Non-subtraction DSA blood vessel segmentation method and system based on differential comparison and anatomical prior

Abstract

The invention belongs to the field of medical image processing, and particularly discloses a non-subtraction DSA blood vessel segmentation method based on differential comparison and anatomical prior. In order to solve the problems that in the prior art, DSA blood vessel segmentation is easily interfered by complex backgrounds, and the recognition rate of tiny blood vessels is low, the method comprises the steps that firstly, through a depth difference comparison embedding module, a main encoder is driven to learn blood vessel specific characteristics by means of a pre-subtraction image Xpre and a post-subtraction image Xpost which are registered, and the blood vessel-background difference is enhanced; 2) through an anatomical consistency regularization module, introducing anatomical prior of label-free Xpre to confront learning constraint background feature distribution and suppress artifact interference; and 3) a three-stage dynamic weighted training strategy is adopted to carry out collaborative optimization on the segmentation task and the self-supervision task. The method has the technical effects that the blood vessel segmentation precision (especially tiny blood vessels) is remarkably improved, the robustness to background interference and motion artifacts is remarkably enhanced, and a segmentation result highly matched with a gold standard is realized on a cerebrovascular DSA data set.

Description

Technical Field

[0001] This invention belongs to the field of medical image processing, and specifically discloses a non-subtraction DSA vessel segmentation method based on differential contrast and anatomical prior. Background Technology

[0002] Digital subtraction angiography (DSA) is the cornerstone of vascular imaging, but its core physical subtraction step is highly susceptible to artifacts caused by minute patient movements, reducing image quality and diagnostic accuracy. To circumvent this problem, researchers have turned to deep learning to directly segment blood vessels from post-angiography images containing background. However, this direct segmentation task faces significant challenges: vascular signals are often obscured by complex anatomical backgrounds (such as bones and tissues) and easily confused with artifacts, especially for small vessels. While existing deep learning methods have made progress, there is still considerable room for improvement in distinguishing fine blood vessels from complex backgrounds, fully utilizing paired pre- and post-angiography image information, and a large number of unlabeled background images (masked images) to enhance the model's understanding of the background and robustness. Therefore, there is an urgent need for an innovative technology that can more intelligently fuse multi-source DSA image information, strengthen vascular feature learning, and suppress background interference to achieve high-precision and highly robust non-subtraction vascular segmentation. Summary of the Invention

[0003] To address the aforementioned problems, this invention discloses a non-subtraction DSA vessel segmentation method based on differential contrast and anatomical prior. Through innovative differential contrast feature learning, anatomical prior regularization, and intelligent collaborative training strategies, this invention significantly improves the accuracy, robustness, and generalization ability of non-subtraction DSA vessel segmentation, demonstrating significant advantages over existing technologies, especially in challenging scenarios such as handling microvessels, complex backgrounds, and data scarcity.

[0004] This invention includes the following technical solutions:

[0005] A non-subtraction DSA vessel segmentation method based on differential contrast and anatomical prior includes the following steps:

[0006] S1. Image preprocessing: Obtain paired pre-subtraction images X pre and subtraction image X post Perform format conversion, intensity normalization, size standardization, and image registration;

[0007] S2, Depth Differential Contrast Embedding: The registered X... pre and X post The main encoder, which receives shared parameters, is used to enhance X through contrastive learning. post Mid-vascular characteristics and inhibition with X pre Consistent background features;

[0008] S3. Anatomical Consistency Regularization: Utilizing Label-Free X pre Train the independent anatomical encoder and constrain the main encoder to X through adversarial learning. post The output features of the background region are consistent with the feature distribution of the anatomical encoder.

[0009] S4. Blood vessel segmentation output: Input the optimized multi-scale features into the decoder to generate a blood vessel probability map and binarize it to obtain a segmentation mask.

[0010] Furthermore, in the above-mentioned non-subtraction DSA vessel segmentation method, the intensity normalization in step S1 uses the following parameters:

[0011] The lower truncation limit is the 0.5 percentile P of the pixel intensity in the dataset. 0.5 ;

[0012] The upper limit of the cutoff is the 99.5th percentile P. 99.5 .

[0013] Furthermore, in the above-mentioned non-subtraction DSA vessel segmentation method, the image preprocessing in step S1 includes:

[0014] Convert DICOM format to NIfTI format;

[0015] Apply window width and window level adjustments and linearly map pixel values ​​to the range [0, 255].

[0016] Truncated minimum-maximum normalization based on dataset statistics;

[0017] Using free-form deformation algorithm for X pre and X post Perform subpixel-level registration.

[0018] Furthermore, in the above-mentioned non-subtraction DSA vessel segmentation method, the window width and window level are set to 800HU and 300HU respectively.

[0019] Furthermore, in the above-mentioned non-subtraction DSA vessel segmentation method, the depth difference contrast embedding in step S2 specifically includes:

[0020] Extracting multi-layer feature maps from the main encoder and

[0021] L2-normalized embedding vectors are generated using the projection head. and

[0022] by As the anchor point, its enhanced view is a positive sample. For negative samples, calculate the multilayer InfoNCE loss to optimize the master encoder.

[0023] Furthermore, in the above-mentioned non-subtraction DSA vessel segmentation method, the anatomical consistency regularization in step S3 specifically includes:

[0024] From the main encoder features Predicted vascular probability map

[0025] according to Extracting background features

[0026] Driven by discriminator Features of Anatomical Encoders Distribution alignment.

[0027] Furthermore, in the above-mentioned non-subtraction DSA vessel segmentation method, the vessel segmentation output in step S4 uses a U-Net architecture decoder to fuse features through the following operations:

[0028] Upsample deep features and concatenate them with the corresponding encoder features;

[0029] The convolutional unit generates a blood vessel probability map with the same resolution as the input.

[0030] Furthermore, the aforementioned non-subtraction DSA vessel segmentation method employs a three-stage dynamic weighted training strategy:

[0031] The first stage mainly uses depth difference contrastive embedding loss and anatomy consistency regularization loss;

[0032] The second phase gradually increases the weighting of monitoring and segmentation losses.

[0033] The third phase focuses on fine-tuning by monitoring and segmenting losses.

[0034] This invention also discloses a non-subtraction DSA vessel segmentation system, comprising:

[0035] A preprocessing module is configured to perform the image preprocessing steps as described in any one of claims 1-8;

[0036] The main encoder module integrates depth differential contrast embedding and anatomical consistency regularization functions;

[0037] The segmentation decoder module is used to generate and binarize the blood vessel probability map.

[0038] The present invention also discloses a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described in the present invention.

[0039] In some embodiments, preferably, the method of the present invention includes the following steps (overall framework as follows) Figure 1 shown):

[0040] S1. Image Preprocessing (DSA Acquisition and Preprocessing)

[0041] The basic data for training the DCE-AINet model is a sequence of digital subtraction angiography (DSA) images, mainly including pre-subtraction images (X-rays). pre (as a mask image) and subtracted image (X) post (i.e., the image after contrast agent injection). X pre The images accurately record the patient's static anatomical background before contrast agent injection, such as bones, soft tissue structures, and any air gaps present. post The image was captured after an X-ray-proof contrast agent was injected into the blood vessel; therefore, it simultaneously contains the vascular network highlighted by the contrast agent (i.e., the segmented target) and the X-ray... pre The same static background anatomical structures in the image. The core objective of this invention, DCE-AINet, is to directly extract from raw, un-physically subtracted X-rays... post The vascular structures were accurately segmented from the images. To ensure high consistency, standardization, and information fidelity of the DSA image data input to the DCE-AINet model, thus providing a solid and reliable foundation for model training, all original acquired images underwent a rigorous and comprehensive standardized preprocessing procedure. This procedure mainly includes the following steps: First, the original DICOM format DSA data was converted to the NIfTI image format, which is easier for deep learning frameworks to process. Second, since the dynamic range of pixel intensity values ​​in the original DSA images may be large, and there may be differences under different devices and acquisition protocols, the window width and window level (WW / WL) commonly used in clinical practice were applied for adjustment. For example, for cerebral vascular DSA images, the parameters were set to WW = 800HU, WL = 300HU, and then the original pixel values ​​P... in The first step is to linearly map the pixel intensity to an 8-bit grayscale range of [0, 255]. The second step is to perform intensity normalization. Specifically, based on the statistical characteristics of the pixel intensity distribution of the entire training dataset, the 0.5 percentile (P0.5) of the dataset is used. 0.5 ) and the 99.5th percentile (P 99.5 The minimum-maximum normalized upper and lower bounds for truncation are:

[0042] X ′ =clamp((XP 0.5 ) / (P 99.5 -P 0.5 ),0,1)

[0043] Then, the preprocessed image (or image patch) is resized to a preset standard size: 512x512 pixels for 2D DSA images and 128x128x128 voxels for 3D DSA images. If the original image size does not match the standard size, bilinear interpolation is used for scaling. For images much larger than the standard size, a sliding window-based patch extraction strategy is used to crop multiple overlapping image patches from the original image as input to the model. The overlap rate is typically set to 25% to 50% to ensure the continuity of contextual information. After patch extraction or image scaling, if blank areas appear at the edges, they should be filled using a symmetrical filling method. Next, image registration is performed, a step specifically for paired X... pre and X post Images. Because patients may move slightly between two acquisitions (mask and contrast images), the background anatomical structures may exhibit spatial displacement and deformation in the two images. Specifically, a B-spline-based Free-Form Deformation (FFD) registration algorithm is used, with the goal of achieving X... ′ pre With X post Subpixel-level precise localization is achieved on shared background anatomical structures. Finally, to enhance the model's generalization ability and robustness to various imaging variations, online data augmentation strategies are applied, including: geometric transformations such as random small-angle rotations (e.g., in the range of [-20, 20] degrees), random scaling (e.g., scale factors between [0.8, 1.2]), random translations (within the range of [-15%, 15%] of the image's width and height), random shearing transformations (shearing angles between [-10, 10] degrees), and nonlinear elastic deformation (by applying a random displacement field to the image with parameters such as control point spacing sigma = 5 and displacement intensity alpha = 40). Intensity and appearance transformations are performed, such as randomly adjusting image brightness (multiplicative factors between [0.7, 1.3]), randomly adjusting contrast (multiplicative factors between [0.6, 1.4]), gamma correction (gamma values ​​between [0.8, 1.2] to simulate different display characteristics), adding Gaussian noise (standard deviation of 0.05, simulating electronic noise), applying Gaussian blur (blur kernel standard deviation between [0.5, 2.0] pixels, simulating focus misalignment or tissue scattering), image sharpening (using the Laplacian operator or Unsharp Mask sharpening), and simulating slight motion blur (by applying blur kernels of specific orientation and length). Domain-specific enhancements are also performed to suit the characteristics of DSA imaging, simulating contrast agent concentration variations and slight misregistration or motion artifacts. After these preprocessing steps, the subsequent model uses the DSA dataset.

[0044] S2, Depth Differential Contrast Embedding (using the Depth Differential Contrast Embedding (DDCE) module, as per the workflow) Figure 2 As shown):

[0045] The goal of the DDCE module is to drive the main encoder network (E main Learning about X post The deep feature representation of vascular structures is highly selectively sensitive and capable of recognizing and suppressing background and artifacts.

[0046] First, X, which has undergone preprocessing in step S1, is... pre (Plain background image) and X post (Background + Blood Vessel Image) Samples are input to the main encoder E. main (ResNet-50, top layer removed), the output of each intermediate layer k (e.g., conv2_x, conv3_x, conv4_x, conv5_x) is denoted as F. (k) For the k-th layer, through the shared parameter E main The leading part yields the feature map of this layer:

[0047]

[0048] Secondly, these feature maps and Feed into the feature projection head The structure is: 1x1 convolutional layer (adjusted to 512 channels) -> Batch Normalization -> ReLU -> GAP -> MLP (two fully connected layers, 512 -> 256 -> 128 dimensions) -> L2 Normalization. The resulting embedding vector is:

[0049]

[0050] Next, construct positive and negative sample pairs and calculate the difference contrast loss. For anchor points... Its positive samples By analyzing X post Applying strong data augmentation operator sets The random transformation T in (such as CutMix, MixUp, GridShuffle, Cutout, and large-amplitude geometric photometric transformations) generates X. ′ post =T(X) post ), and then through and The embedding is obtained. The core negative sample is the paired X. pre Embedded Other negative samples include other X samples within the batch. post Embedding and its enhanced view embedding, as well as other X preEmbedding. The negative sample set is Using InfoNCE loss:

[0051]

[0052] Where sim(u,v)=u T v, τ k This is a temperature parameter (e.g., 0.07 or 0.1). Symmetrical calculation. The total loss of DDCE is the sum of all K ddce Selected levels (e.g., K) ddce =4) Weighted sum of losses:

[0053]

[0054] weight w k This can increase with increasing hierarchy (e.g., w1 = 0.1, w2 = 0.2, w3 = 0.3, w4 = 0.4). This is achieved by minimizing... Optimize E main and The parameters enable specific amplification of vascular signals, suppression of complex backgrounds and artifacts, decoupling of multi-scale features and semantic enhancement, and robustness improvement against subtraction imperfections (such as motion artifacts).

[0055] S3. Anatomical Consistency Regularization (using the Anatomical Consistency Regularization (AIR) module, see workflow) Figure 3 )

[0056] Furthermore, we utilize the AIR module to leverage a large number of label-free X... pre Pure anatomical information improves the main encoder E main Feature representation of the background indirectly improves the accuracy of vessel segmentation. The operation steps are as follows: First, an independent auxiliary anatomical encoder E is introduced. anatomy (Architecture and E) main Same,

[0057] Parameters are trained independently, specifically from a large number of X... pre Learn the anatomical background features without vascular interference. anatomy After selecting an intermediate level s(conv2_x,conv3_x), the output is denoted as... The feature map of ) is considered the "gold standard" background anatomical feature. Secondly, from the main encoder E main Processing X post During this process, its feature representation of the background is dynamically estimated. Specifically, firstly, for E... main Feature maps output from different network layers A preliminary vascular probability map at this level is predicted using a matrix consisting of two 3x3 convolutional layers, one 1x1 convolutional layer, and a sigmoid activation function. Then, X is extracted using this probability graph. post Background-related features in:

[0058]

[0059] Here, ⊙ represents element-wise multiplication. Furthermore, adversarial learning at the feature level is employed to force... Distribution and The distribution is aligned. A discriminator network is introduced for each level s. (A small CNN, including convolution, LeakyReLU, spectral normalization, and a scalar output with GAP+FC at the end). The training objective is to maximize the differentiation from "real" data. and "forgery" The ability. Loss.

[0060]

[0061] Master Encoder E main (and The optimization goal of the generation part is to generate something that can "deceive" others. of Adversarial Generative Loss

[0062]

[0063] Total Anatomical Consistency Regularization Loss It is all selected S air Weighted sum of losses at each level:

[0064]

[0065] weight α s Average settings (e.g., α1 = 0.5, α2 = 0.5). The AIR module is jointly trained with DDCE and the main segmentation task, which improves the stability and refinement of background representation, enhances the identification and suppression of background artifacts, promotes the decoupling of foreground and background features, improves the model's generalization ability, and improves the performance of few-shot learning.

[0066] S4. Vessel Segmentation Output

[0067] This phase integrates the optimized main encoder E from S2 and S3. main Output multi-scale features via blood vessel segmentation decoder D seg Reconstruct pixel-level blood vessel segmentation maps. Specifically, firstly, the blood vessel segmentation decoder D... segThe U-Net decoding path structure is adopted, containing upsampling modules with the same number of stages as the encoder. Each upsampling level k (from deep to shallow): feature upsampling (bilinear interpolation + convolution) takes the output of the previous decoder level. Spatial resolution is doubled. Feature fusion combines the upsampled decoder features with E... main Optimized feature map corresponding to level k splicing along the channel dimension:

[0068]

[0069] The concatenated feature maps are fed into the decoding convolutional unit (DCU, containing two standard convolutional modules: 3x3 convolution + normalization + activation) for processing. The final output layer of the decoder (original image resolution) generates a blood vessel probability map S through a 1x1 convolution (outputting 1 channel) and a sigmoid activation function. post Secondly, the training strategy and loss function are jointly optimized. Total loss. It consists of three weighted parts: supervised segmentation loss In labeled samples The above calculation uses combined loss:

[0070] 1. Focal Loss:

[0071] 2. Dice Loss:

[0072]

[0073] Deep difference contrastive embedding loss Calculated according to S2. Anatomical consistency regularization loss. Calculated according to S3. Total loss function:

[0074]

[0075] Loss weight λ seg (t),λ ddce (t),λ air (t) is dynamically adjusted with training period t, employing a three-stage learning strategy: warm-up and feature pre-learning stage (first 20-30% epochs, high λ) ddce ,λ air low λ seg Task adaptation and multi-task collaborative learning phase (40-50% of epochs, gradually increasing λ). seg ); fine-tuning of segmentation performance (e.g., the last 20-30% of epochs, high λ) seg low λ ddce ,λ airThe model uses the AdamW optimizer (initial learning rate 1×10⁻⁶). -4 weight decay 1×10 -5 Cosine annealing learning rate scheduling, training for hundreds of epochs, batch size determined by GPU memory. Mixed precision training can be enabled. After training, new images... The probability graph obtained through forward propagation of the model The optimal threshold T determined on the validation set is applied. thresh Binarization yields the segmentation mask.

[0076]

[0077] After the model training is completed, its main encoder E main and blood vessel segmentation decoder D seg Parameters are fixed. For new pre-subtraction DSA images (and optional) First, perform the same standardized preprocessing as S1, but without data augmentation, to obtain... (and Subsequently, the preprocessed image is fed into the main encoder E. main Extracting multi-scale features During this process, the DDCE module no longer calculates the loss, and the auxiliary components of the AIR module are typically not activated because the main encoder has already internalized its learned capabilities. Next, Input vessel segmentation decoder D seg Through upsampling and skip connections, the final output is a blood vessel probability map of the same size as the input. Finally, for The optimal threshold T determined on the validation set. thresh Binarization is performed to obtain the segmentation mask.

[0078] Optionally, morphological post-processing (such as opening operations for noise reduction and closing operations for hole filling) can be performed to optimize the results. If a patch strategy is used, the results of each patch need to be stitched and fused. In this way, an accurate vessel segmentation map can be automatically generated for the new DSA image.

[0079] Compared with the prior art, the present invention has the following outstanding advantages:

[0080] This patent proposes a non-subtraction DSA vessel segmentation method and system, whose core advantages are:

[0081] 1. Depth Differential Contrast-Driven Enhancement of Vessel-Specific Features: Unlike traditional methods that directly learn pixel mappings, the DDCE module (S2) of this patent innovatively utilizes paired pre-subtraction (X) images. pre ) and after subtraction (X) post The image, through self-supervised differential contrast learning, forces the main encoder to focus on the difference between the two, namely vascular signals. This allows the model to actively identify and amplify weak, blurry vascular features, while effectively suppressing those with X. pre The consistent complex background structure enables higher sensitivity and specificity in capturing vascular signals, and remains robust even under subtraction problems or artifact interference.

[0082] 2. Anatomical Consistency Regularization-Guided Background Robustness Modeling: The AIR module (S3) cleverly introduces a large number of unlabeled X objects. pre Pure anatomical priors in the image. Through an adversarial learning mechanism, the main encoder is forced to learn X... post At the same time, its feature representation of the background part is consistent with the "gold standard" anatomical background features. This not only deeply stabilizes the model's understanding of complex backgrounds and prevents it from misidentifying background noise or complex anatomical structures as blood vessels, but more importantly, it significantly improves the model's generalization ability and adaptability to unknown background changes in different patients and under different imaging conditions, solving the pain point of performance degradation caused by background variation in many methods.

[0083] 3. Intelligent Training Paradigm of Multi-Task Collaborative Optimization and Curriculum Learning: This patent jointly optimizes the blood vessel segmentation task (S4) with the self-supervised and regularized tasks of DDCE (S2) and AIR (S3) within a unified framework. Through a carefully designed three-stage curriculum learning strategy, the model first fully learns the essence of feature representation (the differences in blood vessels and the consistency of the background) under the guidance of auxiliary tasks, and then gradually shifts its focus to the segmentation task itself. This intelligent training paradigm of "learning commonalities first, then attacking special features" enables the model to converge to a better solution more efficiently and effectively alleviates the problem that the model may get stuck in local optima or become overly dependent on labeled data when directly performing end-to-end supervised learning, especially when labeled data is limited.

[0084] 4. In-depth mining and efficient utilization of raw DSA information: This patented solution fully explores the value of the inherent information in the DSA imaging process. DDCE utilizes X... post With X pre AIR utilizes a large amount of unlabeled X data to perform sophisticated differential learning on pairwise relationships. pre Generalization learning based on background priors. This multi-dimensional and in-depth utilization of raw data information goes far beyond simply applying X... preAs an alternative input channel or a simple differential method, it maximizes the information value of the pre-subtraction DSA image, providing a new and more fundamental solution to the highly challenging problem of non-subtraction vessel segmentation.

[0085] In summary, this patent significantly improves the accuracy, robustness, and generalization ability of non-subtraction DSA vessel segmentation through innovative differential contrast feature learning, anatomical prior regularization, and intelligent collaborative training strategies. It demonstrates significant advantages over existing technologies, especially in challenging scenarios such as handling microvessels, complex backgrounds, and scarce data. Attached Figure Description

[0086] Figure 1 This is a schematic diagram of the overall architecture of a non-subtraction DSA vessel segmentation system according to the present invention;

[0087] Figure 2 This is a schematic diagram of the Deep Differential Contrast Embedding (DDCE) module.

[0088] Figure 3 This is a schematic diagram of the Anatomical Consistency Regularization (AIR) module.

[0089] Figure 4 Here is an example of a DSA image and segmentation results. Detailed Implementation

[0090] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0091] Example

[0092] This embodiment uses a typical post-DSA image of the brain (such as...) Figure 4 The X shown on the left post ) and its paired pre-subtraction image X pre Taking this as an example, the complete processing flow of the DCE-AINet method of the present invention is described in detail.

[0093] Step 1: DSA Image Acquisition and Preprocessing (S1) Raw acquired X post and X pre The images are in DICOM format. First, they are converted to NIfTI format. Next, clinical window settings (WW = 800 HU, WL = 300 HU) are applied to adjust the image contrast, and pixel values ​​are linearly mapped to the [0, 255] range. Then, based on statistics from the entire training dataset (P... 0.5 and P99.5 The image is then intensity-normalized, scaling the effective pixel values ​​to the [0,1] range. Subsequently, the image size is uniformly adjusted to 512x512 pixels. A crucial step is to employ a B-spline-based Freeform Deformation (FFD) algorithm to modify the X... pre and X post Precise registration is performed to eliminate minor patient movements between acquisitions and ensure spatial alignment of background anatomical structures. During model training, a series of online data augmentations, including random rotation, scaling, translation, elastic deformation, and brightness / contrast adjustments, are applied to this registered image pair. During inference (application) phase, only standard preprocessing is performed without data augmentation.

[0094] Steps 2 and 3: Feature extraction and optimization (S2, S3) The preprocessed image pairs (X... post ,X pre The input is fed into the already trained DCE-AINet model.

[0095] 1. Main encoder E main Feature extraction: Image pairs are processed by the main encoder E with shared parameters. main .

[0096] 2. The role of the DDCE module: Because the model is subjected to deep difference contrastive embedding loss during training ( ) optimization, main encoder E main It has learned to identify and magnify X. post and X pre The difference information between them. When X post When (background + blood vessels) pass through, the regions corresponding to blood vessel signals are significantly activated and enhanced in the deep feature map. And when X... pre When a (pure background) model passes through, only the background structure is activated in its feature map. This "differential" effect at the feature level makes the model highly specific for blood vessels, while naturally suppressing the activation of structures related to X. pre Consistent background signals of bones, soft tissues, etc.

[0097] 3. The role of the AIR module: Simultaneously, because the model is subject to anatomical consistency regularization loss during training... Constraints, Master Encoder E main In processing X post When dealing with non-vascular background regions, the generated feature representations are forced in distribution to be consistent with a large number of unlabeled Xg regions. pre The "gold standard" anatomical background features learned from the images remain consistent. This makes the model extremely robust to understanding the background, accurately identifying and filtering out complex anatomical textures, noise, and even slight motion artifacts, preventing them from being misidentified as blood vessels.

[0098] Step 4: Blood vessel segmentation, decoding, and result output (S4) via E main Optimized multi-scale feature map It was fed into the blood vessel segmentation decoder D seg The decoder effectively fuses deep semantic information ("this is a blood vessel") with shallow spatial detail information ("the precise edges of the blood vessel are here") through a series of upsampling operations and skip connections with encoder features. Finally, the decoder outputs a blood vessel probability map S of the same size as the input image. post The value of each pixel on this probability map is between [0,1], representing the probability that the point belongs to a blood vessel.

[0099] Finally, the probability map is binarized by applying a pre-determined optimal threshold to the validation set to generate the final binary segmentation mask. like Figure 4 As shown, from left to right: Original X post The images, the "gold standard" (Ground Truth) DSA image after silhouette reconstruction, the segmentation result output by nnUnet, and the segmentation result output by the method of this invention are shown. It can be seen that the output result of this invention ( Figure 4 (Rightmost) and Gold Standard ( Figure 4 The data (in Chinese) showed a high degree of concordance, accurately outlining the complete vascular tree of the aortic arch, common carotid artery, internal carotid artery, and even small intracranial branches. The background was clean, with no obvious bone artifacts remaining, and the details of tiny blood vessels were well preserved, demonstrating the effectiveness of this method.

[0100] The above are merely a few preferred embodiments of the present invention, described in a relatively specific and detailed manner, but they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention.

Claims

1. A non-subtraction DSA vessel segmentation method based on differential contrast and anatomical prior, characterized in that, Includes the following steps: S1. Image preprocessing: Obtain paired pre-subtraction images X pre and subtraction image X post Perform format conversion, intensity normalization, size standardization, and image registration; S2, Depth Differential Contrast Embedding: The registered X... pre and X post The main encoder, which receives shared parameters, is used to enhance X through contrastive learning. post Mid-vascular characteristics and inhibition with X pre Consistent background features; S3. Anatomical Consistency Regularization: Utilizing Label-Free X pre Train the independent anatomical encoder and constrain the main encoder to X through adversarial learning. post The output features of the background region are consistent with the feature distribution of the anatomical encoder. S4. Blood vessel segmentation output: Input the optimized multi-scale features into the decoder to generate a blood vessel probability map and binarize it to obtain a segmentation mask.

2. The method according to claim 1, characterized in that, The intensity normalization in step S1 uses the following parameters: The lower truncation limit is the 0.5 percentile P of the pixel intensity in the dataset. 0.5 ; The upper limit of the cutoff is the 99.5th percentile P. 99.5 .

3. The method according to claim 1, characterized in that, The image preprocessing in step S1 includes: Convert DICOM format to NIfTI format; Apply window width and window level adjustments and linearly map pixel values ​​to the range [0, 255]. Truncated minimum-maximum normalization based on dataset statistics; Using free-form deformation algorithm for X pre and X post Perform subpixel-level registration.

4. The method according to claim 3, characterized in that, The window width and window position are set to 800HU and 300HU respectively.

5. The method according to claim 1, characterized in that, The depth difference comparison embedding in step S2 specifically includes: Extracting multi-layer feature maps from the main encoder and L2-normalized embedding vectors are generated using the projection head. and by As the anchor point, its enhanced view is a positive sample. For negative samples, calculate the multilayer InfoNCE loss to optimize the master encoder.

6. The method according to claim 1, characterized in that, The anatomical consistency regularization in step S3 specifically includes: From the main encoder features Predicted vascular probability map according to Extracting background features Driven by discriminator Features of Anatomical Encoders Distribution alignment.

7. The method according to claim 1, characterized in that, The blood vessel segmentation output in step S4 uses a U-Net architecture decoder, and features are fused through the following operations: Upsample deep features and concatenate them with the corresponding encoder features; The convolutional unit generates a blood vessel probability map with the same resolution as the input.

8. The method according to claim 1, characterized in that, The method employs a three-stage dynamically weighted training strategy: The first stage mainly uses depth difference contrastive embedding loss and anatomy consistency regularization loss; The second phase gradually increases the weighting of monitoring and segmentation losses. The third phase focuses on fine-tuning by monitoring and segmenting losses.

9. A non-subtraction DSA vessel segmentation system, characterized in that, include: A preprocessing module is configured to perform the image preprocessing steps as described in any one of claims 1-8; The main encoder module integrates depth differential contrast embedding and anatomical consistency regularization functions; The segmentation decoder module is used to generate and binarize the blood vessel probability map.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by a processor, it implements the steps of the method according to any one of claims 1-8.

Images