CN-121998895-A - Intestinal cancer pathological image screening method applied to fluorescence immunohistochemistry

CN121998895ACN 121998895 ACN121998895 ACN 121998895ACN-121998895-A

Abstract

The invention relates to the technical field of medical image processing and artificial intelligence, and particularly discloses a method for screening pathological images of intestinal cancer applied to fluorescence immunohistochemistry. The method comprises the steps of carrying out multichannel acquisition and pretreatment on fluorescent pathological sections, constructing and training an image screening model based on a convolutional neural network, realizing automatic screening and grading of pathological images through a trained deep learning model, and processing massive fluorescent immunohistochemical intestinal cancer pathological images in batches without manual evaluation, thereby greatly shortening the screening period and solving the problems of low efficiency and difficulty in coping with the high-throughput image analysis requirement of the traditional manual screening. The automatic process not only reduces the workload of pathologists, but also avoids fatigue errors caused by long-time manual screening, and provides high-efficiency and stable data screening support for follow-up accurate quantitative analysis and AI model training, thereby assisting the efficient propulsion of clinical pathology research and diagnosis.

Inventors

Bu Lingbin
ZHANG RUI
LIU XIUCAI
HU ZICHAO

Assignees

江苏阔然生物医药科技有限公司

Dates

Publication Date: 20260508
Application Date: 20251204

Claims (7)

1. The method for screening the pathological image of the intestinal cancer applied to fluorescence immunohistochemistry is characterized by comprising the following steps of: S1, carrying out multichannel image acquisition on a fluorescence immune histochemical intestinal cancer pathological section to obtain an original image, and carrying out pretreatment operations comprising denoising, background enhancement and brightness homogenization on the original image to obtain a pretreated image; S2, constructing and training a pathological image screening model based on a convolutional neural network, extracting features of the pathological image through supervised learning, outputting image quality evaluation indexes comprising fluorescent signal intensity, background noise proportion and cell structure definition, generating an approximate synthetic pathological image based on a generation countermeasure network technology, and expanding a training data set; S3, automatically screening and grading the preprocessed intestinal cancer pathological images to be screened by using a trained deep learning model; and S4, displaying screening results and image grading basis through a visual interface, and feeding back and optimizing the results of the screening model.
2. The method for screening the intestinal cancer pathological image applied to fluorescence immunohistochemistry according to claim 1 is characterized in that in the step S1, a laser confocal microscope is adopted for multi-channel image acquisition to acquire intestinal cancer pathological sections after fluorescence immunohistochemical treatment, a combination of median filter primary screening and frequency domain noise reduction optimization is adopted for denoising, a top hat transformation and a background deduction method are adopted for background enhancement treatment to enhance contrast ratio of fluorescent signals and background, and brightness uniformity treatment is based on region self-adaptive histogram equalization to achieve image brightness uniformity.
3. The method for screening pathological images of intestinal cancer for fluorescence immunohistochemistry according to claim 1, wherein said step S2 specifically comprises: 201. The construction and marking of a real pathological image data set comprises the steps of obtaining a plurality of preprocessed fluorescence immune constitutive intestinal cancer pathological images, presetting medical standards, carrying out quantitative label marking on the preprocessed fluorescence immune constitutive intestinal cancer pathological images, dividing the marked pathological images into a basic training set and a verification set according to a ratio of 7:3, constructing the basic training set into a real image library for GAN training, and carrying out format standardization treatment on the real image library; 202. training and constructing a basic CNN screening model; 203. Constructing a GAN model and performing countermeasure training; 204. screening and constructing a high-quality synthetic data set; 205. construction of a fusion data set and secondary training optimization of a CNN model: Mixing the basic training set and the high-quality synthetic data set according to a preset proportion to form a fusion training set, And stopping training when the predicted mean square error of the quantized label of the CNN model on the verification set meets a preset threshold value, so as to obtain a final CNN screening model.
4. The method for screening pathological images of intestinal cancer applied to fluorescence immunohistochemistry according to claim 3, wherein the training and construction of the basic CNN screening model comprises the following steps: 202-1, constructing a CNN model architecture, namely selecting a convolutional neural network CNN as a basic model, wherein the convolutional neural network CNN comprises an input layer, a feature extraction layer, a full-connection layer and a multi-task output layer, the input layer receives a standardized multi-channel pathological image, the feature extraction layer consists of a convolutional layer and a pooling layer alternately, the pooling layer adopts maximum pooling or average pooling operation, and the full-connection layer maps a high-dimensional feature image output by the feature extraction layer into a one-dimensional feature vector; 202-2, initializing parameters, namely adopting a mean square error as a loss function, adopting an Adam optimizer, setting an initial learning rate to be 0.001, setting a batch size to be 16, and setting iteration times to be 80-100 rounds; 202-3, forward propagation, namely inputting a basic training set into a CNN model, and obtaining an output result of the model through calculation of a convolution layer, a pooling layer and a full connection layer; 202-4, calculating a loss function, namely calculating an error by using the loss function according to the output result of the model and the real label marked by the expert; 202-5, back propagation and parameter updating, namely, back propagation is carried out according to the calculated loss function value, the gradient of each layer is calculated, and an Adam optimizer is used for updating parameters of a CNN model according to the gradient, so that the model can better fit training data; 202-6, iterative training and verification, namely repeating the processes of forward propagation, calculation of a loss function, reverse propagation and parameter updating, and performing repeated iterative training, wherein in the training process, the performance of the model is estimated by periodically using a verification set, and the super parameters of the model are adjusted according to the estimation result on the verification set so as to optimize the training effect of the model and obtain a basic CNN screening model.
5. A method for screening pathological images of intestinal cancer for fluorescence immunohistochemistry according to claim 3, wherein the screening construction of the high quality synthetic dataset comprises: 203-1, constructing a GAN model architecture, wherein the GAN comprises a generator G and a discriminator D, wherein the generator G is a network structure formed by alternately transposing convolution layers and up-sampling layers, inputs 100-dimensional random noise vectors, and outputs synthetic pathology images with consistent pathology image specifications; 203-2, initializing model parameters, namely randomly initializing the weights of a generator and a discriminator, taking a binary cross entropy loss function as a loss function, adopting an Adam optimizer as the optimizer, setting the initial learning rate to be 0.0002, setting the batch size to be 32, and setting the iteration times to be 50-80 rounds; 203-3, training a discriminator D, namely sampling a batch of real pathological images x from a real image library for GAN training, inputting the samples into the discriminator D, and calculating the real image loss ; The generator generates a permit a leave image Inputting into a discriminator D, calculating false image loss ; The total loss of the discriminator is The parameter of the discriminator D is updated by back propagation based on the total loss of the discriminator; 203-4, training generator G, generator generates permit a leave images Input to a discriminator D, calculate generator loss Back propagation based on generator loss, updating parameters of generator G; 203-5, iterating until convergence, and repeating the step 203-3 and the step 203-4 until the GAN reaches a nash equilibrium state.
6. The method for screening pathological images of intestinal cancer applied to fluorescence immunohistochemistry according to claim 4, wherein the specific process of screening and constructing the high-quality synthetic data set comprises the following steps: 204-1, controlling the trained generator G to receive a preset number of random noise vectors and generating synthetic pathological images in batches; 204-2, inputting the synthesized pathological images generated in batch into a basic CNN screening model obtained in the step 202-6, and outputting a quantization label of each synthesized pathological image by the basic CNN screening model; 204-3, screening the synthesized pathological images according to a preset screening threshold value, and reserving synthesized images with quantitative labels meeting the conditions; 204-4, manually rechecking the screened synthetic images, randomly extracting 10% -20% of the screened synthetic images, removing the synthetic images with abnormal morphology, and finally obtaining a high-quality synthetic data set.
7. The method for screening pathological images of intestinal cancer for fluorescence immunohistochemistry according to claim 1, wherein the image screening and grading comprises: inputting the preprocessed intestinal cancer pathology image to be screened into a trained CNN screening model, automatically performing forward propagation calculation on each image by the model, and outputting various corresponding quality index values; comprehensively calculating the index according to a preset weighting algorithm to obtain a comprehensive quality score; Comparing and analyzing the comprehensive quality score of the pathological image of the intestinal cancer to be screened with a corresponding preset threshold value to realize image classification, wherein the method specifically comprises the following steps: when the comprehensive quality score of the pathological images of the intestinal cancer to be screened is larger than a comparison threshold value TH1, judging that the pathological images of the intestinal cancer are high-quality images; when the comprehensive quality score of the pathological images of the intestinal cancer to be screened is between the comparison threshold values TH1 and TH2, judging the pathological images to be qualified images; And when the comprehensive quality score of the pathological images of the intestinal cancer to be screened is smaller than the comparison threshold value TH2, judging that the images are inferior.

Description

Intestinal cancer pathological image screening method applied to fluorescence immunohistochemistry Technical Field The invention relates to the technical field of medical image processing and artificial intelligence, in particular to a method for screening pathological images of intestinal cancer, which is applied to fluorescence immunohistochemistry. Background Fluorescent immunohistochemical technology has become one of the important tools of tumor pathology research and clinical diagnosis, realizes the fluorescent detection of target protein by labeling specific antibodies, and provides important basis for the molecular typing of cancer and the analysis of therapeutic targets. However, in the pathological research of intestinal cancer, obtaining high-quality fluorescence histochemical images faces the following technical bottlenecks: 1. The image quality difference is large, and the intensity and consistency of fluorescent signals are poor due to the influences of factors such as slice preparation, fluorescent microscope parameter setting, heterogeneity of samples and the like. 2. The manual screening efficiency is low, the high-throughput pathological image screening mainly depends on manual operation, the time consumption is long, the subjectivity is strong, and the consistency of the screening quality is difficult to ensure. 3. The high-quality data for the deep learning is insufficient, a large amount of high-quality annotation image data is needed for training the deep learning model, but the problems of noise and serious background interference exist in the data in the existing image library, the data requirement of the training model cannot be met, and the stable application of the model is affected. Therefore, an efficient and automatic pathological image screening method is urgently needed to improve the quality of fluorescence immunohistochemical data and support the pathological research and application of deep learning. Disclosure of Invention The invention aims to overcome the defects of strong subjectivity and low efficiency of the prior art relying on manual screening of pathological images and provides an automatic, objective and efficient intestinal cancer pathological image screening method. The invention aims to realize quantitative evaluation, automatic grading and efficient screening of the quality of fluorescence immune histochemical images by constructing an intelligent system integrating image preprocessing, automatic evaluation of a deep learning model and expert feedback optimization. The invention aims at realizing the following technical scheme that the method for screening the pathological image of the intestinal cancer applied to fluorescence immunohistochemistry comprises the following steps: S1, image acquisition and preprocessing, namely performing multi-channel image acquisition on a fluorescence immune-histochemical intestinal cancer pathological section to obtain an original image, and performing preprocessing operations comprising denoising, background enhancement and brightness homogenization on the original image to obtain a preprocessed image; S2, training and optimizing a deep learning model, namely constructing and training a pathological image screening model based on a Convolutional Neural Network (CNN), extracting characteristics of the pathological image through supervised learning, outputting image quality evaluation indexes comprising fluorescent signal intensity, background noise proportion and cell structure definition, introducing a generation countermeasure network (GAN) technology to generate an approximate synthetic pathological image, and expanding a training data set; s3, image screening and grading, namely automatically screening and grading the preprocessed intestinal cancer pathological images to be screened by using a trained deep learning model; And S4, visualization and verification of the screening result, namely providing a visual interface, displaying the screening result and the image grading basis, introducing an expert manual verification mechanism, and feeding back and optimizing the result of the screening model according to clinical expert review and medical standard verification. Compared with the prior art, the invention has the following beneficial effects: The invention adopts a laser confocal microscope to collect original images in multiple channels, designs a multi-step collaborative pretreatment flow, accurately distinguishes effective signals and noise through a combined denoising method combining median filtering and frequency domain noise reduction, keeps cell structure details while inhibiting interference, enhances the contrast ratio of fluorescent signals and background through top hat transformation and background deduction, reduces the interference of nonspecific fluorescence, realizes brightness uniformity through a region self-adaptive histogram equalization technology, and solves the problem of uneven brightness of