CN-121120732-B - Time sequence brain image calibration method based on multitasking countermeasure learning

Abstract

A time sequence brain image calibration method based on multi-task countermeasure learning belongs to the technical field of medical images and comprises the following steps of preprocessing a pre-acquired time sequence brain image to remove irrelevant information, dividing a brain image phenotype calibration network into two subtasks (local detail reconstruction and global structure transformation), learning in different modes respectively, and restraining and generating a better calibration brain image by adopting a discriminator in a countermeasure learning strategy. The invention is very valuable for improving the correlation performance of brain images and genes, and can effectively observe and understand the state of human brain.

Inventors

• WANG MEILING
• LIU QINGSHAN

Assignees

• 南京邮电大学

Dates

Publication Date

20260505

Application Date

20250820

Claims (1)

1. 1. A time sequence brain image calibration method based on multitasking countermeasure learning is characterized by comprising the following specific processes: (1) Preprocessing a pre-acquired time sequence brain image to remove irrelevant information; (2) Maintaining consistency between calibration brain image and template brain image generated from constraint on pixel level, and constructing reconstruction loss term ; The step (2) of reconstructing the loss term The method comprises the following steps: ; Wherein, it is assumed that And Respectively represent the calibration brain images Time brain images, all of which are of the size Here, where 、 Representing the length and width of the brain image phenotype, respectively; is an identical network and is encoded by an encoder And decoder The input and output parts are the same template brain image phenotype; (3) Constructing a feature transformation loss term from regression loss of the brain image phenotype characteristic at time n to the template brain image phenotype characteristic ; The step (3) is characterized by transforming loss terms The method comprises the following steps: ; Wherein, the And Respectively representing the length and the width of the characteristic brain image; A representation feature transformer; (4) Constraint network generates more real calibration brain image phenotype, and constructs counterdamage item ; The countermeasures against loss term of the step (4) The method comprises the following steps: ; Wherein, the Is a transformation network and is encoded by an encoder Characteristic transformer And decoder Composition, input and output are respectively A moment brain image phenotype and a corresponding calibration brain image phenotype, Representing a arbiter in the antagonism network; (5) Dividing a brain image phenotype calibration network into two subtasks, wherein the two subtasks are local detail reconstruction and global structure transformation respectively, learning in different modes respectively, and restraining and generating a better calibration brain image by adopting a discriminator in an countermeasure learning strategy; The implementation process of the step (5) is as follows: (51) According to step (2), step (3) and step (4), assume that And Respectively represent the calibration brain images A moment in time brain image, And Representing a generator and a arbiter, respectively, in a countermeasure network architecture, in the proposed calibration method the generator comprises 3 sub-modules, an encoder Characteristic transformer Decoder Wherein, the encoder is responsible for encoding the detail information of the input brain image phenotype, and the feature encoding is responsible for The feature expression of the brain image phenotype in the hidden space at the moment is mapped into the feature space corresponding to the calibration brain image phenotype, and the decoder is responsible for reconstructing the calibration brain image according to the feature expression, and in summary, the loss function of the proposed missing feature filling method comprises a plurality of loss terms, namely: ; Wherein, the And Is a weight parameter that balances the importance of each item; (52) Training two generators simultaneously in a multitasking manner And Wherein Is an identical network and is encoded by an encoder And decoder The input and output parts are the same template brain image phenotype, and Is a transformation network and is encoded by an encoder Characteristic transformer And decoder Composition, input and output are respectively A moment brain image phenotype and a corresponding calibration brain image phenotype, And In (a) encoder And decoder Is parameter-sharing to share the ability to encode/decode detailed information, the network constrains the feature transformer by a feature regression loss term Learning slave Mapping of moment brain image phenotype feature space to template brain image phenotype feature space, and without considering how to encode detail information, a discriminator is also introduced to distinguish the calibrated brain image phenotype generated by the generator from the template brain image phenotype, thereby constraining the generator to generate a better calibrated brain image phenotype.

Description

Time sequence brain image calibration method based on multitasking countermeasure learning Technical Field The invention belongs to the technical field of medical images, and particularly relates to a time sequence brain image calibration method based on multitasking countermeasure learning. Background As an important research content of complex brain state change analysis, gene and brain image association analysis is a typical multidisciplinary cross subject combining the machine learning field, the neural image and the gene sequencing technology, and needs the participation of scientific researchers in different fields, including mathematics, computers, biology and the like. The association analysis of the genes and the brain images can mine the association relation and evolution rule between the neural image and the gene data and the brain state, and find the biomarker related to the brain state, thereby providing data-dependent explanation for the development mechanism of the complex brain state and providing basic support for realizing how the genes affect the neural structure and function of the brain. Clinically, the time sequence brain image data can reflect and reveal the change condition of the brain structure and function along with time, including the changes of the whole brain volume, gray matter volume, ventricle size, frontal temporal lobe cortex thickness and surface area, hippocampus and amygdala volume and the like. Time sequence brain image and gene association analysis effectively reveals the influence mechanism of genes on brain structure or function change in the disease process by establishing association relations between brain images and genes at different moments, and can more comprehensively understand the essence and individual difference of brain diseases. Considering that patients receive imaging scans from different devices or doctors at different points in time, it is likely that time-series brain image data will exhibit heterogeneity in time-series, i.e., heterogeneity between phenotypes. This heterogeneity is not only due to inherent differences in the image acquisition, processing and reconstruction techniques of the imaging device itself, but is further exacerbated by the specific parameters selected by the person during scanning, the maintenance calibration state of the device, and the operational differences that may exist. The time sequence heterogeneity of the time sequence brain images can bring a plurality of challenges to the training and prediction of the subsequent association analysis model, such as the reduction of the stability of the model, the difficulty of training and the like. Therefore, the invention provides a time sequence brain image calibration method based on multi-task countermeasure learning, which divides a brain image phenotype calibration network into two subtasks (firstly, local detail reconstruction and global structure transformation), learns in different modes respectively, and constrains and generates a better calibration brain image by adopting a discriminator in a countermeasure learning strategy. According to the invention, the brain image phenotypes at different moments are calibrated into the same structural space, so that the heterogeneity of time sequence brain image data is reduced, the method is very valuable for improving the correlation performance of brain images and genes, and the states of human brains can be effectively observed and understood. Compared with the prior art, the method has the following differences: compared with the technology of patent CN 112308833A' one-shot brain image segmentation method based on cycle consistency correlation 1. The patent CN112308833A aims at brain image segmentation tasks, adopts an LT-NET network model, and improves segmentation efficiency by matching forward mapping and backward mapping with supervision loss. Aiming at a time sequence brain image calibration task, the brain image phenotype calibration network is divided into two subtasks of local detail reconstruction and global structure transformation, and a better calibration image is generated by adopting an countermeasure learning strategy so as to improve the correlation performance of the brain image and genes. 2. Patent CN112308833a focuses on improving the one-way correlation learning performance to optimize the image segmentation effect. The invention optimizes the local and global characteristics through multi-task learning respectively, focuses on the calibration of time sequence brain images, and can more effectively observe and understand the state of human brain. Compared with the technology of patent CN 112348786B' one-shot brain image segmentation method based on bidirectional correlation 1. Patent CN112348786B builds an image transformation model for brain image segmentation to learn the bi-directional mapping, constraining the forward mapping by the backward mapping to improve the forward mapping accuracy. Aiming at