CN-121983252-A - Cardiac ultrasonic image processing method based on federal learning

CN121983252ACN 121983252 ACN121983252 ACN 121983252ACN-121983252-A

Abstract

The application discloses a heart ultrasonic image processing method based on federal learning, which relates to the technical field of medical image processing and federal learning, and comprises the steps of acquiring a global federal model issued by a server as a local sub-model, training the local sub-model on a local sample set, and outputting a prediction result; constructing local prototype parameters in training and uploading the local prototype parameters to a server, acquiring global prototype parameters obtained by the server based on the local prototype parameters of all clients, calculating a loss function based on a prediction result and the global prototype parameters, updating local sub-model parameters and uploading the local sub-model parameters to the server, aggregating sub-model parameters of each client through the server to update a global federal model, and processing a target heart ultrasonic image through the converged global federal model to obtain an image processing result. The application solves the problem of data isomerism across hospitals through the federal learning technology, and can improve the early detection capability of congenital heart disease on the premise of protecting the privacy of patients.

Inventors

DU BO
WU QINGQING
YE MANG
Lei Wenjia
HUANG WENKE
Liao Yangxu
WEN CHI
LI HE

Assignees

武汉大学
首都医科大学附属北京妇产医院

Dates

Publication Date: 20260505
Application Date: 20260127

Claims (10)

1. A federal learning-based cardiac ultrasound image processing method, characterized by being applied to a client, the method comprising: acquiring a global federation model issued by a server; taking the global federation model as a local sub-model, training the local sub-model on a local sample set, and outputting a prediction result; Constructing and obtaining local prototype parameters in the training process, and uploading the local prototype parameters to the server; acquiring global prototype parameters generated by the server based on local prototype parameter aggregation of all clients, and calculating a loss function based on the prediction result and the global prototype parameters; updating the sub-model parameters of the local sub-model based on the loss function, and uploading the updated sub-model parameters to the server; Carrying out global model aggregation on sub-model parameters of each client through the server to construct a new global federation model; Judging whether the new global federation model is converged or not through the server, and using the converged global federation model for processing the target heart ultrasonic image to obtain an image processing result.
2. The method for processing cardiac ultrasound images based on federal learning according to claim 1, wherein the construction process of the local sample set comprises: acquiring a plurality of local sample heart ultrasonic images; Acquiring the image type of each sample heart ultrasonic image, and extracting various medical concept parameters corresponding to each sample heart ultrasonic image; Obtaining an image type label of each sample heart ultrasonic image based on the image type, and obtaining a medical concept parameter label of each sample heart ultrasonic image based on the various medical concept parameters; the local sample set is constructed based on the plurality of sample cardiac ultrasound images, the corresponding image type tags, and the medical concept parameter tags.
3. The method for processing cardiac ultrasound image based on federal learning according to claim 2, wherein the training the local sub-model on the local sample set, outputting the prediction result, comprises: Processing each sample heart ultrasonic image of a local sample set through a local sub-model, and respectively processing to obtain a prediction result of the sample heart ultrasonic image, wherein the prediction result comprises an image type prediction result and a medical concept parameter prediction result; In the processing process, extracting the characteristics of each sample heart ultrasonic image through the local submodel respectively to obtain an image characteristic vector; and respectively inputting each sample heart ultrasonic image into a function for extracting the medical conceptual parameter feature vector to obtain the medical conceptual parameter feature vector.
4. A federally learned cardiac ultrasound image processing method according to claim 3, wherein the local prototype parameters include medical concept prototype parameters and image feature prototype parameters; the constructing and obtaining the local prototype parameters in the training process comprises the following steps: Constructing and obtaining medical concept prototype parameters based on the medical concept parameter feature vectors, wherein the medical concept prototype parameters comprise positive prototype parameters corresponding to each type of medical concept prototype parameters and negative prototype parameters corresponding to each type of medical concept prototype parameters; And clustering the image feature vectors of each image type based on the image type prediction result, and obtaining the image feature prototype parameters of the current client corresponding to each image type based on the average value of all the image feature vectors under each image type.
5. The method for processing cardiac ultrasound image based on federal learning according to claim 4, wherein the step of obtaining global prototype parameters generated by the server based on local prototype parameter aggregation of all clients comprises: Obtaining local prototype parameters of each client at a server; The method comprises the steps that positive prototype parameters corresponding to each type of medical conceptual parameters learned by each client are aggregated through a server, and global positive prototype parameters corresponding to the type of medical conceptual parameters are obtained; The negative prototype parameters corresponding to each type of medical concept prototype parameters learned by each client are aggregated through a server, so that global negative prototype parameters corresponding to the type of medical concept parameters are obtained; Aggregating the image feature prototype parameters corresponding to each image type learned by each client through the server to obtain global image feature prototype parameters corresponding to each image type The global prototype parameters are issued to each client through a server; The global prototype parameters comprise global negative prototype parameters and global positive prototype parameters of each type of medical conceptual parameters and global image characteristic prototype parameters corresponding to each image type.
6. The method of claim 5, wherein the loss function comprises a default loss term, a concept regularization loss term, a federal clinical concept alignment loss term, and a federal disease prototype comparison loss term, wherein each loss term is weighted and summed to obtain a total loss function, and the formula is applied: ; Wherein, the The loss function is represented by a function of the loss, Representing the default loss term(s) in question, Representing the concept regularization loss term, Represents the federal clinical concept alignment loss term, Representing the federal disease prototype comparative loss term.
7. The federal learning-based cardiac ultrasound image processing method according to claim 6, wherein the calculation process of the default loss term applies the formula: ; ; Wherein, the Representing the default penalty term; a function representing a calculated cross entropy loss; The method comprises the steps of predicting an image type corresponding to an ith sample heart ultrasonic image of a kth client; An image type label of an ith sample cardiac ultrasound image for a kth client; Representing a total number of sample cardiac ultrasound images of the kth client; the image feature vector corresponding to the ith sample heart ultrasonic image of the kth client is obtained; A structure mask preset in the construction process of the local sample set is adopted; representing a function for calculating an attention map.
8. The heart ultrasonic image processing device based on federal learning is characterized by comprising a client and a server: the client is used for acquiring a global federation model issued by the server; The client is also used for taking the global federation model as a local sub-model, training the local sub-model on a local sample set and outputting a prediction result; The client is also used for constructing and obtaining local prototype parameters in the training process and uploading the local prototype parameters to the server; The client is also used for acquiring global prototype parameters generated by the server based on the local prototype parameters of all the clients in an aggregation mode, and calculating a loss function based on the prediction result and the global prototype parameters; The client is further used for updating the sub-model parameters of the local sub-model based on the loss function, and uploading the updated sub-model parameters to the server; the server is also used for carrying out global model aggregation on the sub-model parameters of each client so as to construct a new global federation model; the server is also used for judging whether the new global federation model is converged or not, and the client side is used for processing the target heart ultrasonic image by using the converged global federation model to obtain an image processing result.
9. An electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, characterized in that the processor implements the steps of the method according to any one of claims 1 to 7 when the program is executed by the processor.
10. A non-transitory computer readable storage medium, on which a computer program is stored, characterized in that the computer program, when being executed by a processor, implements the steps of the method according to any one of claims 1 to 7.

Description

Cardiac ultrasonic image processing method based on federal learning Technical Field The application relates to the technical field of medical image processing and federal learning, in particular to a heart ultrasonic image processing method based on federal learning. Background Congenital heart disease (Congenital HEART DEFECT, CHD) is the most common congenital defect, and achieving early CHD detection is critical to an effective intervention plan. Currently, prenatal detection of CHD relies mainly on ultrasound examination during the second pregnancy (18-22 weeks of gestation), this relatively late detection window limiting the choice of intervention, with the possibility of missed detection, which complicates the decision to terminate pregnancy if severe CHD is identified later in gestation, and is detrimental to effective intervention. Thus, advancing the detection window to the first pregnancy (11-14 weeks of gestation) detection represents a critical but unmet need for technology. However, the first pregnancy data is limited in both size and availability, and CHD detection during the first pregnancy depends on the experience of the technician, thereby limiting the effectiveness of CHD detection. Analysis of the first pregnancy data may be performed by aggregating medical samples from multiple institutions to enable technicians to better perform CHD detection during the first pregnancy, but this approach poses serious risks to patient privacy and is subject to significant ethical concerns. Some related technologies can process an ultrasonic image of a first pregnancy through a deep learning model training mode, however, training of the deep learning model for CHD detection faces a plurality of difficulties, on the premise that the scale and the availability of data of the first pregnancy are limited, a single mechanism often lacks a large number of high-quality training samples, more time is required for collecting the training samples and training the model, and the accuracy of the deep learning model for CHD detection cannot be guaranteed. Thus, there is currently a lack of an ultrasound image processing method that enables early CHD detection with patient privacy preservation. Disclosure of Invention The application provides a heart ultrasonic image processing method based on federal learning, which is used for solving the defects of the related technology, and the technical scheme is as follows: In a first aspect, the present application provides a cardiac ultrasound image processing method based on federal learning, applied to a client, the method comprising: acquiring a global federation model issued by a server; taking the global federation model as a local sub-model, training the local sub-model on a local sample set, and outputting a prediction result; Constructing and obtaining local prototype parameters in the training process, and uploading the local prototype parameters to the server; acquiring global prototype parameters generated by the server based on local prototype parameter aggregation of all clients, and calculating a loss function based on the prediction result and the global prototype parameters; updating the sub-model parameters of the local sub-model based on the loss function, and uploading the updated sub-model parameters to the server; Carrying out global model aggregation on sub-model parameters of each client through the server to construct a new global federation model; Judging whether the new global federation model is converged or not through the server, and using the converged global federation model for processing the target heart ultrasonic image to obtain an image processing result. In an alternative of the first aspect, the constructing the local sample set includes: acquiring a plurality of local sample heart ultrasonic images; Acquiring the image type of each sample heart ultrasonic image, and extracting various medical concept parameters corresponding to each sample heart ultrasonic image; Obtaining an image type label of each sample heart ultrasonic image based on the image type, and obtaining a medical concept parameter label of each sample heart ultrasonic image based on the various medical concept parameters; the local sample set is constructed based on the plurality of sample cardiac ultrasound images, the corresponding image type tags, and the medical concept parameter tags. In an alternative aspect of the first aspect, the training the local sub-model on the local sample set, outputting the prediction result, includes: Processing each sample heart ultrasonic image of a local sample set through a local sub-model, and respectively processing to obtain a prediction result of the sample heart ultrasonic image, wherein the prediction result comprises an image type prediction result and a medical concept parameter prediction result; In the processing process, extracting the characteristics of each sample heart ultrasonic image through the local submodel respect