CN113576508B - Cerebral hemorrhage auxiliary diagnosis system based on neural network - Google Patents

Abstract

The invention discloses a cerebral hemorrhage auxiliary diagnosis system based on a neural network, which belongs to the technical field of medical image processing and segmentation, and comprises the following components: the acquisition module is used for acquiring a plurality of original CT images and marking a focus area; the processing module is used for carrying out image enhancement on the plurality of original CT images to obtain a training set; the training module is used for training the neural network by utilizing the training set; the first diagnosis module is used for acquiring a CT image shot for the second time, preprocessing the CT image, inputting the processed image into a trained neural network, and obtaining a gray level image which is consistent with the size of the input image and represents the confidence coefficient of each pixel; the second diagnosis module is used for sequentially carrying out threshold segmentation, morphological closing operation and hole elimination treatment on the gray level image; and according to the processed gray level image, cerebral hemorrhage focus detection and hemorrhage amount calculation are realized. Therefore, the invention can improve the efficiency and accuracy of cerebral hemorrhage diagnosis and treatment.

Description

A neural network-based auxiliary diagnosis system for cerebral hemorrhage

Technical Field

The present invention belongs to the technical field of medical image processing and segmentation, and more specifically, relates to a neural network-based auxiliary diagnosis system for cerebral hemorrhage.

Background technique

Cerebral hemorrhage refers to primary non-traumatic bleeding in the brain parenchyma, also known as spontaneous cerebral hemorrhage. It is the type of disease with the highest mortality rate among acute cerebrovascular diseases, accounting for 20% to 30% of acute cerebrovascular diseases. At present, acute cerebrovascular-related diseases have become the third leading cause of death. The determination of the amount of bleeding in cerebral hemorrhage can serve as an important reference for the subsequent treatment measures taken by doctors.

At present, X-ray computed tomography (CT) is widely used in clinical practice as the main screening method for cerebral hemorrhage, which has the advantages of convenient operation, economy and non-invasiveness. However, since the hemorrhage area in the CT image of cerebral hemorrhage is easily confused with other normal brain tissues, it is difficult for diagnostic doctors to make qualitative observations and quantitative analyses; at the same time, cerebral hemorrhage occurs rapidly and progresses rapidly, and the symptoms reach a peak within minutes or hours after onset. However, since pathological diagnosis requires microscopic operation and the workload of naked eye observation is large, CT reports generally take 6-24 hours to be obtained, the waiting time is too long, and the inspection report is easily affected by the experience and fatigue status of the pathologist. At the same time, some remote areas are limited by the distribution of medical resources, especially the lack of experienced doctors, and it is difficult to quickly determine the CT diagnosis, which can easily cause patients to miss the best time for treatment.

With the development of machine learning technology and its wide application in the medical field, in the field of auxiliary diagnosis of cerebral hemorrhage, the establishment of auxiliary analysis tools for cerebral hemorrhage CT and pathological slice images based on machine learning methods can utilize the powerful image processing and matrix computing capabilities of computers to improve image analysis efficiency, reduce the workload of doctors, and help to alleviate the imbalance of regional medical resources to a certain extent. However, for the sake of protecting patient privacy, the CT images taken by the CT device are stored in the internal system of the hospital, and generally only support printing into films. It is difficult to directly open the data interface to external programs, making it difficult to apply the above-mentioned auxiliary analysis program based on machine learning to the original captured images. One way to solve this problem is to take a second shot of the printed CT film and use it as the input of the program, but the quality of the second-shot CT image is greatly affected by the shooting angle and lighting conditions, making it difficult for conventional auxiliary analysis programs to obtain accurate cerebral hemorrhage diagnosis results. Therefore, designing an effective auxiliary diagnosis system for cerebral hemorrhage based on the second-shot CT image is a problem that needs to be solved at present, and it has important practical application value.

Summary of the invention

In response to the above defects or improvement needs of the prior art, the present invention provides a neural network-based auxiliary diagnosis system for cerebral hemorrhage. It only requires using a mobile phone to take a CT image of the patient and upload it, and then the trained neural network can be used to complete bleeding lesion detection, bleeding volume calculation, etc., thereby helping doctors determine the patient's cerebral hemorrhage condition and improving diagnostic efficiency.

To achieve the above object, the present invention provides a neural network-based cerebral hemorrhage auxiliary diagnosis system, comprising the following modules:

An acquisition module is used to acquire multiple original CT images and mark the lesion area;

A processing module, used for performing image enhancement on the plurality of original CT images to obtain a training set;

A training module, used for training a neural network using the training set;

The first diagnostic module is used to obtain the secondary CT image and perform preprocessing, and input the processed image into the trained neural network to obtain a grayscale image representing the confidence of each pixel that is consistent with the size of the input image;

The second diagnosis module is used to perform threshold segmentation, morphological closing operation and hole elimination processing on the grayscale image in sequence; and realize cerebral hemorrhage lesion detection and bleeding volume calculation based on the processed grayscale image.

Furthermore, the training module is also used to import multiple adjacent images into the input layer of the neural network for joint training.

Furthermore, the training module is also used to calculate the loss between the grayscale image output by the neural network and the annotated original CT image according to the selected cross entropy loss function, and back-propagate the loss to update the neural network weight parameters.

Furthermore, the second-shot CT image is obtained and preprocessed, including: obtaining the second-shot CT image, and performing image center cropping and image compression to make the processed image consistent with the shape and size of the original CT image.

Furthermore, after performing image center cropping and image compression, the first diagnosis module is also used to perform grayscale processing on the image.

Furthermore, the image enhancement methods adopted include: grayscale random jitter, contrast adjustment, multiple noise simulations, filtering operations and random bright spot superposition.

Furthermore, the neural network is a fully convolutional semantic segmentation neural network.

In general, the above technical solutions conceived by the present invention can achieve the following beneficial effects:

(1) The brain hemorrhage auxiliary diagnosis system provided by the present invention takes into account that in actual applications, the CT images taken by the CT device cannot be exported immediately, and it is impossible to quickly use the trained neural network to complete the detection of bleeding lesions, the calculation of the amount of bleeding, etc., so it is chosen to perform a second shot of the original CT image through a user device such as a mobile phone. Since there are differences between the second-shot image and the original CT image, the grayscale distribution, image shape and other features of the original CT image in the training set are changed through an image enhancement method to make it closer to the features of the second-shot image, thereby improving the generalization ability of the learned model. After obtaining the corresponding grayscale image based on the trained neural network, threshold segmentation, morphological closing operation and hole elimination processing are performed in sequence to restore the actual brain hemorrhage situation. In this way, the present invention can improve the efficiency and accuracy of brain hemorrhage diagnosis and treatment.

(2) Before the image of the training set is input into the neural network, the present invention designs a program for importing multiple adjacent images at the same time, firstly sorting the images in order, and before the current image is input into the neural network, querying the adjacent images before and after the current image, and importing them and the current image into the input layer of the neural network together for joint training. This can reduce the false alarm rate of the final result, making the final segmentation result more credible.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a block diagram of a neural network-based cerebral hemorrhage auxiliary diagnosis system provided by an embodiment of the present invention;

FIG. 2 is a visualization effect diagram of cerebral hemorrhage detection provided by an embodiment of the present invention.

Detailed ways

In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

In the present invention, the terms "first", "second", etc. (if any) in the present invention and the accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence.

As shown in FIG1 , the neural network-based cerebral hemorrhage auxiliary diagnosis system 100 provided by the present invention includes:

An acquisition module 110 is used to acquire multiple original CT images and mark the lesion area;

Specifically, the JSON tag files marked by professionals are batch converted into masks for model input, that is, jpg binary images of the same size as the CT image. Mask is used as the correctly labeled data input model, where the binary values of Mask represent the areas with cerebral hemorrhage lesions and the areas without cerebral hemorrhage lesions.

A processing module 120, configured to perform image enhancement on the plurality of original CT images to obtain a training set;

Specifically, since the grayscale image of the original CT is directly acquired, but the system actually uses the CT image taken by the user device such as a mobile phone, there is a problem of inconsistent distribution of the training set data and the test set data for the neural network. For this problem, after analyzing the image histograms of the original CT image and the second-shot CT image, we have targetedly selected reasonable image enhancement methods to change the grayscale distribution, image shape and other features of the original CT image in the training set to make it closer to the features of the second-shot image, thereby improving the generalization ability of the learned model. The image enhancement methods used include: grayscale random jitter, contrast adjustment (selecting appropriate contrast parameters for the histogram of the second-shot image), multiple noise simulations (simulating noise interference in shooting), filtering operations (simulating jitter blur that may occur in shooting), and random bright spot superposition (used to simulate reflections in shooting).

A training module 130, configured to train a neural network using the training set;

Specifically, the neural network is a fully convolutional semantic segmentation neural network, which has a significant effect on the segmentation problem of medical images; the network is mainly composed of an encoder, a decoder, and a skip connection: the encoder acts as a downsampler, converting the image into a feature map containing high semantic information; the decoder acts as an upsampler, converting the high semantic information into a score map for pixel-level classification; the skip connection is used to connect the front and back feature maps at the same scale. The network output is a grayscale image representing the confidence of each pixel with the same size as the input image. According to the selected cross entropy loss function, the loss between the grayscale image output by the neural network and the annotated original CT image is calculated, and the loss is back-propagated to update the neural network weight parameters.

Furthermore, before the training set images are input into the neural network, a program is designed to import multiple adjacent images at once. The images are sorted in order, and before the current image is input into the neural network, the adjacent images before and after the current image are queried, and the images are imported into the input layer of the neural network together with the current image for joint training. This can reduce the false alarm rate of the final result and make the final segmentation result more credible.

The first diagnosis module 140 is used to obtain the second-shot CT image and perform preprocessing, and input the processed image into the trained neural network to obtain a grayscale image representing the confidence of each pixel that is consistent with the size of the input image;

Specifically, doctors take stable and high-precision CT images under good lighting conditions and upload them; obtain secondary CT images and perform preprocessing, including: image center cropping (cropping to a square shape consistent with the training set image), image compression (compressing the image to a size consistent with the training set to reduce the amount of calculations during the test), and image grayscale.

The second diagnosis module 150 is used to perform threshold segmentation, morphological closing operation and hole elimination processing on the grayscale image in sequence; and to detect cerebral hemorrhage lesions and calculate the amount of bleeding based on the processed grayscale image.

Specifically, the grayscale image is subjected to threshold segmentation processing (preliminarily obtaining the foreground representing the cerebral hemorrhage lesion and the background representing the lesion-free lesion), morphological closing operation (used to eliminate isolated points after segmentation and merge narrow connected areas), and hole elimination operation (by finding connected areas, the holes in the lesion foreground are eliminated to restore the real cerebral hemorrhage situation) to obtain the final predicted cerebral hemorrhage image Mask. The amount of bleeding is calculated based on the cerebral hemorrhage image Mask of a patient, and it is determined whether cerebral hemorrhage occurs. When the amount of bleeding is greater than the set acceptable error range, it is determined as cerebral hemorrhage and the amount of cerebral hemorrhage is output.

In addition, a packaged auxiliary diagnosis applet can be set up on the WeChat side of mobile phones and other user devices. After the WeChat side of the mobile phone starts the applet, the patient's basic information and the content of the second CT image are interactively input, and the trained model is used to output whether the patient has cerebral hemorrhage and the calculation results of the amount of bleeding, etc.; according to the interactive area selection on the applet page, the location of different bleeding points on the brain CT, the amount of bleeding and other information can be manually enlarged to output the preliminary diagnosis conclusion.

The visualization effect diagram of cerebral hemorrhage detection provided by an embodiment of the present invention is shown in Figure 2, which can accurately locate the bleeding lesions. In the identified image, the white area is the area with cerebral hemorrhage lesions, and the black area is the area without cerebral hemorrhage lesions. At the same time, the conclusion of whether there is cerebral hemorrhage and the amount of bleeding is output.

In practical applications, the present invention can simultaneously perform linkage diagnosis on multiple brain diseases to form a comprehensive and integrated diagnosis and treatment result, breaking through the limitations of independent diagnosis and treatment and limited scope of each department in existing medicine, reducing the probability of repeated medical consultation and examination, and making the diagnosis more accurate and comprehensive.

It will be easily understood by those skilled in the art that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims (5)

1. A neural network-based cerebral hemorrhage auxiliary diagnosis system, characterized by comprising the following modules: An acquisition module is used to acquire multiple original CT images and mark the lesion area; A processing module is used to perform image enhancement on the plurality of original CT images to obtain a training set; the image enhancement means used include: grayscale random jitter, contrast adjustment, multiple noise simulations, filtering operations, and random bright spot superposition, so as to make it closer to the characteristics of the secondary shot image; A training module, used for training a neural network using the training set; The first diagnostic module is used to obtain and preprocess the CT image taken twice, and input the processed image into the trained neural network to obtain a grayscale image representing the confidence of each pixel that is consistent with the size of the input image; the CT image taken twice is a CT image taken stably by the doctor using a mobile phone; The second diagnosis module is used to perform threshold segmentation, morphological closing operation and hole elimination processing on the grayscale image in sequence; and to detect cerebral hemorrhage lesions and calculate the amount of bleeding according to the processed grayscale image; The training module is also used to calculate the loss between the grayscale image output by the neural network and the annotated original CT image according to the selected cross entropy loss function, and use the loss back propagation to update the neural network weight parameters. 2. According to the neural network-based cerebral hemorrhage auxiliary diagnosis system of claim 1, it is characterized in that the training module is also used to import multiple adjacent images into the input layer of the neural network for joint training. 3. According to claim 1, the neural network-based cerebral hemorrhage auxiliary diagnosis system is characterized in that the CT image shot twice is obtained and preprocessed, including: obtaining the CT image shot twice, and performing image center cropping and image compression to make the processed image consistent with the shape and size of the original CT image. 4. The neural network-based cerebral hemorrhage auxiliary diagnosis system according to claim 3 is characterized in that after image center cropping and image compression, the first diagnosis module is also used to grayscale the image. 5. The neural network-based cerebral hemorrhage auxiliary diagnosis system according to claim 1 is characterized in that the neural network is a fully convolutional semantic segmentation neural network.