module d: The confidence of the pseudo-label is evaluated by using the difference in prediction scores between the final model and the optimal model of the trained teacher network, and the pseudo-label with high confidence is further refined to solve the problem of its blurred boundary and lack of semantic information; Generate cascaded pixel random change module e: According to the number of categories in the pseudo-label, set the probability of triggering each sub-data enhancement method to generate cascaded pixel random changes, thereby alleviating the category imbalance problem in the pseudo-label data; Model building and training module f: Through the above-mentioned elimination of confusing categories of low-confidence pseudo-labels, pseudo-label confidence evaluation and refinement, and cascaded pixel random changes, a semi-supervised semantic segmentation model for pseudo-labels of colorectal cancer pathology images is constructed. The model is weighted on the training set, and the trained weights are used for testing and verification on the test set and validation set.

[0042] Through the above-mentioned detailed description of a method for semi-supervised semantic segmentation of colorectal cancer pathology images with regulated pseudo-labels in this specification, those skilled in the art can clearly know the method and system for semi-supervised semantic segmentation of colorectal cancer pathology images with regulated pseudo-labels in this embodiment. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the description of the method part.

[0043] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for semi-supervised semantic segmentation of colorectal cancer pathology images with controlled pseudo labels, characterized in that: The method comprises: S1 acquires images of the target pathological area through an image acquisition device to obtain an image data set of the target pathological area; S2 performs semantic segmentation mask annotation on the acquired image dataset and crops it into small-sized images, which are then divided into training set, validation set, and test set in proportion; S3 sorts and reduces the categories of low-confidence pseudo-labels according to their confidence scores during the consistency regularization process of training the teacher model, and enhances the certainty of low-confidence pseudo-labels by excluding confusing categories; S4 uses the difference in prediction scores between the final model of the trained teacher network and the optimal model to evaluate the confidence of the pseudo labels, and further refines the pseudo labels with high confidence; S5 sets the probability of triggering each sub-data enhancement method according to the number of categories in the pseudo-label, and generates cascaded pixel random changes; S6 constructs a semi-supervised semantic segmentation model for pseudo-labels of colorectal cancer pathology images by eliminating confusing categories of low-confidence pseudo-labels, pseudo-label confidence evaluation and refinement, and cascaded pixel random changes.

2. The method for semi-supervised semantic segmentation of colorectal cancer pathology images with controlled pseudo labels according to claim 1, characterized in that: The S1 includes: Determining a pixel range based on an image acquisition boundary of an image acquisition device; Within the pixel range, marking the target pathological region and obtaining data of several pathological regions; The images corresponding to the plurality of pathological region data are integrated to obtain an image data set of the target pathological region.

3. The method for semi-supervised semantic segmentation of colorectal cancer pathology images with controlled pseudo labels according to claim 1, characterized in that: The S2 includes: Preprocess the acquired image data, use annotation tools to perform semantic segmentation mask annotation on the preprocessed image, and accurately identify the boundaries and category information of the pathological area; The labeled images are cropped into small-sized images suitable for model training and divided into training set, validation set and test set in proportion.

4. The method for semi-supervised semantic segmentation of colorectal cancer pathology images with controlled pseudo labels according to claim 1, characterized in that: The S3 includes: Set the pseudo-label category confidence threshold; The pixels whose maximum category confidence value in the pseudo-label is greater than the set threshold are identified as credible pixels, and the cross entropy calculation is performed directly; The pixels whose maximum category confidence value in the pseudo-label is lower than the set threshold are identified as untrustworthy pixels, the category with the second highest confidence value of the pixel is excluded, and the cross entropy is recalculated on the remaining categories after exclusion.

5. The method for semi-supervised semantic segmentation of colorectal cancer pathology images with controlled pseudo labels according to claim 1, characterized in that: The S4 includes: Determine whether the pseudo-label is reliable by calculating the difference in prediction scores between the final model of the network and the optimal model of the network; The rough pseudo-label is used to capture the feature information of each pixel independent of its surrounding pixels, and then the spatial relationship and color correlation between pixels are captured according to the original image, and the rough pseudo-label is refined; Morphological operations are used to remove small holes and small objects, and then the image is repaired by filling the holes.

6. The method for semi-supervised semantic segmentation of colorectal cancer pathology images with controlled pseudo labels according to claim 1, characterized in that: The S5 includes: Multiple data augmentation operations are cascaded and images are randomly cut out and overlapped to achieve diverse training data augmentation. Make the new pseudo-label consist of half of the pseudo-label and half of the real label to reduce the probability of single-class pseudo-label appearing; The probability of using each sub-data augmentation operation is set according to the number of pseudo-label categories.

7. A semi-supervised semantic segmentation system for colorectal cancer pathology images with controlled pseudo labels, characterized in that: The system comprises: Image acquisition module: collects images of the target pathological area through an image acquisition device to obtain an image data set of the target pathological area; Dataset division module: semantic segmentation mask annotation is performed on the acquired image dataset and it is cropped into small-size images, and then divided into training set, validation set and test set according to proportion; Category sorting and reduction module: sorts and reduces the categories of low-confidence pseudo-labels according to the confidence scores, and enhances the certainty of low-confidence pseudo-labels by excluding confusing categories; Confidence evaluation and refinement module: The confidence of pseudo labels is evaluated by using the difference in prediction scores between the final model and the optimal model of the trained teacher network, and pseudo labels with high confidence are further refined; Generate cascaded pixel random change module: according to the number of categories in the pseudo-label, set the probability of triggering each sub-data enhancement method to generate cascaded pixel random changes; Model building and training module: By eliminating confusing categories of low-confidence pseudo-labels, evaluating and refining pseudo-label confidence, and cascading pixel random changes, a semi-supervised semantic segmentation model for pseudo-labels of colorectal cancer pathology images is constructed. The model is weighted on the training set, and the trained weights are used for testing and verification on the test set and validation set.

Images