CN-121982294-A - Unsupervised domain adaptive medical image segmentation method based on dynamic topology modeling and frequency domain consistency

Abstract

The invention relates to the field of medical image processing, in particular to an unsupervised domain adaptive medical image segmentation method based on dynamic topology modeling and frequency domain consistency. The method is used for solving the problem of insufficient precision caused by neglecting the spatial topological relation of the anatomical structure, the inter-domain feature distribution difference and the imaging protocol deviation in the cross-domain segmentation of the existing method. The method comprises the following steps of firstly constructing a sparse adjacency matrix through a dynamic matrix characteristic enhancement module, explicitly modeling long-distance space topology dependence, then aligning characteristic distribution of a source domain and a target domain by adopting multi-core maximum mean value difference (MK-MMD) to generate domain invariant characteristics, and finally introducing double consistency constraint based on wavelet transformation and random transformation, and combining frequency domain enhancement and appearance transformation to improve prediction confidence and robustness of a model on a label-free target domain. The method is mainly used for realizing accurate and automatic segmentation of the cross-mode and cross-equipment medical images and assisting clinical diagnosis.

Inventors

• ZHANG HONGLIANG
• ZHANG HONGYING
• WU BIN

Assignees

• 西南科技大学

Dates

Publication Date

20260505

Application Date

20260302

Claims (5)

1. 1. The unsupervised domain adaptive medical image segmentation method based on dynamic topology modeling and frequency domain consistency is characterized by constructing a double-component framework, comprising a dynamic matrix characteristic enhancement module for explicitly modeling anatomical space topology dependence and a wavelet enhancement consistency constraint mechanism for relieving frequency domain deviation, wherein the method specifically comprises five parts of medical image data preprocessing, network foundation construction based on antagonism learning, characteristic enhancement and alignment based on a dynamic matrix, target domain wavelet enhancement consistency constraint and network optimization based on multiple loss functions: The first part comprises two steps: Step 1, downloading public medical datasets MMWHS, brats and Pro12 in a format of nii or nii.gz, wherein MMWHS comprises medical images of CT and MR modes, brats comprises medical images of T2 and Flair modes, pro12 comprises medical images of T2 modes of BIDMC and HK medical centers, and unifying the datasets for normalization operation; Step 2, slicing the image processed in the step 1 along the depth direction, respectively setting the resolutions of the data set samples to 144× 144,128 ×128 and 256×256, and dividing the data set into source domain medical image data sets with pixel-level labels according to different distributions of the training set samples And a label-free target domain medical image dataset Forming a final required training set sample; the second part comprises two steps: Step 3, training the target domain training sample set obtained in the step 2 The random enhancement stream processing is applied, and is implemented by applying random color dithering and Gaussian blur transformation operation to generate random enhancement images To simulate conventional data disturbance and increase data diversity; step 4, training the target domain training sample set obtained in the step 2 Wavelet enhanced stream processing is applied, and the method is concretely implemented as follows: Step 4.1, firstly decomposing an image into 3 stages by using a Haar wavelet basis function; Step 4.2, preserving low frequency approximation coefficients to maintain anatomical structure, linearly enhancing high frequency detail coefficients in horizontal, vertical and diagonal directions, enhancement factors in frequency domain processing Set to 1.5; step 4.3, applying soft threshold function to the enhanced high frequency coefficient to remove high frequency noise, and finally reconstructing the image by inverse wavelet transformation to generate a wavelet enhanced image ; The third part comprises four steps: step 5, the data set obtained in step 2 And In the shared encoder E of the input DeepLabV architecture, the middle layer feature map is extracted And ; Step 6, the intermediate layer characteristic diagram obtained in the step 5 is processed And As an input to the dynamic matrix feature enhancement module DMFE, a feature map enhanced by explicitly modeling long-range spatial topology dependencies of the anatomy, the dynamic matrix feature enhancement module DMFE is shown in fig. 3, and is specifically implemented as follows: step 6.1, calculating a similarity matrix among all the space nodes in the feature map; Step 6.2, utilizing the threshold value Step 6.2, introducing a unit matrix to perform self-connection enhancement, and performing bilateral symmetry normalization on the adjacent matrix; Step 6.4, aggregating the normalized matrix and the original features, and outputting the enhanced feature map And ; Step 7, in the enhanced feature map And On the basis, the distance between the source domain and the target domain features in the regenerated kernel Hilbert space is calculated by using a multi-core maximum mean difference loss function MK-MMD, the number of Gaussian kernels is set to be 5, the bandwidth multiplier is set to be 2.0, and the distance is minimized Explicit alignment of source domain and target domain feature distribution is realized; In the step 8 of the method, the step of, Constructing a discriminator and an encoder to perform countermeasure training, wherein the discriminator identifies characteristic sources, and the encoder optimizes countermeasure loss Generating domain invariant features, and further reducing domain differences; The fourth part comprises a step of: step 9, performing dual consistency constraint based on wavelet transformation and random transformation on the target domain medical image, wherein the method is specifically implemented as follows: Step 9.1, randomly enhancing the images obtained in the step 3 and the step 4 And wavelet enhanced images As inputs to DeepLabV's 2 split network, predictive probability maps are obtained, respectively And ; Step 9.2, calculating the consistency loss between the two by means of a mean square error loss function MSE The constraint forces the network to maintain the consistency of the prediction result for the physical-based wavelet domain enhancement and the geometric-based random transformation, thereby improving the robustness and the prediction confidence of the model to the frequency deviation in the unlabeled target domain; The fifth part comprises three steps: Step 10, defining a total loss function: wherein For the source domain with supervised loss including cross entropy loss and Dice loss, setting network super-parameters, balancing weight 、 、 The segmentation model optimizer is SGD, the learning rate is 2.5X10 −4 , the arbiter optimizer is Adam, the learning rate is 10 −4 , the batch size is 4, and the iteration times are not fixed by adopting an advanced stop strategy; Step 11, inputting the training set sample in the step 2 into the network from the step 3 to the step 10, and obtaining a final medical image segmentation pre-training model by utilizing the combined training of the source domain data and the label-free target domain data; Step 12, inputting the common test set into the pre-training model of step 11, wherein the model can output the fine segmentation result of the medical image.
2. 2. The method for segmenting the unsupervised domain-adaptive medical image based on dynamic topology modeling and frequency domain consistency according to claim 1, wherein when wavelet enhancement stream processing is applied in the step 4, a multi-stage multi-scale decomposition mode is adopted to obtain a low-frequency approximate component and a high-frequency detail component of an image, the obtained high-frequency detail component is subjected to linear enhancement and denoising processing, and the obtained low-frequency approximate component is subjected to retaining operation.
3. 3. The method for segmenting the unsupervised domain-adaptive medical image based on dynamic topological modeling and frequency domain consistency according to claim 1, wherein step 6 is used for constructing a sparse adjacency matrix based on the spatial similarity of the feature map, modeling the spatial topological relation of the anatomical structure is realized, the expressive power of features is enhanced, and the feature map with the spatial structure enhancement is output.
4. 4. The method for unsupervised domain-adaptive medical image segmentation based on dynamic topology modeling and frequency domain consistency according to claim 1, wherein in step 7, the distribution distance of the features of the source domain and the target domain in the projection space is calculated by using a multi-core maximum mean difference loss function, and the explicit distribution alignment of the feature level is realized by minimizing the distance.
5. 5. The method for segmenting the unsupervised domain-adaptive medical image based on dynamic topology modeling and frequency domain consistency according to claim 1, wherein step 9 inputs the image sequences processed by different enhancement strategies into a segmentation network to obtain corresponding prediction results, calculates consistency loss among different prediction results, forces the network to keep stable output under different disturbance, and improves the confidence of the model on unlabeled data.

Description

Unsupervised domain adaptive medical image segmentation method based on dynamic topology modeling and frequency domain consistency Technical Field The invention relates to an image processing technology, in particular to an unsupervised domain adaptive medical image segmentation method based on dynamic topology modeling and frequency domain consistency. Background Medical image segmentation is a vital task in medical image processing that aims to precisely locate and isolate specific anatomical structures or lesion areas from medical images. The deep learning technology has made remarkable progress in medical image segmentation, so that the efficiency and accuracy of acquiring key information from complex medical images are obviously improved. In the prior art, the encoder-decoder model represented by U-Net and DeepLab can automatically extract layering characteristics and realize high-precision segmentation under the scene with a large amount of high-quality annotation data. Existing deep-learning segmentation models are typically based on the assumption that training data (source domain) and test data (target domain) are independently co-distributed. In practical clinical applications, however, medical images often originate from different imaging devices, different medical centers, or employ different imaging protocols. These different sources of data have significant differences in gray scale distribution, texture detail and contrast, and because of the huge time and expert costs required to re-label each new target domain, the dilemma of lack of labeling of target domain data is often faced clinically. At this time, if the model trained only on the source domain is directly applied to the unlabeled target domain data, the segmentation performance thereof will be greatly reduced, and the clinical requirements cannot be satisfied. In such a context, unsupervised domain adaptation (Unsupervised Domain Adaptation, UDA) techniques have evolved, the core goal of which is to exploit knowledge of the source domain to achieve efficient generalization of the model over the target domain, with the target domain completely unlabeled. The current mainstream UDA method is mainly based on countermeasure training, and a segmentation network is prompted to extract domain invariant features with consistent source domain and target domain distribution by introducing a discriminator. Nevertheless, existing countermeasure training strategies have significant limitations in handling unsupervised tasks due to the complexity of the medical anatomy. The conventional countermeasure DANN, cyclegan network focuses mainly on the alignment of global statistical distributions, but ignores the spatial topological relationships inherent to medical anatomies. Without the constraint of the target domain label, the model is very prone to losing key details. SEASA introducing predictive consistency regularization to compensate for the deficiency of the target domain labels, and insufficient predictive confidence of the model on the target domain is caused by insufficient mining of the inherent invariance of the data on the frequency domain. Thus, for the problems presented in the above-mentioned medical image UDA, it is still difficult for the existing single countermeasure or geometric consistency method to meet the requirement of high precision segmentation. Therefore, how to design a medical image segmentation model which can not only explicitly model the anatomical space topological relation, but also overcome imaging protocol deviation through frequency domain consistency has important research significance under the constraint of no labeling of a target domain. Disclosure of Invention The invention aims to solve the problem of weak generalization capability caused by insufficient feature expression and frequency domain characterization deviation when the unsupervised domain adaptive medical image segmentation method faces to unmarked target domain data. The invention provides an unsupervised domain adaptive medical image segmentation method based on dynamic topology modeling and frequency domain consistency, which realizes accurate segmentation of a cross-mode medical image by mining the spatial topology relation and the frequency domain consistency characteristics of data. In order to achieve the above purpose, the invention provides an unsupervised domain adaptive medical image segmentation method based on dynamic topology modeling and frequency domain consistency, which mainly comprises 4 parts, wherein the first part is used for preprocessing a medical image data set, the second part is used for carrying out double-stream data enhancement on training data, the third part is used for carrying out characteristic enhancement and cross-domain distribution alignment based on a dynamic matrix on extracted characteristics, the fourth part is used for carrying out double consistency constraint based on wavelet transformation and random