CN114820431A - Channel non-equilibrium cancer cell image detection method based on reinforcement learning - Google Patents

Abstract

The invention discloses a channel unbalanced cancer cell image detection method based on reinforcement learning, which is characterized in that cancer cells are respectively segmented on red, green and blue color channels, confrontation network expansion part green channel data is generated through GAN, the expanded green channel is recombined with information of other two channels to expand a cancer cell data set with irregular morphology, and the green channel strength is reinforced through a reinforcement algorithm, so that the characteristics of the cancer cells are more obvious. The invention can complete the training task by a simple convolutional neural network through the expansion of the unbalanced data set and the reinforcement learning algorithm, and does not need to complete the identification of the malformed cancer cells by constructing a complex network structure or carrying out multiple detections, thereby greatly improving the detection efficiency. The off-line detection of the cancer cell image can be completed by deploying the lightweight network.

Description

A Channel Unbalanced Cancer Cell Image Detection Method Based on Reinforcement Learning

technical field

The present invention relates to the field of cancer cell monitoring, and more particularly to a method for detecting images of cancer cells with unbalanced channels based on reinforcement learning.

Background technique

As an important means of preventing and controlling cancer, cancer cell detection technology has many applications in both cancer prevention and cancer treatment. The current cancer cell image detection technology is gradually combined with the deep learning network to identify and judge, and has achieved good results. Cancer cell detection methods based on various deep learning networks such as Faster R-CNN, Mask R-CNN, and improved U-net have appeared, but deep learning networks often require a balanced cancer cell dataset for training, and the collection of CTC images It is often more difficult to collect than non-CTC images, and there are a few irregular CTC images, which leads to deep learning networks tend to be biased towards extracting non-CTC features, and the recognition accuracy of CTC by the network model is lower than that of non-CTC images. Identification accuracy, and it is difficult to identify CTCs with irregular shapes. What the present invention needs to do is to improve on the original basis to solve the unbalanced cancer cell data set, and the cancer cell detection technology based on the deep learning network has low recognition accuracy for cancer cells with irregular cell shapes, and is easy to miss the key. question.

Among them, the publication number is CN201811068427, the cancer cell recognition method based on the improved U-net convolutional neural network model: the trained neural network model has a high recognition rate only for cancer cells with good shape in medical images, and for a small number of cells. Irregularly shaped cancer cells, it is difficult for the neural network model to correctly classify and identify them;

A cancer cell detection method based on Faster R-CNN with publication number CN201810174376: It is necessary to add image coordinates to the label for training, and the model training process is complicated, time-consuming and labor-intensive. Through two cancer cell detections, the efficiency is low, and the detection model is difficult to have a high recognition rate for a few cancer cells with irregular shapes;

The intelligent detection method of cervical cancer cells based on deep learning with publication number CN202011259420: using the U-Net segmentation model to segment the nucleus, there is a certain loss of information. Expansion and category segmentation of classified data using active learning methods are labor-intensive. The trained ResNeSt classification model is difficult to accurately classify and identify a few irregular nuclei.

Therefore, how to provide a channel non-balanced cancer cell image detection method based on reinforcement learning with high recognition effect, high recognition accuracy and fast detection efficiency.

SUMMARY OF THE INVENTION

In view of this, the present invention provides a channel unbalanced cancer cell image detection method based on reinforcement learning.

To achieve the above object, the present invention provides the following technical solutions, which mainly include:

A reinforcement learning-based channel unbalanced cancer cell image detection method specifically includes the following steps:

S1: Get a certain number of cancer cell images and classify them;

S2: Separate the deformed CTC image based on the channel level; obtain the irregular green channel image of the deformed CTC, and send it to the GAN generation network, and send the generated image and the real image to the discrimination network for discrimination. The result of the network performs reverse training on the generation network and the discriminant network. After the training, the expansion of the green channel image can be realized through the generation network;

S3: Separate the typical CTC image based on the channel level to obtain the red and blue channel image of the typical CTC, and randomly combine it with the expanded green channel image with irregular shape to obtain a new deformed CTC image;

S4: Separate the color weak CTC image based on the channel level, and obtain the red and blue channel image and the green channel image of the color weak CTC image respectively;

S5: Rotate, fold, mirror and other operations on the expanded and enhanced CTC image, perform secondary expansion, and mark it as True, obtain a certain number of non-CTC images and mark it as False, and obtain the ratio of positive samples and negative samples as r The data set of cancer cells is divided into training set and test set with a certain ratio (the ratio is generally 0.2 to 0.3);

S7: Finally, the offline detection of cancer cell images can be completed through the trained network model parameters and the deployed network.

Preferably, in the classification of step S1, it is divided into typical CTCs, chromatic weak CTCs (generally weak green) and atypical (malformed) CTCs.

Preferably, the piecewise linear transformation is performed on the green channel image in the step S4, and the pixel points less than 30 are assigned as 0, which can effectively remove background noise; the pixels of 30-180 are stretched and transformed to 60- 255, the features of the green channel image are enhanced; the pixels larger than 180 are assigned as 255, and the influence of background noise is also eliminated. The red and blue channel images and the enhanced green channel images are combined to obtain an enhanced CTC image.

Preferably, in the step S5, the training set data is sent to the DQN reinforcement learning network for training. For images of cancer cells with a small proportion, after each environmental interaction, according to whether the classification result of the agent is correct or not, the environment will be The agent will be given a large positive (negative) reward to guide the agent to learn more about the classification of cancer cells. Among them, the value network chooses a simple CNN convolutional neural network. The value network undergoes multiple iterative training to continuously update the network parameters. After every certain number of iterations, the judgment accuracy of the network is tested on the test set until the training is completed; after each test, the model parameters will be in the form of a .ckpt file. Save it.

It can be seen from the above technical solutions that, compared with the prior art, the present invention enhances and expands the data set of deformed cancer cells based on the color channel level, and proposes a new method for processing unbalanced data sets; The processing of the adversarial network and the image color channel effectively expands the data set of the deformed CTC, and improves the recognition accuracy of the deformed CTC by the deep neural network; the present invention uses a simple CNN after data expansion, image enhancement and other processing. The convolutional neural network can fully express the characteristics of typical CTCs and deformed CTCs, simplifies model deployment, realizes offline operation, and improves identification efficiency; the present invention obtains the expanded unbalanced cancer cell data set through the DQN reinforcement learning algorithm The classification model, combined with the convolutional neural network, effectively improves the recognition accuracy of the model for cancer cells, and solves the problem that CTC images are not easy to obtain.

Description of drawings

In order to explain the embodiments of the present invention or the technical solutions in the prior art more clearly, the following briefly introduces the accompanying drawings that need to be used in the description of the embodiments or the prior art. Obviously, the accompanying drawings in the following description are only It is an embodiment of the present invention. For those of ordinary skill in the art, other drawings can also be obtained according to the provided drawings without creative efforts.

FIG. 1 is a schematic diagram of the classification of cancer cells according to the present invention.

FIG. 2 is a structural diagram of the GAN generative adversarial network of the present invention.

FIG. 3 is a schematic diagram of separation and recombination of typical and atypical CTCs based on the channel level of the present invention.

FIG. 4 is a schematic diagram of enhancing the color weak CTC based on the channel level according to the present invention.

FIG. 5 is a structural diagram of the DQN reinforcement learning network of the present invention.

FIG. 6 is a structural diagram of a convolutional neural network of the present invention.

Detailed ways

The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

Example

As shown in Figure 1-6;

First, obtain the image of cancer cells collected by the sampling needle provided by Delutong Biotechnology Co., Ltd. as the test sample we need.

1. Cancer cell image classification: The original cancer cell dataset is manually classified into three categories: typical CTCs, color-weak CTCs, and atypical (malformed) CTCs.

2. Use GAN to generate a new green channel image: Separate the deformed CTC images based on the channel level, obtain multiple green channel images of deformed CTCs, send them to the GAN generation network, and combine the generated images with the real images. It is sent to the discriminant network for discrimination, and the generation network and the discriminant network are reverse-trained according to the results of the discriminant network, until the discriminant network cannot distinguish the generated image from the real image, and the training ends. After training, a new green channel image is obtained through the generative network.

3. Separate the typical CTC images based on the channel level, obtain the red and blue channel images of typical CTCs, and randomly combine them with the expanded green channel images with irregular shapes to obtain new deformed CTC images.

4. Image enhancement based on the channel level: Channel separation is performed on the color-weak CTC image to obtain the red and blue channel images and green channel images of the color-weak CTC, and piecewise linear transformation is performed on the green channel image, and the pixels less than 30 are assigned as 0 ; Scale the pixels of 30-180 to 60-255; assign the pixels larger than 180 to 255 to obtain an enhanced single-channel image. The enhanced single-channel image is recombined with the red and blue channel images to obtain an enhanced CTC image.

5. Rotate, fold, mirror and other operations on the expanded and enhanced CTC image, perform secondary expansion, and mark it as True, get a certain number of non-CTC images and mark it as False, and get the ratio of positive samples and negative samples as r The cancer cell data set is divided into training set and test set with a certain ratio (the ratio is generally 0.2 to 0.3). The training set data is sent to the DQN reinforcement learning network for training. After each environmental interaction, if the agent correctly classifies the CTC, the environment gives the agent a positive return p, and the next environment interaction is performed; if the agent misclassifies the CTC , the environment gives the agent a negative return -p, ends the interaction, calculates the total return, and proceeds to the next iteration; if the agent correctly classifies non-CTC, the environment gives the agent a positive return rp, and performs the next environment interaction; Misclassify CTC, the environment gives the agent a negative reward -rp, end the interaction, calculate the total reward, and proceed to the next iteration. Combined with the convolutional neural network to maximize the return, the convolutional neural network is trained for 1000 iterations, and the network parameters are continuously updated. The network performance is tested on the test set every 50 times of training, and the updated model parameters will be updated after each test. The .ckpt file is saved in the specified path. After the training is completed, a mature convolutional neural network for cancer cell recognition can be generated.

6. Load the model parameters in the .ckpt file, use the trained 40×40 convolutional network window to traverse the input large image containing CTC and non-CTC with a stride of 2, and identify cancer cells in the large image And save the image information of each identified cancer cell.

7. Accurate identification of typical CTC and deformed CTC cells can be achieved by summarizing all the identified CTC image information

The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments, and the same and similar parts between the various embodiments can be referred to each other. As 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 part can be referred to the description of the method.

The above description of the disclosed embodiments enables any person skilled in the art to make or use the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the generic principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Thus, the present invention is not intended to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims (4)

1. a channel non-equilibrium cancer cell image detection method based on reinforcement learning, is characterized in that, specifically comprises the following steps: S1: Get a certain number of cancer cell images and classify them; S2: Separate the deformed CTC image based on the channel level; obtain the irregular green channel image of the deformed CTC, and send it to the GAN generation network, and send the generated image and the real image to the discrimination network for discrimination. The result of the network performs reverse training on the generation network and the discriminant network. After the training, the expansion of the green channel image can be realized through the generation network; S3: Separate the typical CTC image based on the channel level to obtain the red and blue channel image of the typical CTC, and randomly combine it with the expanded green channel image with irregular shape to obtain a new deformed CTC image; S4: Separate the color weak CTC image based on the channel level, and obtain the red and blue channel image and the green channel image of the color weak CTC image respectively; S5: Rotate, fold, mirror and other operations on the expanded and enhanced CTC image, perform secondary expansion, and mark it as True, obtain a certain number of non-CTC images and mark it as False, and obtain the ratio of positive samples and negative samples as r The data set of cancer cells is divided into training set and test set with a certain ratio (the ratio is generally 0.2 to 0.3); S7: Finally, the offline detection of cancer cell images can be completed through the trained network model parameters and the deployed network. 2. a kind of channel disequilibrium cancer cell image detection method based on reinforcement learning according to claim 1, is characterized in that, the classification of described step S1, it is divided into typical CTC, color weak CTC (generally is green weak) and atypical (malformed) CTCs. 3. A reinforcement learning-based channel unbalanced cancer cell image detection method according to claim 1, characterized in that, in the step S4, the green channel image is subjected to piecewise linear transformation, and the pixel points less than 30 The value of 0 can effectively remove the background noise; the pixels of 30-180 are stretched and transformed to 60-255 to enhance the features of the green channel image; the pixels larger than 180 can be assigned to 255, which also eliminates the background noise Impact. The red and blue channel images and the enhanced green channel images are combined to obtain an enhanced CTC image. 4. a kind of channel disequilibrium cancer cell image detection method based on reinforcement learning according to claim 1, is characterized in that, in described step S5, the training set data is sent into DQN reinforcement learning network for training, for accounting With relatively few images of cancer cells, after each environmental interaction, according to whether the classification result of the agent is correct or not, the environment will give the agent a larger positive (negative) reward to guide the agent to learn more about cancer cells. Classification. Among them, the value network chooses a simple CNN convolutional neural network. The value network is trained for many times to continuously update the network parameters. After every certain number of iterations, the judgment accuracy of the network is tested on the test set until the training is completed; after each test, the model parameters will be in the form of a .ckpt file. Save it.

Images