CN113435607B - Disease screening method based on federal learning - Google Patents

Abstract

The invention discloses a disease screening method based on federal learning, which comprises the following steps: 1) Constructing a shared data set; 2) Pre-training; 3) Calculating an accuracy average value; 4) Calculating the data quantity; 5) And (5) convergence. The invention belongs to the technical field of federal learning and deep learning, in particular to a disease screening method based on federal learning, which combines knowledge of federal learning and deep learning, can more fully utilize data of various places and can efficiently and accurately diagnose diseases.

Description

A disease screening method based on federated learning

technical field

The invention belongs to the technical field of federated learning and deep learning, and specifically refers to a disease screening method based on federated learning.

Background technique

A medical image is an image that reflects the internal structure or internal function of an anatomical region, and it is composed of a group of image elements—pixels (2D) or voxels (3D). Medical images are discrete image representations generated by sampling or reconstruction, which can map values to different spatial locations. The number of pixels is used to describe medical imaging under a certain imaging device, and it is also a way of expressing the details of anatomy and its function. The specific value expressed by the pixel is determined by the imaging equipment, imaging protocol, image reconstruction and post-processing. But sometimes even as many as a thousand images are examined per patient, and even more, assuming a concentrated outbreak of disease. Relying solely on doctor's diagnosis leads to slow efficiency. Federated learning combines deep learning to process image data, and can comprehensively utilize data to quickly analyze images and make accurate diagnoses.

The techniques closest to the present invention are:

1. Image classification algorithm based on deep learning: Image classification is to distinguish different types of images according to the semantic information of images, which is an important basic problem in computer vision. In deep learning, the Convolution Neural Network (CNN) is mainly used for image classification. The pixel information of the image is used as input, and the feature extraction and high-level abstraction are performed through the convolution operation. The model output is directly the result of image recognition. Currently common image classification CNN networks include Lenet, Alxnet, Vgg series, Resnet series, Inception series, Densenet series, Googlenet, etc. However, this method cannot comprehensively utilize the data of hospitals in various places for comprehensive processing, resulting in low accuracy and poor performance.

2. The DeCoVNet network is a weakly supervised learning algorithm. It is of heavier practical significance in cases where the distribution of disease cases is uneven and the overall number is small. It has the characteristics of fast calculation speed and less requirements for data labeling, but it still has the disadvantage of low accuracy so that misdiagnosis is easy to occur.

Contents of the invention

In view of the above situation, in order to overcome the defects of the existing technology, the present invention provides a disease screening method based on federated learning, which combines the knowledge of federated learning and deep learning, can make full use of data from various places, and perform disease diagnosis efficiently and accurately.

The technical scheme that the present invention takes is as follows: a kind of disease screening method based on federated learning of the present invention, comprises the following steps:

1) The server builds a shared data set, and sends the initial model (U-Net++, DeCoVNet) to the participating training clients;

2) When the client runs the first epoch, use the unsupervised learning training to obtain the marked mask and the marked training set to pre-train the U-Net++ model; after all the training sets are pre-trained, U -Net++ model to obtain the masks of all training sets; all the masks, training sets and labels of all training sets enter the judgment model DeCoVNet to train the training set; all the masks of the test set and the test set enter the judgment model DeCoVNet for testing; calculation Local accuracy, and upload to the server;

3) The server collects the accuracy of the clients participating in the training and calculates the mean value of the accuracy;

4) Before running the second epoch, if the precision of a certain client is less than the average precision, send m pieces of data to the client. The formula is as follows: m=i×n, where i=(acc _avg -acc _i )/(acc _avg -acc _min ), n is the size of the shared data set, and the distributed data set should not be larger than the data volume of the client itself ;

5) From the second epoch onwards, the client with the shared data set will train the model together with the shared data set and its own data set until the overall model converges. .

The beneficial effects obtained by the present invention by adopting the above-mentioned structure are as follows: This scheme is a disease screening method based on federated learning, which establishes a model for medical images through a deep learning convolutional neural network, effectively improving the accuracy of disease diagnosis; using federated learning to comprehensively utilize various Hospital node data to improve the generalization ability of the model; a dynamic fusion strategy is proposed to improve the overall accuracy of the system.

Description of drawings

Fig. 1 is a block diagram of a federated learning fusion method for a disease screening method based on federated learning in the present invention;

Fig. 2 is a block diagram of a medical image deep learning method for a disease screening method based on federated learning in the present invention.

The accompanying drawings are used to provide a further understanding of the present invention, and constitute a part of the description, and are used together with the embodiments of the present invention to explain the present invention, and do not constitute a limitation to the present invention.

Detailed ways

The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only some of the embodiments of the present invention, not all of them; based on The embodiments of the present invention and all other embodiments obtained by persons of ordinary skill in the art without making creative efforts belong to the protection scope of the present invention.

As shown in Figure 1-2, the disease screening method based on federated learning of the present invention includes the following steps:

1) The server builds a shared data set, and sends the initial model (U-Net++, DeCoVNet) to the participating training clients;

2) When the client runs the first epoch, use the unsupervised learning training to obtain the marked mask and the marked training set to pre-train the U-Net++ model; after all the training sets are pre-trained, U -Net++ model to obtain the masks of all training sets; all the masks, training sets and labels of all training sets enter the judgment model DeCoVNet to train the training set; all the masks of the test set and the test set enter the judgment model DeCoVNet for testing; calculation Local accuracy, and upload to the server;

3) The server collects the accuracy of the clients participating in the training and calculates the mean value of the accuracy;

4) Before running the second epoch, if the precision of a certain client is less than the average precision, send m pieces of data to the client. The formula is as follows: m=i×n, where i=(acc _avg -acc _i )/(acc _avg -acc _min ), n is the size of the shared data set, and the distributed data set should not be larger than the data volume of the client itself ;

5) From the second epoch onwards, the client with the shared data set will train the model together with the shared data set and its own data set until the overall model converges.

In specific use, the user first sets up a shared data set, and then in the first epoch, after the client performs n iterations, the accuracy is uploaded to the server, and the server collects all the accuracy of the client participating in the training, and calculates the mean value. If the precision of the i-th client is lower than the average precision, calculate the data volume of the shared data to be distributed (the data volume is not greater than the data volume of the client itself), and if it is higher than the average precision, the shared data value will not be issued. Finally, the clients with the shared dataset keep training for the remaining epochs until the model converges. In view of the characteristics of densely populated cities that lead to a large number of daily diagnostic medical images, it will cause doctors to face the situation that the judgment speed of many medical images is slow, so there is an urgent need for artificial intelligence models to help doctors judge diseases. We propose a model to use the U-Net++ model to segment the medical image (judging the disease area), and then use the judgment model (DeCoVNet) to judge whether there is a disease in the segmented part. The above is the overall workflow of the present invention. The following Repeat this step for the first use.

It should be noted that in this article, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply that there is a relationship between these entities or operations. any such actual relationship or order exists between them. Furthermore, the term "comprises", "comprises" or any other variation thereof is intended to cover a non-exclusive inclusion such that a process, method, article, or apparatus comprising a set of elements includes not only those elements, but also includes elements not expressly listed. other elements of or also include elements inherent in such a process, method, article, or apparatus.

Although the embodiments of the present invention have been shown and described, those skilled in the art can understand that various changes, modifications and substitutions can be made to these embodiments without departing from the principle and spirit of the present invention. and modifications, the scope of the invention is defined by the appended claims and their equivalents.

The present invention and its implementations have been described above, and this description is not limiting. What is shown in the drawings is only one of the implementations of the present invention, and the actual structure is not limited thereto. All in all, if a person of ordinary skill in the art is inspired by it, without departing from the inventive concept of the present invention, without creatively designing a structure and an embodiment similar to the technical solution, it shall fall within the scope of protection of the present invention.

Claims (1)

1. A disease screening method based on federal learning, characterized in that; the method comprises the following steps:

1) The server builds a shared data set and transmits the initial model to the training client;

2) When a client runs a first epoch, performing pre-training on a U-Net++ model by using an unsupervised learning training to obtain a mask with a mark and a training set with the mark; obtaining masks of all training sets by the U-Net++ model after the training sets are pre-trained; the mask of all training sets, the training sets and the labels enter a judgment model DeCoVNet together, and training is carried out on the training sets; the masks of all the test sets and the test sets enter a judgment model DeCoVNet for testing; calculating local precision and uploading the local precision to a server;

3) The server collects the accuracy of the clients participating in training and calculates an accuracy average value;

4) Before the second epoch is operated, if the precision of a certain client is smaller than the precision average value, m pieces of data are issued to the client, and the formula is as follows: m=i×n, where i= (acc) _avg -acc _i )/(acc _avg -acc _min ) N is the size of the shared data set, and the issued data set is not larger than the data volume of the client;

5) After the second epoch, the client with the issued shared data set trains the model together with the own data set until the whole model converges.