CN-121982309-A - Full heart CT image segmentation method and system for coping with contrast agent induced domain offset

CN121982309ACN 121982309 ACN121982309 ACN 121982309ACN-121982309-A

Abstract

The invention provides a full heart CT image segmentation method and system for coping with contrast agent induced domain offset, comprising the following steps of S1, obtaining a heart CT angiography three-dimensional image to be processed and preprocessing, S2, inputting the preprocessed image into a coding path of a segmentation model to extract intermediate features, executing space consistency compensation operation on the intermediate features, inputting the features subjected to space consistency compensation into a decoding path of the segmentation model to obtain an initial segmentation prediction result and corresponding deep features, S3, introducing an uncertainty modeling mechanism based on an evidence theory to the initial segmentation prediction result, constructing a reliability mask, S4, under the constraint of the reliability mask, carrying out sample-level online calibration by utilizing the deep features output in the step S2 to obtain a sample-level calibration prediction result, and carrying out weighted fusion with the initial segmentation prediction result to obtain a final segmentation result. The invention can realize stable segmentation of the invisible contrast mode under the condition of no target domain labeling.

Inventors

PAN LIN
XU CHANGJUN
YANG MINGJING
HUANG LIQIN

Assignees

福州大学

Dates

Publication Date: 20260505
Application Date: 20260121

Claims (8)

1. A method for segmenting a full-heart CT image in response to contrast agent induced domain shift, comprising the steps of: s1, acquiring a heart CT angiography three-dimensional image to be processed and performing preprocessing operation; S2, inputting the preprocessed image into a coding path of a segmentation model based on a three-dimensional convolutional neural network, extracting intermediate features representing heart anatomy structure space and semantic information, executing space consistency compensation operation on the intermediate features, and inputting the features subjected to space consistency compensation into a decoding path of the segmentation model to obtain an initial segmentation prediction result and corresponding deep features; in the training stage, the segmentation model performs on-line domain statistics coordination operation on the characteristics after the space consistency compensation, and takes the characteristics after the on-line domain statistics coordination as the input of a segmentation model decoding path; s3, introducing an uncertainty modeling mechanism based on an evidence theory to construct a reliability mask for the initial segmentation prediction result; S4, under the constraint of a reliability mask, carrying out sample-level online calibration by utilizing the deep features output in the step S2 to obtain a sample-level calibration prediction result, and carrying out weighted fusion on the sample-level calibration prediction result and the initial segmentation prediction result to obtain a final segmentation result.
2. A method of whole-heart CT image segmentation against contrast-induced domain shifts according to claim 1, wherein said preprocessing operations include resampling, intensity normalization and heart region cropping operations of the image.
3. The method of claim 1, wherein the performing spatial consistency compensation on the intermediate features includes parallel modeling of local detail enhancement branches and global guidance branches and residual reinjection, and is characterized by: Wherein, the Representing the intermediate characteristics of the features, The parameters of the convolution kernel are represented, Representing a three-dimensional convolution operation, An example normalization operation is shown and, Representing a non-linear activation function, 、、 The results of the intermediate processing are respectively the results of the intermediate processing, Is a feature after spatial consistency compensation.
4. The method for whole-heart CT image segmentation with contrast-induced domain offset according to claim 1, wherein the performing of the online domain statistical coordination on the spatial consistency compensated features is as follows: Spatial consistency compensated features for current samples Calculating an instance mean value corresponding to the current sample according to the channel dimension Standard deviation from the examples And on-line maintaining global statistics by adopting an exponential moving average mode to obtain global mean variance Global variance ; And introducing random disturbance to generate a characteristic re-parameterization coefficient: Wherein, the Beta represents a characteristic heavy parameterized bias term introduced in an online domain statistical coordination process and is used for adjusting the mean value position of the characteristic; The feature re-parameterization scale adjustment item introduced in the online domain statistics coordination process is shown and used for adjusting the scale amplitude of the feature; And (3) with Respectively expressed by corresponding variances 、 Taking the square root to obtain the statistical fluctuation amplitude; For characteristics of Executing normalization and affine transformation to obtain output characteristics after on-line domain statistics coordination: Wherein, the Representing the features after the online domain statistics reconciliation, Smoothing constants to prevent numerical instability.
5. The method for whole-heart CT image segmentation as recited in claim 4, wherein the global statistics are maintained online by exponential moving average to obtain a global mean variance Global variance The method is characterized by comprising the following steps: maintaining first order statistics of example statistics of characteristic mean values by adopting exponential moving average mode And second order statistics : Wherein the method comprises the steps of Is an exponential moving average coefficient; calculating global mean variance according to the maintained first-order and second-order statistics: Maintaining first order statistics of example statistics of characteristic standard deviation in exponential moving average mode And second order statistics And calculates the global variance: 。
6. The method for segmenting a full heart CT image against contrast agent induced domain shift according to claim 1, wherein for the initial segmentation prediction result, an uncertainty modeling mechanism based on evidence theory is introduced to construct a reliability mask, specifically as follows: For any voxel position, set Representing any element In the first place Initial predicted logic values on the individual segmentation classes, the corresponding evidence amounts are defined as: Wherein, the Representing voxels In the first place The corresponding evidence strengths are predicted on the individual segmentation classes, Adjusting parameters for evidence mapping, wherein the parameters are used for controlling the numerical scaling degree when predicting the mapping of logit to evidence quantity; The number of the divided categories; constructing Dirichlet distribution parameters based on evidence quantities : Defining voxel level uncertainty metrics In which a single voxel Uncertainty of (2) Expressed as: constructing a reliability mask by combining the prediction category information: Wherein, the A category index representing the segmentation task; for indicating the function, the value is 1 when the condition in the brackets is satisfied, otherwise, the value is 0; represent the first Reliability masks corresponding to the individual segmentation classes.
7. The method for segmenting a full heart CT image against contrast agent induced domain shift according to claim 6, wherein the sample-level online calibration is performed by using the deep features output in step S2 under the constraint of a reliability mask, and a sample-level calibration prediction result is obtained, specifically as follows: For deep features under reliability mask constraints Weighted aggregation to construct the first Dynamic classifier weight vector corresponding to each segmentation class : Wherein, the Represent the first Dynamic classifier weight vectors for each of the segmentation classes; representing a voxel index; Representing deep features at the first Feature vectors at voxel locations; represent the first Reliability mask corresponding to each partition category is at the first A value at the voxel location; based on the dynamic classifier weight vector, the deep features are classified into the deep features Classifying to obtain a sample-level calibration prediction result : 。
8. A whole heart CT image segmentation system responsive to contrast induced domain shifts, comprising a processor, a memory and a computer program stored on said memory, said processor, when executing said computer program, performing in particular the steps of the whole heart CT image segmentation method as claimed in any of claims 1-7.

Description

Full heart CT image segmentation method and system for coping with contrast agent induced domain offset Technical Field The invention belongs to the technical field of image segmentation, and particularly relates to a full heart CT image segmentation method and system for coping with contrast agent induced domain offset. Background At present, the full heart CT image segmentation technology is to segment various anatomical structures such as a left atrium, a right atrium, a left ventricle, a right ventricle, a cardiac muscle, an ascending aorta, a pulmonary artery and the like through end-to-end learning by adopting a three-dimensional full convolution network and various variants thereof (such as 3D U-Net, nnU-Net and an improved network architecture based on the three-dimensional full convolution network) aiming at a CTA (CT angiography) image. In order to improve the cross-center generalization capability of a segmentation method, the prior art method is generally improved from the following directions of 1, enhancing a data/style layer, simulating domain differences through modes of normalization, intensity disturbance, style migration and the like, 2, geometry topology constraint, introduction of shape prior, topology consistency or geometry similarity learning to relieve instability caused by complex heart structure and fuzzy boundary, 3, generating modeling, utilizing a diffusion generating framework to jointly model an image and mask distribution to adapt to different appearances, and 4, self-adapting in test, carrying out entropy minimization, self-supervision task or batch normalization (Batch Normalization, BN) statistical calibration and the like update by using unmarked test data in an reasoning stage to adapt to a target domain. However, the above technique still suffers from the significant disadvantage of a multi-centered CTA scene of "contrast induced domain shift", defect 1-lack of explicit modeling of the complex enhancement mode caused by the "contrast injection protocol/pass timing/scanning device differences". In the actual cross center, significant gray level statistical drift and local enhancement non-uniformity can occur in the same anatomical structure, thereby generating prediction bias. It is often difficult to cover statistical fluctuations caused by true enhancement dynamics with style disturbance alone. The defect 2:3D heart segmentation is often limited by a video memory, the training batch size is smaller, the existing characteristic enhancement and normalization method which depends on the current small batch statistics is easy to generate deviation, global domain statistics is difficult to reflect, and the cross-domain generalization difficulty is high. Defect 3. Local structure boundaries are susceptible to contrast agent inhomogeneities and are "blurred and broken", and it is often difficult to maintain good structural continuity and clear boundary details on fine blood vessels and thin-walled structures at the same time by means of global style alignment or geometric constraints only. Defect 4 existing adaptation at test methods typically rely on back-propagation update model parameters, or build on a single hypothesis (e.g., high confidence predictions or BN calibratability). In the medical image segmentation task, the method is easy to have the problems of instability, high calculation cost, negative migration and the like, and reliable region self-adaption for a single sample is difficult to realize. In view of the above, the present invention proposes a method and a system for segmenting whole heart CT images in response to contrast agent induced domain shifts. Disclosure of Invention The invention aims to provide a full heart CT image segmentation method and system for coping with contrast agent induced domain offset, which follow the general technical thought of generalization during training and self-adaption cooperation during testing, and realize stable segmentation of an unobserved contrast mode under the condition of no target domain labeling. In order to achieve the above purpose, the technical scheme of the invention is as follows: a full heart CT image segmentation method for coping with contrast agent induced domain shift specifically comprises the following steps: s1, acquiring a heart CT angiography three-dimensional image to be processed and performing preprocessing operation; S2, inputting the preprocessed image into a coding path of a segmentation model based on a three-dimensional convolutional neural network, extracting intermediate features representing heart anatomy structure space and semantic information, executing space consistency compensation operation on the intermediate features, and inputting the features subjected to space consistency compensation into a decoding path of the segmentation model to obtain an initial segmentation prediction result and corresponding deep features; in the training stage, the segmentation model performs on-lin