An intelligent craniocerebral vascular wall image processing system
By employing a hierarchical modulation system and an adaptive vascular feature perception design, the problems of noise deviation and loss of fine-grained features in the cranial vascular wall image processing system are solved, thereby improving processing accuracy and robustness and adapting to image processing capabilities under different conditions.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- HANGZHOU PANORAMIC MEDICAL IMAGING DIAGNOSIS CO LTD
- Filing Date
- 2026-03-30
- Publication Date
- 2026-07-07
AI Technical Summary
Existing brain vascular wall image processing systems are prone to introducing noise, deviating from the characteristics of the vascular wall, and losing fine-grained features, resulting in low processing accuracy. Furthermore, local fine-grained features are easily interfered with, local feature invariance is ignored, information is not fully utilized, and processing effect is poor.
A hierarchical modulation system was designed, which focuses on preserving the core features of the vascular wall through light modulation and simulates real clinical variations through heavy modulation. A dual-pool directional modulation system was constructed, and an adaptive vascular feature perception weight was introduced to design the total loss of heavy modulation, avoid interference, and improve adaptability and robustness.
It improves the accuracy and effectiveness of cranial vascular wall image processing, enhances adaptability to different scanning equipment, patients and acquisition conditions, and ensures the stability and robustness of fine-grained feature extraction.
Smart Images
Figure CN121962838B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of image processing, specifically to an intelligent image processing system for the cerebral blood vessel wall. Background Technology
[0002] A brain vascular wall image processing system is an image processing system used for image preprocessing, feature extraction, lesion detection, or segmentation and classification of medical images related to brain vascular walls. However, general brain vascular wall image processing systems suffer from problems such as easily introducing noise, deviating from the characteristics of the vascular wall, and easily losing fine-grained features, leading to low processing accuracy. Furthermore, these systems often suffer from the susceptibility of local fine-grained features to interference, neglect of local feature invariance, and insufficient information utilization, resulting in poor processing performance. Summary of the Invention
[0003] To address the aforementioned issues and overcome the shortcomings of existing technologies, this invention provides an intelligent cranial vascular wall image processing system. Addressing the problems of conventional cranial vascular wall image processing systems, such as the easy introduction of noise, deviation from vascular wall characteristics, and loss of fine-grained features, leading to low processing accuracy, this solution designs a layered modulation system. Light modulation focuses on preserving the core features of the vascular wall, while heavy modulation focuses on simulating real clinical variations. The heavy modulation constructs a dual-pool directional modulation system to improve the adaptability of cranial vascular wall image processing to different scanning equipment, patients, and acquisition conditions. By constructing a dual-layer modulation system, a secondary modulation is performed on the lightly modulated image to cover variations at the feature level. This approach avoids background interference leading to inaccurate extraction of vessel wall features, thereby improving the accuracy of cranial vascular wall image processing. Addressing the issues of poor processing results caused by the susceptibility of local fine-grained features to interference, neglect of local feature invariance, and insufficient information utilization in general cranial vascular wall image processing systems, this solution employs a two-level coordination constraint to enable the system to learn the essential features of the vessel wall under changes in grayscale stretching, blurring, and elastic deformation, thus enhancing its fine-grained robustness to clinical variations. An adaptive vascular feature perception weight design is introduced to remodulate the total loss, avoiding interference from remodulation on the learning of basic vascular wall features, ensuring the stability of fine-grained vascular feature extraction, and ultimately improving the image processing performance of cranial vascular walls.
[0004] The technical solution adopted by the present invention is as follows: The present invention provides an intelligent cranial blood vessel wall image processing system, including an image acquisition module, a cranial blood vessel wall image processing model design module, a label-free set modulation strategy design module, a secondary modulation module, a remodulation feature fusion module, a model training optimization module, and a real-time cranial blood vessel wall image processing module.
[0005] The image acquisition module acquires clinical images of the cerebral blood vessel walls, and after preprocessing, constructs labeled and unlabeled sets.
[0006] The brain vascular wall image processing model design module is based on preprocessed clinical brain vascular wall images. It designs the model architecture and processing flow, which includes feature extraction, label-free set modulation, secondary modulation, remodulation feature fusion, training optimization and final output.
[0007] The unlabeled set modulation strategy design module generates lightly modulated and heavily modulated images for the unlabeled set.
[0008] The secondary modulation module performs secondary modulation on the lightly modulated image;
[0009] The remodulation feature fusion module fuses the features and original predictions of the two remodulated images respectively;
[0010] The model training and optimization module designs loss functions for labeled and unlabeled images, and trains and optimizes the cranial blood vessel wall image processing model to complete the establishment of the cranial blood vessel wall image processing model.
[0011] The real-time cranial blood vessel wall image processing module performs real-time cranial blood vessel wall image processing based on the established cranial blood vessel wall image processing model.
[0012] Furthermore, the image acquisition module acquires clinical cerebral vascular wall HR-MRI images. The labeled set consists of cerebral vascular wall images labeled pixel by pixel with three categories of tags: vascular wall, lumen, and background. The unlabeled set consists of unlabeled cerebral vascular wall images of the same modality. Preprocessing is performed, including resampling to a uniform voxel size, grayscale normalization, and ROI cropping.
[0013] Furthermore, the brain vascular wall image processing model design module is based on preprocessed clinical brain vascular wall images and adopts a lightweight encoder-decoder architecture. The encoder is responsible for extracting multi-scale features, and the decoder is responsible for feature upsampling and pixel-level segmentation prediction. The output layer adopts the Softmax activation function and outputs a pixel-level three-class segmentation probability map. The model processing flow includes: feature extraction - unlabeled set modulation - secondary modulation - remodulation feature fusion - training optimization - final output.
[0014] Furthermore, the unlabeled set modulation strategy design module specifically includes:
[0015] Lightly modulated image generation: A slightly randomized image is generated by applying Gaussian noise to an unlabeled image;
[0016] Remodulation dual-pool generation: Construct two remodulation pools to generate remodulated images for unlabeled images.
[0017] Furthermore, the secondary modulation module specifically includes:
[0018] Feature extraction; features are extracted from the slightly modified image using an encoder;
[0019] Secondary modulation operation: Simultaneously perform path dimension and domain dimension modulation on the extracted features, with each modulation operation assigned to an independent path, generating three modulated features: including: path dimension modulation; domain dimension net noise modulation; and domain dimension masking modulation.
[0020] Generating pseudo-labels for the lightly modulated image; generating independent predictions from the three modulated features using the decoder; filtering the original predictions of the lightly modulated image using a confidence threshold, and retaining pixels with probability values not lower than the confidence threshold as reliable pseudo-labels;
[0021] Consistency-based loss is used; reliable pseudo-labels are used to guide the prediction of the three paths, and a loss function is designed.
[0022] Furthermore, the readjustment feature fusion module specifically includes:
[0023] Remodulation feature extraction: The encoder extracts features from the two remodulated images, and then the decoder generates the original prediction.
[0024] Feature fusion; Alpha fusion is performed on the two remodulated features; the fused features are then decoded to generate predictions;
[0025] The prediction results are fused; and the original predictions of the two remodulations are fused with the same proportion of alpha.
[0026] Two-level coordination constraints: global layer constraint, using reliable pseudo-labels from light modulation to guide the original predictions of the two remodulations, and designing a local layer constraint loss for the remodulation; local layer constraint, using mean squared error to constrain the coordination between the fused feature predictions and the fused prediction results.
[0027] The total loss is retuned; by integrating global and local layer constraints and introducing an adaptive vascular feature sensing function, a retuned total loss is obtained.
[0028] Furthermore, the model training optimization module specifically includes:
[0029] Guided loss design calculates the error between the labeled image and the true label;
[0030] Without guided loss, for unlabeled images, the lightly modulated loss and the heavily modulated two-level constraint loss are fused to obtain the total model loss;
[0031] Training optimization strategy design: The AdamW optimizer is adopted, and the learning rate is scheduled using a cosine annealing strategy; the Dice similarity coefficient of the validation set is used as an indicator, and training is stopped if there is no improvement for 20 consecutive epochs; thus, the image processing model of the cranial blood vessel wall is established.
[0032] Furthermore, the real-time cranial blood vessel wall image processing module applies the established cranial blood vessel wall image processing model to clinically unlabeled cranial blood vessel wall images and designs a standardized processing flow.
[0033] The beneficial effects achieved by the present invention using the above solution are as follows:
[0034] (1) To address the problems of general cranial vascular wall image processing systems, such as the easy introduction of noise, deviation from vascular wall characteristics, and loss of fine-grained features, which leads to low processing accuracy, this solution designs a layered modulation system. Light modulation focuses on preserving the core features of the vascular wall, while heavy modulation focuses on simulating real clinical variations. Heavy modulation constructs a dual-pool directional modulation system to improve the adaptability of cranial vascular wall image processing to different scanning equipment, different patients, and different acquisition conditions. By constructing a dual-layer modulation system, secondary modulation is performed on the basis of light modulation images to cover feature-level variations. Background interference is avoided, which leads to inaccurate extraction of vascular wall features. This improves the accuracy of cranial vascular wall image processing.
[0035] (2) To address the problems of poor processing results caused by the susceptibility of local fine-grained features to interference, neglect of local feature invariance, and insufficient information utilization in general cranial vascular wall image processing systems, this solution uses a two-level coordination degree constraint to enable the system to learn the essential features of the vascular wall under changes in grayscale stretching, blurring, and elastic deformation, thereby improving the fine-grained robustness to clinical variations; it introduces an adaptive vascular feature perception weight design to remodulate the total loss, avoiding interference of remodulation on the learning of basic vascular wall features; it ensures the stability of fine-grained vascular feature extraction; and thus improves the image processing effect of cranial vascular wall. Attached Figure Description
[0036] Figure 1 A flowchart illustrating an intelligent cranial blood vessel wall image processing system provided by the present invention;
[0037] Figure 2 Images of unlabeled cerebral blood vessel walls;
[0038] Figure 3 This is a processed image of the cerebral blood vessel wall.
[0039] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. Detailed Implementation
[0040] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.
[0041] In the description of this invention, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.
[0042] Example 1, see Figure 1 The present invention provides an intelligent cranial blood vessel wall image processing system, including an image acquisition module, a cranial blood vessel wall image processing model design module, a label-free set modulation strategy design module, a secondary modulation module, a remodulation feature fusion module, a model training optimization module, and a real-time cranial blood vessel wall image processing module.
[0043] The image acquisition module acquires clinical images of the intracranial blood vessel walls, preprocesses them to construct labeled and unlabeled sets, and then sends the data to the intracranial blood vessel wall image processing model design module.
[0044] The brain vascular wall image processing model design module designs the model architecture and processing flow based on preprocessed clinical brain vascular wall images. The processing flow includes feature extraction, unlabeled set modulation, secondary modulation, remodulation feature fusion, training optimization, and final output; and sends the data to the unlabeled set modulation strategy design module.
[0045] The unlabeled set modulation strategy design module generates lightly modulated and heavily modulated images for the unlabeled set; and sends the data to the secondary modulation module.
[0046] The secondary modulation module performs secondary modulation on the lightly modulated image and sends the data to the remodulation feature fusion module.
[0047] The readjustment feature fusion module fuses the features and original predictions of the two readjustment images respectively, and sends the data to the model training and optimization module.
[0048] The model training and optimization module designs loss functions for labeled and unlabeled images, trains and optimizes the cranial blood vessel wall image processing model, and completes the establishment of the cranial blood vessel wall image processing model; and sends the data to the real-time cranial blood vessel wall image processing module.
[0049] The real-time cranial blood vessel wall image processing module performs real-time cranial blood vessel wall image processing based on the established cranial blood vessel wall image processing model.
[0050] Example 2, see Figure 1 This embodiment is based on the above embodiment. The image acquisition module acquires clinical cranial blood vessel wall HR-MRI images. The labeled set consists of cranial blood vessel wall images labeled pixel by pixel with three categories of labels: blood vessel wall (including lesion area), lumen, and background. The unlabeled set consists of unlabeled cranial blood vessel wall images of the same modality. Preprocessing is performed, including resampling to a uniform voxel size, grayscale normalization, and ROI cropping, retaining only the area from the skull base to the middle cerebral artery, and removing invalid background.
[0051] Example 3, see Figure 1 This embodiment is based on the above embodiment. The brain vascular wall image processing model design module is based on preprocessed clinical brain vascular wall images and adopts a lightweight encoder-decoder architecture to balance fine-grained feature extraction of brain vascular wall images with model inference efficiency. The encoder is responsible for extracting multi-scale features, and the decoder is responsible for feature upsampling and pixel-level segmentation prediction. The model architecture includes: the encoder uses ResNet-34 as the basic backbone, removes fully connected layers, and adds dilated convolution modules (dilation rate = 2 / 4) to expand the receptive field while retaining fine-grained features of the vascular wall and avoiding the loss of thin structural features; the decoder adopts a nested upsampling structure of U-Net++, introduces cross-layer feature fusion, and repairs the lumen-wall boundary features lost during encoder downsampling; the output layer uses the Softmax activation function to output pixel-level three-class segmentation probability maps (vascular wall / lumen / background); the model processing flow includes: feature extraction - unlabeled set modulation - secondary modulation - remodulation feature fusion - training optimization - final output.
[0052] Example 4, see Figure 1 This embodiment is based on the above embodiment. The unlabeled set modulation strategy design module generates a lightly modulated version (minor modulation, preserving the core features of the blood vessel wall, used to generate reliable pseudo-labels) and a heavily modulated version (targeted modulation, simulating various variations in clinical images, used to improve model robustness) for the unlabeled set. The modulation strategy perfectly matches the image features of the cranial blood vessel wall, avoiding feature distortion caused by general modulation. Specifically, it includes:
[0053] Slightly modulated image generation: Slight random cropping + Gaussian noise (low intensity) is applied to unlabeled images to generate slightly modulated images; low intensity Gaussian noise simulates scanning noise of the blood vessel walls of the brain, and slight cropping simulates positional shift of image acquisition. These are all minor variations that are common in clinical practice, and the pseudo-labels have high reliability.
[0054] Remodulation dual-pooling generation: Construct two remodulation pools to generate remodulation images from unlabeled images. and To meet the dual requirements of grayscale adaptation and robustness of tubular structures, both operations unrelated to the cerebral vascular wall are eliminated in general modulation: remodulation pool 1 ( ) Perform adaptive grayscale stretching (to adapt to the grayscale heterogeneity of the cerebral vascular wall), Gaussian blur (Gaussian kernel size 3×3 / 5×5, Gaussian kernel standard deviation σ∈[0.5,1.0], simulating the difference in scan resolution), and region blending (CutMix, only blending vascular regions); readjustment pool 2 ( Perform gamma correction (correction coefficient 0.8~1.2, adapting to the gray distribution of cerebral blood vessel walls), elastic deformation (elastic deformation smoothness coefficient 0.5~1.0, simulating slight morphological deformation caused by vascular pulsation), and region blending (CutMix, blending only vascular regions).
[0055] Example 5, see Figure 1 This embodiment is based on the above embodiment. The secondary modulation module is designed to avoid the loss of fine-grained features of the vessel wall and insufficient robustness of domain-dimensional features in the extraction of intracranial vessel wall features due to single-path modulation. It performs both path and domain-dimensional modulation on the encoder features of the lightly modulated image and employs multi-path independent guided training to ensure that each modulation dimension effectively improves the model's adaptability to variations in vessel wall features while maintaining compatibility with the baseline method. Specifically, it includes:
[0056] Feature extraction; lightly adjusted images via encoder Feature extraction ;
[0057] Secondary modulation operation; for feature extraction Simultaneously perform link dimension and domain dimension modulation, with each modulation operation assigned to an independent path, generating three modulated features; including: link dimension modulation, using two-dimensional channel Dropout (nn.Dropout2d), randomly discarding 10%~30% of channels to simulate random missing features in the link, represented as: Domain-level noise modulation for feature generation and extraction. A noise tensor N of the same size (U(−0.2,0.2) uniformly distributed) is used to avoid excessive disturbance to the fine-grained features of the lightly activated pipe wall, and is expressed as: ; Domain-dimensional masking modulation; First extract features Summing and normalizing along the path dimension yields the domain surface activation graph. Regenerate the mask Randomly masking 10%-30% of low-activation background regions forces the model to focus on the core features of the blood vessel wall, as shown below: ;in, It is the feature map after the connection dimension is modulated; It is a link dimension modulation operation; It is a feature map after the net noise modulation of the domain plane dimension; It is the activation threshold; It is a uniformly distributed random number; It is the feature map after the domain dimension masking modulation;
[0058] The pseudo-label is generated by light modulation; three modulation features—connection dimension modulation, domain dimension net noise modulation, and domain dimension masking modulation—are generated independently and predicted by the decoder h(·), and are denoted as follows: , and Original prediction for lightly modulated images After confidence threshold (Values range from 0.8 to 0.9) Filtering: Pixels with a probability value not lower than the confidence threshold are retained as reliable pseudo-labels, denoted as... Filter out noisy pixels with low confidence;
[0059] Coordination-constrained loss; reliable pseudo-labels guide the prediction of the three paths respectively, and the loss function is expressed as: ; It is a slight adjustment that results in a loss; It refers to the batch size; It is an indicator function; it returns 1 if the condition is true, and 0 otherwise. yes The maximum value of the three probabilities for a single pixel; and This is the path weight coefficient, with a value ranging from 0.5 to 1.0; It is a Dice loss.
[0060] By performing the above operations, this solution addresses the problems of general cranial vascular wall image processing systems, such as the easy introduction of noise, deviation from vascular wall characteristics, and loss of fine-grained features, leading to low processing accuracy. It designs a layered modulation system: light modulation focuses on preserving the core features of the vascular wall, while heavy modulation focuses on simulating real clinical variations. Heavy modulation constructs a dual-pool directional modulation system to improve the adaptability of cranial vascular wall image processing to different scanning equipment, patients, and acquisition conditions. By constructing a dual-layer modulation system, secondary modulation is performed on the lightly modulated image to cover feature-level variations, avoiding background interference that leads to inaccurate vascular wall feature extraction, thereby improving the accuracy of cranial vascular wall image processing.
[0061] Example 6, see Figure 1This embodiment, based on the above embodiment, fuses the features of the two remodulation pools using a remodulation feature fusion module, and constructs a two-level coordination constraint at the global and local levels: the global layer constraint ensures the global coordination between the remodulation prediction and the pseudo-label, while the local layer constraint mines hidden common information in the remodulation features; specifically including:
[0062] Remodulation feature extraction: The encoder extracts features from the two remodulated images, and then the decoder generates the original prediction.
[0063] Feature fusion; Alpha fusion is performed on the two remodulated features; the fused features are then decoded to generate predictions;
[0064] The prediction results are fused; and the original predictions of the two remodulations are fused with the same proportion of alpha, as shown below: ; It is the Alpha fusion function; It is the fusion ratio factor, with a value ranging from 0.3 to 0.7; and These are the original predictions of two remodulations;
[0065] Two-level coordination constraints; global level constraints, using reliable pseudo-labels from lightly tuned variables to guide the original predictions of the two retuned variables, ensuring global coordination, expressed as: Local layer constraints, using mean squared error to limit the coordination between fused feature predictions and fused prediction results, uncover hidden common information in the readjusted features, and ensure that the model captures the invariance of blood vessel wall features under readjustment, expressed as: ;in, It is a global layer constraint loss for readjustment; It is a local layer-constrained loss of the readjustment; It is the mean squared error loss; It is a segmentation prediction based on fused features. The segmentation probability map output by the decoder;
[0066] The overall loss function was redesigned; global and local layer constraints were integrated, and an adaptive vascular feature perception function was introduced. The weights are dynamically adapted to the learning state of blood vessel features during the training process, which is represented as: ; ;in, It is the total loss due to readjustment; is the incremental growth rate factor, with a value ranging from 0.15 to 0.25; t is the current training round; T is the total number of training rounds; This is the initial activation round threshold for the weights, ranging from 0.2T to 0.3T; This is the weight saturation round threshold, ranging from 0.7T to 0.8T; This is the round offset coefficient, with a value ranging from 0.1 to 0.15; It is the Dice coefficient of the blood vessel wall in the t-th round of validation set; It is the upper limit of the weight, ranging from 0.8 to 1.0; to avoid the interference of remodulation on the learning of basic features of the blood vessel wall; and to ensure the stability of fine-grained blood vessel feature extraction.
[0067] By performing the above operations, this solution addresses the problems of poor processing results caused by the susceptibility of local fine-grained features to interference, neglect of local feature invariance, and insufficient information utilization in general cranial vascular wall image processing systems. It employs a two-level coordination constraint to enable the system to learn the essential features of the vascular wall under changes in grayscale stretching, blurring, and elastic deformation, thereby improving its fine-grained robustness to clinical variations. Furthermore, it introduces adaptive vascular feature perception weights to design a remodulation total loss, avoiding interference from remodulation on the learning of basic vascular wall features, ensuring the stability of fine-grained vascular feature extraction, and ultimately improving the image processing performance of cranial vascular walls.
[0068] Example 7, see Figure 1 This embodiment is based on the above embodiment. The model training optimization module integrates the losses from labeled and unlabeled images to train and optimize the cranial blood vessel wall image processing model; specifically, it includes:
[0069] The guided loss design calculates the error between the labeled image and the true label, ensuring the model learns the basic blood vessel wall segmentation features, represented as: ; It is the guided loss of the labeled image; This refers to the batch size of the labeled images; These are the actual labels on the annotated images; It is segmentation prediction of labeled images;
[0070] For unlabeled images, without guided loss, the light-tuning loss and the heavy-tuning two-level constraint loss are fused, represented as: ;
[0071] Total loss is ; It is the total loss of the unlabeled image; It is the weighting coefficient of the lightly adjusted loss. This is the readjustment loss weighting coefficient, with a value ranging from 0.8 to 1.2; It is the total training loss;
[0072] Training optimization strategy design: The AdamW optimizer is used, with an initial learning rate of 1e. -4 The weight decays to 1e -5The learning rate is scheduled using a cosine annealing strategy; training stops if there is no improvement after 20 consecutive epochs, using the Dice similarity coefficient of the validation set as the indicator; the batch ratio of labeled images to unlabeled images is 1:4; and thus the image processing model of the cranial blood vessel wall is established.
[0073] Example 8, see Figure 1 This embodiment is based on the above embodiment. The real-time cranial blood vessel wall image processing module applies the established cranial blood vessel wall image processing model to clinically unlabeled cranial blood vessel wall images. A standardized processing flow is designed to ensure the accuracy and stability of the segmentation results. The specific operation is as follows: For newly acquired clinical cranial blood vessel wall images, preprocessing is performed, and the preprocessed image is input into the trained model to output pixel-level three-class segmentation probability maps (blood vessel wall / lumen / background); morphological closing operation is used to repair the small holes in the blood vessel wall segmentation results (adapting to the segmentation defects of thin blood vessel walls); connected component analysis is used to remove isolated small-area noise regions and retain the continuous tubular structure of the blood vessel wall; and a visualized segmentation mask is output (superimposed on the original image).
[0074] Example 9, see Figure 1 , Figure 2 and Figure 3 This embodiment is based on the above embodiment. Figure 2 These are unlabeled images of the walls of blood vessels in the brain. Figure 3This is the result of processing an image of the cerebral blood vessel wall. The image features a red background, a cyan main body, and red spots. The encoder expands the receptive field while preserving fine-grained features of the thin-walled vessel structure through dilated convolution (the edges of the cyan areas and the outlines of small branch vessels are clear, without obvious breaks or blurring, indicating that the thin-walled structure was not lost). The decoder repairs the lumen-wall boundary (the boundary between the cyan area and the red background is sharp, and the distinction between the blood vessel and the background is high, indicating that the boundary features lost during downsampling are effectively recovered). Light modulation (slight cropping + low-intensity Gaussian noise): The cyan main body area in the image is intact, without obvious artifacts or incorrect segmentation, indicating that the basic segmentation guided by pseudo-labels is accurate. Remodulation pool 1: The cyan brightness is uniform in different areas of the image, without abrupt grayscale changes, indicating that grayscale stretching / blurring improves the model's performance. It adapts to differences in imaging parameters; Remodulation pool 2 (morphology + structure): The small branch vessels in the cyan area of the image have natural morphology without twisting or breakage, indicating that morphology modulation improves the model's robustness to vessel morphology; Secondary modulation: Through multi-path modulation, the model is forced to focus on the core features of the vessel wall (the red spots in the image may be small vessels or vessel wall details, which are not interfered with by background noise, indicating that feature masking effectively shields the low-activation background, and channel dropout enhances feature robustness); Global layer constraint: Ensures the consistency between global segmentation and pseudo-labels (the overall distribution of the cyan subject and the red background in the image conforms to anatomical logic, with no large-area incorrect segmentation); Local layer constraint: The small details (spots) in the cyan area of the image are clearer after the fusion of the two remodulation pools, indicating that local commonalities are effectively captured.
[0075] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention.
[0076] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.
Claims
1. An intelligent image processing system for the cerebral blood vessel wall, characterized in that: The system includes an image acquisition module, a brain blood vessel wall image processing model design module, a label-free set modulation strategy design module, a secondary modulation module, a remodulation feature fusion module, a model training optimization module, and a real-time brain blood vessel wall image processing module. The image acquisition module acquires clinical images of the cerebral blood vessel walls, and after preprocessing, constructs labeled and unlabeled sets. The brain vascular wall image processing model design module is based on preprocessed clinical brain vascular wall images. It designs the model architecture and processing flow, which includes feature extraction, label-free set modulation, secondary modulation, remodulation feature fusion, training optimization and final output. The unlabeled set modulation strategy design module generates lightly modulated and heavily modulated images for the unlabeled set. The secondary modulation module performs secondary modulation on the lightly modulated image; The remodulation feature fusion module fuses the features and original predictions of the two remodulated images respectively; The model training and optimization module designs loss functions for labeled and unlabeled images, and trains and optimizes the cranial blood vessel wall image processing model to complete the establishment of the cranial blood vessel wall image processing model. The real-time cranial blood vessel wall image processing module performs real-time cranial blood vessel wall image processing based on the established cranial blood vessel wall image processing model. The unlabeled set modulation strategy design module specifically includes: Lightly modulated image generation: Perform slight random cropping + Gaussian noise on the unlabeled image to generate a lightly modulated image; Remodulation dual-pool generation: Construct two remodulation pools to generate remodulated images for unlabeled images; The secondary modulation module specifically includes: Feature extraction; features are extracted from the slightly modified image using an encoder; Secondary modulation operation: Simultaneously perform path dimension and domain dimension modulation on the extracted features, with each modulation operation assigned to an independent path, generating three modulated features: including: path dimension modulation; domain dimension net noise modulation; and domain dimension masking modulation. Generating pseudo-labels for the lightly modulated image; generating independent predictions from the three modulated features using the decoder; filtering the original predictions of the lightly modulated image using a confidence threshold, and retaining pixels with probability values not lower than the confidence threshold as reliable pseudo-labels; Consistency-based loss constraint; design loss functions by using reliable pseudo-labels to guide the prediction of the three paths respectively; The remodulation feature fusion module specifically includes: Remodulation feature extraction: The encoder extracts features from the two remodulated images, and then the decoder generates the original prediction. Feature fusion; Alpha fusion is performed on the two remodulated features; the fused features are then decoded to generate predictions; The prediction results are fused; and the original predictions of the two remodulations are fused with the same proportion of alpha. Two-level coordination constraints: global layer constraint, using reliable pseudo-labels from light modulation to guide the original predictions of the two remodulations, and designing a local layer constraint loss for the remodulation; local layer constraint, using mean squared error to constrain the coordination between the fused feature predictions and the fused prediction results. The total loss is retuned; by integrating global and local layer constraints and introducing an adaptive vascular feature sensing function, a retuned total loss is obtained.
2. The intelligent cranial blood vessel wall image processing system according to claim 1, characterized in that: The brain vascular wall image processing model design module is based on preprocessed clinical brain vascular wall images and adopts a lightweight encoder-decoder architecture. The encoder is responsible for extracting multi-scale features, and the decoder is responsible for feature upsampling and pixel-level segmentation prediction. The output layer uses the Softmax activation function to output a pixel-level three-class segmentation probability map; The model processing flow includes: feature extraction - unlabeled set modulation - secondary modulation - re-modulation feature fusion - training optimization - final output.
3. The intelligent cranial blood vessel wall image processing system according to claim 2, characterized in that: The image acquisition module acquires clinical HR-MRI images of the cerebral blood vessel walls, and the annotation set consists of images of the cerebral blood vessel walls labeled with three categories: blood vessel wall, lumen, and background, labeled pixel by pixel. The unlabeled set consists of unlabeled images of the same modality of cerebral blood vessel walls; Preprocessing is performed, including resampling to a uniform voxel size, grayscale normalization, and ROI cropping.
4. The intelligent cranial blood vessel wall image processing system according to claim 3, characterized in that: The model training optimization module specifically includes: Guided loss design calculates the error between the labeled image and the true label; Without guided loss, for unlabeled images, the lightly modulated loss and the heavily modulated two-level constraint loss are fused to obtain the total model loss; Training optimization strategy design: The AdamW optimizer is adopted, and the learning rate is scheduled using a cosine annealing strategy; the Dice similarity coefficient of the validation set is used as an indicator, and training is stopped if there is no improvement for 20 consecutive epochs; thus, the image processing model of the cranial blood vessel wall is established.
5. The intelligent cranial blood vessel wall image processing system according to claim 4, characterized in that: The real-time cranial blood vessel wall image processing module applies the established cranial blood vessel wall image processing model to clinically unlabeled cranial blood vessel wall images and designs a standardized processing flow.