Generative artificial intelligence cerebral artery lesion detection method based on plain-scan brain CT (Computed Tomography)

Abstract

The invention discloses a generative artificial intelligence cerebral artery lesion detection method based on plain-scan brain CT. The method comprises the steps that 1, a plain-scan brain CT image to be processed is acquired; 2, converting the plain scanning brain CT image into a CTA image through an adaptive noise elimination network; and 3, carrying out multi-modal lesion detection on the basis of the CTA image generated in the step 2. According to the method, a CTA image is generated through an adaptive noise elimination network (ANE-NET), and a real-time lesion feature analysis engine (RTAL-FE) is embedded in the generation process, so that real-time classification prediction of lesion types is realized. And furthermore, a detection result is output through a dynamic weight decision model (DWD-M), so that the generation quality and the detection efficiency are remarkably improved. The problems that a traditional method is low in generation quality and lags behind detection are solved, and the method is particularly suitable for low-dose and non-invasive cerebrovascular disease screening scenes and has important clinical application value.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical image processing, and in particular to a generative artificial intelligence cerebral arterial lesion detection method based on plain brain CT. BACKGROUND

[0002] Cerebral arterial lesions (such as aneurysm, stenosis or malformation, etc.) are common cerebrovascular diseases, and early and accurate diagnosis is crucial for preventing stroke and other serious neurological complications. At present, the widely used cerebral arterial imaging techniques in clinical practice mainly include digital subtraction angiography (DSA) and computed tomography angiography (CTA).

[0003] DSA has long been considered the standard for diagnosing cerebral vascular lesions and can provide high-resolution dynamic images of blood vessels. However, this technique is an invasive procedure that requires arterial catheterization for contrast agent injection, and it is complex to operate and carries a certain degree of surgical risk, including vascular injury, hematoma, embolization events, and contrast agent-related allergic reactions. In addition, both the patient and the medical staff need to receive a high dose of ionizing radiation during the DSA examination, and long-term or frequent examinations may pose a risk of cumulative radiation exposure.

[0004] CTA, as a relatively high non-invasive alternative, avoids arterial puncture, but still requires intravenous injection of iodine contrast agents, which may pose a risk of nephrotoxicity, and its applicability is limited for patients with renal insufficiency or contrast agent allergies. Similarly, CTA imaging is also associated with significant ionizing radiation. In addition, both of the above methods rely on large imaging equipment, which has high examination costs and strict requirements for medical resources, limiting their widespread application in some primary medical institutions or specific populations (such as pregnant women, children, and patients with renal dysfunction).

[0005] Existing generative models (such as U-Net, GAN) can attempt to generate CTA from plain CT, but have the following defects: 1. Insufficient generation quality: traditional models have limited ability to restore subtle features such as vessel wall thickness and lumen shape; 2. Diagnostic lag: usually requires generating a complete CTA image first for secondary analysis, lacking real-time performance; 3. Weak generalization ability: insufficient robustness for low-resolution images or complex lesions. SUMMARY

[0006] In view of the above problems, the present invention provides a generative artificial intelligence method for detecting cerebral artery lesions based on plain brain CT scan. This method introduces a Structured State Space Diffusion Engine (SSSD-Eng) and a Dynamic Weighted Decision Model (DWD-M), and combines v-prediction parameterization and a Min-SNR weighted training strategy to significantly improve the generation quality and diagnostic efficiency.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0008] In a first aspect, the present invention provides a generative artificial intelligence method for detecting cerebral artery lesions based on plain brain CT scans, comprising:

[0009] Step 1: Acquire plain CT images of the brain to be processed;

[0010] Step 2: Convert the plain CT brain images into CTA images using an adaptive noise cancellation network. The specific process includes:

[0011] The input plain brain CT images are dynamically normalized.

[0012] High-dimensional semantic features of plain brain CT images are extracted using latent feature mapping generation units.

[0013] During the forward noise diffusion stage, Gaussian noise is gradually added through adaptive cosine scheduling to transform the plain brain CT image into a pure noise state. The cosine scheduling is defined as follows:

[0014]

[0015] Where d is the resolution of the input image, t is the number of diffusion time steps, ∈ t Here is the Gaussian noise vector sampled at diffusion time step t, and sec is the secant function;

[0016] In the reverse denoising stage, noise is gradually removed through a multi-scale feature fusion engine to generate a high-quality CTA image;

[0017] The multi-scale feature fusion engine includes a cross-layer attention module for extracting vascular wall, lumen morphology, and hemodynamic features;

[0018] A real-time lesion feature analysis engine is embedded in the denoising process. The real-time lesion feature analysis engine is used to predict the type of cerebral artery lesion in real time based on the feature information of the current decoding stage.

[0019] Step 3: Based on the CTA images generated in Step 2, perform multimodal lesion detection, including:

[0020] Extract morphological-dynamic combined features of vascular regions, including vessel diameter, wall thickness, and blood flow velocity gradient;

[0021] The features are input into a dynamic weighted decision model, which outputs detection results through a two-stage classification strategy. The first stage of the two-stage classification strategy determines the presence of lesions, and the second stage determines the type and severity of the lesions.

[0022] In one embodiment of the present invention, the adaptive noise cancellation network includes:

[0023] The latent feature map generation unit includes a multi-level convolutional-downsampling structure for mapping the input image to the latent space;

[0024] The adaptive cosine scheduling module is used to dynamically adjust the noise scheduling curve based on the resolution of the input image to ensure the consistency of image generation at different scales.

[0025] A multi-scale feature fusion engine, including a cross-layer attention module, is used to extract local and global vascular features through a self-attention mechanism;

[0026] Integrated residual skip connections are used to preserve anatomical information from the original plain CT brain images;

[0027] The real-time lesion feature analysis engine, embedded in the intermediate layer of the denoising stage, is used to extract lesion-related features through a lightweight Transformer structure, output the probability distribution of lesion types, and determine the final prediction result through a dynamic threshold determination mechanism.

[0028] In one embodiment of the present invention, the dynamic weight decision model includes:

[0029] First-stage classifier: Determines the presence of lesions based on morphological features;

[0030] Second-stage classifier: If a lesion is present, a multi-task learning framework is used to distinguish between stenosis and aneurysm types and quantify the severity of the lesion;

[0031] The multi-task learning framework includes a shared feature extraction layer and a task-specific classification head;

[0032] A hybrid activation function combining Softmax and Sigmoid is used to process the class probability and continuous score output, respectively.

[0033] In one embodiment of the present invention, the training strategy of the generative model includes:

[0034] vPrediction parameterization: Define the noise prediction target as

[0035]

[0036] Where, σ t Let x be the standard deviation of the noise at diffusion time step t. t This refers to the intermediate latent image features or states after noise perturbation;

[0037] Min-SNR weighted loss function: through

[0038]

[0039] Reducing the weights of low-noise step sizes improves model robustness; where γ is a preset threshold to accelerate model convergence and reduce noise sensitivity. SNR(t) is the signal-to-noise ratio at diffusion time step t, used as a weighting coefficient to balance the loss contribution at different time steps.

[0040] Multi-resolution distillation training: Distillation loss of low-resolution images is introduced in the pre-training stage to accelerate the model’s learning of the global structure.

[0041] In one embodiment of the present invention, the adaptive noise cancellation network includes:

[0042] Structured State Space Diffusion Engine: A structured state space model is introduced into the forward diffusion and backward denoising stages of the diffusion model, and the temporal-spatial dynamic features of blood vessels are modeled through matrix decomposition.

[0043] The structured state-space model includes a state transition matrix A, an input matrix B, an output matrix C, and a jump connection matrix D. The parameters are optimized using gradient descent, as shown in the following formula:

[0044]

[0045] Where x t For the potential state, u t For input noise, y t For output features;

[0046] The structured state-space model and the diffusion model are jointly trained, and the generation process is accelerated through a non-autoregressive SaShiMi architecture.

[0047] In one embodiment of the present invention, the reverse denoising stage embeds a condition guidance module:

[0048] The lesion type label is injected into the diffusion model as conditional information through a cross-attention mechanism;

[0049] Conditional information is extracted by the CLIP encoder to obtain text embeddings, which are then fused with image features through attention, using the following formula:

[0050]

[0051] Where Q, K, and V are the query, key, and value matrices, respectively, and d k For feature dimensions;

[0052] The condition guidance module supports multimodal condition input, enabling dynamic control of the generation process.

[0053] In one embodiment of the present invention, the generative model supports a lightweight fine-tuning strategy:

[0054] The LoRA module is used to fine-tune the network parameters, updating only the low-rank matrix instead of all parameters, thus reducing computational resource consumption.

[0055] The LoRA module is embedded in the Unet structure of the diffusion model, using the rank decomposition formula:

[0056] ΔW=αβ T

[0057] Where W0 is the original parameter, W is the new weight actually used in this layer after fine-tuning, and α and β are low-rank matrices.

[0058] Secondly, the present invention provides a generative artificial intelligence cerebral artery lesion detection system based on plain brain CT scan, for implementing the method described above, the system comprising:

[0059] The image preprocessing module is used to perform dynamic normalization and resolution adaptation.

[0060] The adaptive noise cancellation module includes a latent feature map generation unit, an adaptive cosine scheduling module, and a multi-scale feature fusion engine.

[0061] A structured state-space diffusion engine is used to convert plain CT brain images into CTA images;

[0062] The real-time lesion analysis module, embedded in the generation process, is used to provide immediate lesion prediction;

[0063] The multimodal detection module includes a dynamic weighted decision model for outputting the final detection results.

[0064] The full-process automation controller coordinates the operation of various modules to achieve end-to-end processing from raw image input to detection output.

[0065] Thirdly, the present invention provides a computer-readable storage medium storing computer instructions which are executed by a processor using the method described above.

[0066] Fourthly, the present invention provides a computer program product, the computer program product storing computer instructions, the computer instructions being executed by a processor using the method described above.

[0067] The beneficial effects achieved by this invention are as follows:

[0068] This invention provides a generative artificial intelligence method for detecting cerebral artery lesions based on plain brain CT scans. Through adaptive cosine scheduling and v-prediction parameterization, the Fréchet distance (FD) of CTA images is reduced by approximately 90% compared to the traditional U-Net. The multi-scale feature fusion engine significantly enhances the ability to reconstruct vessel wall thickness and lumen morphology. The real-time lesion feature analysis engine of this invention enables simultaneous generation and detection, reducing secondary processing time. The dynamic weighted decision model, through a two-stage classification strategy, improves the detection accuracy to over 95%. This invention supports efficient generation and detection of low-resolution images, adapting to mobile medical devices. The Min-SNR weighted training strategy reduces the model's sensitivity to noise and improves robustness in complex lesion scenarios.

[0069] This invention generates CTA images using an Adaptive Noise Cancellation Network (ANE-NET) and embeds a Real-Time Lesion Feature Analysis Engine (RTAL-FE) during the generation process to achieve real-time classification and prediction of lesion types. Furthermore, the detection results are output through a Dynamic Weighted Decision Model (DWD-M), significantly improving both generation quality and detection efficiency. This invention solves the problems of low generation quality and detection lag in traditional methods, and is particularly suitable for low-dose, non-invasive cerebrovascular disease screening scenarios, possessing significant clinical application value. Attached Figure Description

[0070] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the embodiments of the present invention will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the drawings without creative effort.

[0071] Figure 1 This is a flowchart of the generative artificial intelligence method for detecting cerebral artery lesions based on plain brain CT scan, which is based on the present invention.

[0072] Figure 2 This is a structural diagram of the adaptive noise cancellation network (ANE-NET) of this invention.

[0073] Figure 3This is a structural diagram of the Structured State Space Diffusion Engine (SSSD-Eng) of the present invention.

[0074] Figure 4 This is a schematic diagram of the LoRA fine-tuning strategy of the present invention.

[0075] Figure 5 This is a structural diagram of the Dynamic Weighted Decision Model (DWD-M) of this invention.

[0076] Figure 6 This is a diagram showing the relationship between the structured state-space diffusion engine and the LoRA fine-tuning strategy of this invention. Detailed Implementation

[0077] The embodiments of the technical solution of the present invention will now be described in detail with reference to the accompanying drawings. These embodiments are merely illustrative of the technical solution of the present invention and are therefore intended to limit the scope of protection of the present invention.

[0078] Unless otherwise defined, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains; the terminology used in this invention is for the purpose of describing particular embodiments only and is not intended to limit the invention; the terms “comprising” and “having”, and any variations thereof, in the specification, claims and foregoing description of the drawings are intended to cover non-exclusive inclusion.

[0079] In the description of the embodiments of this invention, technical terms such as "first" and "second" are used only to distinguish different objects and should not be construed as indicating or implying relative importance or implicitly specifying the number, specific order, or primary and secondary relationship of the indicated technical features. In the description of the embodiments of this invention, "multiple" means two or more, unless otherwise explicitly defined.

[0080] In this invention, the reference to "embodiment" means that a specific feature, structure, or characteristic described in connection with an embodiment may be included in at least some embodiments of the invention. The appearance of this phrase in various places in the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described in this invention can be combined with other embodiments.

[0081] In the description of the embodiments of this invention, the term "and / or" is merely a description of the relationship between associated objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, in this invention, the character " / " generally indicates that the preceding and following associated objects have an "or" relationship.

[0082] Example 1:

[0083] Please refer to Figures 1 to 6 This embodiment provides a generative artificial intelligence method for detecting cerebral artery lesions based on plain brain CT scans. The method includes the following steps:

[0084] Step 1: Acquire plain CT images of the brain to be processed;

[0085] Step 2: Convert the plain CT brain images into CTA images using an adaptive noise cancellation network. The specific process includes:

[0086] The input plain brain CT images are dynamically normalized to adapt to the input range of the generated model.

[0087] High-dimensional semantic features of plain brain CT images are extracted using latent feature mapping generation units.

[0088] During the forward noise diffusion stage, Gaussian noise is gradually added through adaptive cosine scheduling to transform the plain brain CT image into a pure noise state. The cosine scheduling is defined as follows:

[0089]

[0090] Where d is the resolution of the input image, t is the number of diffusion time steps, ∈ t Here is the Gaussian noise vector sampled at diffusion time step t, and sec is the secant function;

[0091] In the reverse denoising stage, noise is gradually removed through a multi-scale feature fusion engine to generate a high-quality CTA image;

[0092] The multi-scale feature fusion engine includes a cross-layer attention module for extracting vascular wall, lumen morphology, and hemodynamic features;

[0093] A real-time lesion feature analysis engine is embedded in the denoising process. The real-time lesion feature analysis engine is used to predict the type of cerebral artery lesion in real time based on the feature information of the current decoding stage. The lesion types include, but are not limited to, cerebral artery stenosis and aneurysm.

[0094] Step 3: Based on the CTA images generated in Step 2, perform multimodal lesion detection, including:

[0095] Extract morphological-dynamic combined features of vascular regions, including vessel diameter, wall thickness, and blood flow velocity gradient;

[0096] The features are input into a dynamic weighted decision model, which outputs detection results through a two-stage classification strategy. The first stage of the two-stage classification strategy determines the presence of lesions, and the second stage determines the type and severity of the lesions.

[0097] Optionally, the adaptive noise cancellation network includes:

[0098] The latent feature map generation unit includes a multi-level convolutional-downsampling structure for mapping the input image to the latent space;

[0099] The adaptive cosine scheduling module is used to dynamically adjust the noise scheduling curve based on the resolution of the input image to ensure the consistency of image generation at different scales.

[0100] The multi-scale feature fusion engine includes a cross-layer attention module, which is used to extract local and global vascular features through a self-attention mechanism;

[0101] Integrated residual skip connections are used to preserve anatomical information from the original plain CT brain images;

[0102] The real-time lesion feature analysis engine, embedded in the intermediate layer of the denoising stage, is used to extract lesion-related features through a lightweight Transformer structure, output the probability distribution of lesion types, and determine the final prediction result through a dynamic threshold determination mechanism.

[0103] Optionally, the dynamic weight decision model includes:

[0104] The first-stage classifier determines the presence of lesions based on morphological features (such as the rate of change in vessel diameter).

[0105] The second-stage classifier: If a lesion is present, a multi-task learning framework is used to further distinguish between stenosis and aneurysm types and quantify the severity of the lesion;

[0106] The multi-task learning framework includes a shared feature extraction layer and a task-specific classification head;

[0107] A hybrid activation function combining Softmax and Sigmoid is used to process the class probability and continuous score output, respectively.

[0108] Optionally, the training strategy for the generative model includes:

[0109] vPrediction parameterization: Define the noise prediction target as

[0110]

[0111] Instead of traditional noise prediction; where σ t Let x be the standard deviation of the noise at diffusion time step t. t This refers to the intermediate latent image features or states after noise perturbation;

[0112] Min-SNR weighted loss function: through

[0113]

[0114] Reducing the weights of low-noise step sizes improves model robustness; where γ is a preset threshold to accelerate model convergence and reduce noise sensitivity. To balance the weighting coefficients of the loss contribution at different time steps, SNR(t) is the signal-to-noise ratio at the diffusion time step t;

[0115] Multi-resolution distillation training: Distillation loss of low-resolution images is introduced in the pre-training stage to accelerate the model’s learning of the global structure.

[0116] The method provided by this invention realizes a multi-stage collaborative processing flow from raw plain brain CT images to CTA image generation and then to lesion detection. The training of the generation model adopts the above-mentioned v prediction parameterization and Min-SNR weighted joint strategy.

[0117] Optionally, the adaptive noise cancellation network (ANE-NET) includes:

[0118] Structured State Space Diffusion Engine (SSSD-Eng): Introduces a Structured State Space Model (SSM) into the forward diffusion and backward denoising stages of the diffusion model, and models the temporal-spatial dynamic characteristics of blood vessels through matrix decomposition (such as HiPPO theory).

[0119] The structured state-space model includes a state transition matrix A, an input matrix B, an output matrix C, and a skip connection matrix D. The parameters are optimized using gradient descent, as shown in the following formula:

[0120]

[0121] Where x t For the potential state, u t For input noise, y t For output features;

[0122] The structured state-space model and the diffusion model are jointly trained, and the generation process is accelerated by a non-autoregressive SaShiMi architecture.

[0123] Optionally, the reverse denoising stage embeds a Conditional Guidance Module (CGM):

[0124] The lesion type label (such as stenosis / aneurysm) is injected into the diffusion model as conditional information through a cross-attention mechanism;

[0125] Conditional information is extracted into text embeddings using the CLIP encoder (Contrastive Language-Image Pretraining) and then fused with image features through attention, as shown in the formula:

[0126]

[0127] Where Q, K, and V are the query, key, and value matrices, respectively, and d k For feature dimensions;

[0128] The condition guidance module supports multimodal condition input (such as text description and target trajectory) to achieve dynamic control of the generation process.

[0129] Optionally, the generative model supports a lightweight fine-tuning strategy:

[0130] The LoRA (Low-Rank Adaptation) technique is used to fine-tune the network parameters, updating only the low-rank matrix instead of all parameters, thus reducing computational resource consumption.

[0131] The LoRA module is embedded in the Unet structure of the diffusion model, using the rank decomposition formula:

[0132] ΔW=αβ T

[0133] Where W0 is the original parameter, W is the new weight actually used in this layer after fine-tuning, and α and β are low-rank matrices, which significantly reduce training costs.

[0134] Example 2: Method Implementation

[0135] (1) Input preprocessing:

[0136] Obtain plain brain CT images (512×512×128), perform dynamic normalization processing, and limit the pixel value range to [0,1].

[0137] (2) Adaptive noise cancellation network (ANE-NET) generates CTA images:

[0138] a. Encoder: 5-level convolutional-downsampling structure to extract high-dimensional semantic features;

[0139] Forward diffusion: Gaussian noise is gradually added through adaptive cosine scheduling to generate a latent representation;

[0140] b. Reverse denoising: CLAM is used in MSFFE to extract the morphology and hemodynamic features of the vessel wall and lumen, and the decoder (5-level convolution-upsampling) restores the CTA image;

[0141] c.RTAL-FE: Embed a lightweight Transformer in the denoising stage to output the probability distribution of stenosis (0.8) and aneurysm (0.2).

[0142] (3) Dynamic Weighted Decision Model (DWD-M) testing:

[0143] Phase 1: Determining the presence of lesions based on morphological characteristics (rate of change in vessel diameter);

[0144] Phase 2: Differentiate lesion types using MTL-F and quantify the degree of stenosis as 70%.

[0145] Example 3: System Deployment

[0146] Image acquisition module: Connects to the hospital's PACS system to acquire plain CT data in real time;

[0147] ANE-NET module: Deployed on GPU clusters, it supports batch processing;

[0148] RTAL-FE module: Embedded in the generation process, outputting prediction results in real time;

[0149] DWD-M module: Employs a pre-trained CNN model (such as ResNet-50) and combines it with labeled data to optimize detection accuracy;

[0150] Output interface: Outputs test reports via DICOM standard and integrates them into the hospital's imaging system.

[0151] Example 4: Application of a Structured State-Space Diffusion Engine

[0152] (1) Input preprocessing:

[0153] Obtain plain brain CT images (512×512×128) and perform dynamic normalization processing.

[0154] (2) SSSD-Eng generates CTA images:

[0155] a. State-space modeling: Initialize the state transition matrix A using HiPPO theory, and extract the spatiotemporal dynamic features of blood vessels using SSM;

[0156] b. Non-autoregressive generation: The SaShiMi architecture is used to reduce the number of denoising steps to 50, increasing the generation speed by 30%;

[0157] c. Output results: The generated CTA images clearly restore the vessel wall thickness and lumen morphology, with a Fréchet distance (FD) of 0.78.

[0158] Example 5: Multimodal Control of Conditional Guidance Module

[0159] (1) Conditional input:

[0160] Input text description: "Aneurysm located in the anterior communicating artery, diameter >5mm";

[0161] Extract text embeddings using the CLIP encoder.

[0162] (2) CGM generates CTA images:

[0163] By performing cross-attention fusion of text embeddings and image features, an aneurysm image that matches the description is generated;

[0164] Output results: The location and size of the aneurysm are consistent with the input description, and the detection accuracy is improved to 96%.

[0165] Example 5: Parameter Optimization of Lightweight Fine-Tuning Strategy

[0166] Pre-trained model: A pre-trained diffusion model (such as ADM) is used.

[0167] LoRA fine-tuning: Embedding the LoRA module into the Unet structure, updating only the low-rank matrices A and B, reduces training time from 72 hours to 18 hours, generating a quality loss of <5%.

[0168] Therefore, the Adaptive Noise Elimination Network (ANE-NET) of this invention:

[0169] The forward diffusion process based on the Adaptive Cosine Schedule is defined as follows:

[0170]

[0171] Dynamically adjust the noise intensity during the diffusion process to adapt to input images of different resolutions;

[0172] By using v-prediction parameterization, the noise prediction target is defined as...

[0173]

[0174] Optimize denoising stability and improve CTA image generation quality to Fréchet distance (FD) < 0.8;

[0175] The system integrates a multi-scale feature fusion engine (MSFFE) and utilizes a cross-layer attention module (CLAM) with residual skip connections to accurately restore key features such as vessel wall thickness and lumen morphology.

[0176] The Real-Time Lesion Feature Analysis Engine (RTAL-FE) of this invention:

[0177] A lightweight Transformer structure is embedded in the inverse denoising stage of the diffusion model to extract lesion-related features (such as the rate of change of vessel diameter and blood flow velocity gradient) in real time.

[0178] A dynamic thresholding mechanism is adopted to dynamically adjust the classification confidence threshold based on image quality, ensuring robustness of detection in low-resolution scenes.

[0179] The Dynamic Weighted Decision Model (DWD-M) of this invention:

[0180] A two-stage classification strategy is adopted: the first stage determines the presence of lesions based on morphological features, and the second stage distinguishes between stenosis and aneurysm types and quantifies the degree of stenosis (e.g., 70%) using a multi-task learning framework (MTL-F).

[0181] A Softmax-Sigmoid hybrid activation function is introduced to accommodate both categorical probability output and continuous scoring requirements, thus adapting to the multi-dimensional detection of complex lesions.

[0182] The Structured State Space Diffusion Engine (SSSD-Eng) and Lightweight Fine-Tuning Strategy of this Invention:

[0183] By modeling the spatiotemporal dynamics of blood vessels using a structured state-space model (SSM) driven by HiPPO theory, and combining it with a non-autoregressive SaShiMi architecture, the number of generation steps is reduced from 1000 to 50, increasing the speed by 30%.

[0184] This invention employs LoRA (Low-Rank Adaptation) technology to fine-tune the pre-trained diffusion model, updating only the low-rank matrix.

[0185] ΔW=AB T

[0186] Reduce training costs by 70% and generate quality loss of less than 5%.

[0187] The method of this invention was validated on the TopCoW dataset (89 patients) and an external test set (11 patients), and the results showed:

[0188] Generation quality: FD score is 90% lower than traditional U-Net (0.816 vs. 10.518);

[0189] Detection efficiency: Stenosis detection sensitivity reaches 92%, aneurysm identification specificity reaches 88%;

[0190] Generalization capability: Supports the generation of 128×128 low-resolution images, adapting to mobile medical devices.

[0191] This invention provides an efficient tool for non-invasive screening, emergency auxiliary detection, and multimodal image integration of cerebrovascular diseases, significantly reducing the missed diagnosis rate (<5%) and shortening the detection cycle, and has significant clinical value and industrial application potential.

[0192] Furthermore, the present invention also provides a generative artificial intelligence cerebral artery lesion detection system based on plain brain CT scan, the system comprising:

[0193] The image preprocessing module is used to perform dynamic normalization and resolution adaptation.

[0194] The Adaptive Noise Cancellation Module (ANE-NET) includes a latent feature map generation unit, an adaptive cosine scheduling module, and a multi-scale feature fusion engine.

[0195] A structured state-space diffusion engine is used to convert plain CT brain images into CTA images;

[0196] The Real-Time Lesion Analysis Module (RTAL-FE), embedded in the generation process, is used to provide immediate lesion prediction;

[0197] The multimodal detection module (DWD-M) includes a dynamic weighted decision model, which is used to output the final detection results;

[0198] The full-process automation controller coordinates the operation of various modules to achieve end-to-end processing from raw image input to detection output.

[0199] In summary, this invention provides a generative artificial intelligence method for detecting cerebral artery lesions based on plain brain CT scans. It generates CTA images using an Adaptive Noise Cancellation Network (ANE-NET) and embeds a Real-Time Lesion Feature Analysis Engine (RTAL-FE) during the generation process to achieve real-time classification and prediction of lesion types. Furthermore, the detection results are output through a Dynamic Weighted Decision Model (DWD-M), significantly improving both generation quality and detection efficiency. This invention solves the problems of low generation quality and detection lag in traditional methods and has significant clinical application value.

[0200] This invention also has the following advantages:

[0201] Improved generation quality: Through adaptive cosine scheduling and v prediction parameterization, the Fréchet distance (FD) of CTA images is reduced by about 90% compared with the traditional U-Net (FD = 0.816 vs. 10.518); the multi-scale feature fusion engine significantly enhances the ability to restore vessel wall thickness and lumen morphology.

[0202] Detection efficiency optimization: The real-time lesion feature analysis engine enables simultaneous generation and detection, reducing secondary processing time; the dynamic weight decision model improves detection accuracy to over 95% through a two-stage classification strategy.

[0203] Enhanced clinical applicability: Supports efficient generation and detection of low-resolution images (e.g., 128×128), adapting to mobile medical devices; Min-SNR weighted training strategy reduces the model's sensitivity to noise and improves robustness in complex disease scenarios.

[0204] This invention can be widely applied in the field of medical image analysis, including but not limited to:

[0205] Cerebrovascular disease screening: Based on non-invasive CTA image generation from plain brain CT scan, it is suitable for early detection of cerebral artery stenosis and aneurysm in primary healthcare institutions.

[0206] Emergency auxiliary detection: In patients with acute stroke, CTA images can be generated quickly and lesion warnings can be output simultaneously, shortening the detection process to 30 seconds per case.

[0207] Multimodal image integration: Combined analysis with MRI or DSA data provides a comprehensive assessment of vascular morphology and hemodynamics, assisting in surgical planning.

[0208] Telemedicine support: Deployed on mobile devices using a lightweight, fine-tuning strategy to enable real-time lesion screening and remote expert consultations in remote areas.

[0209] In some embodiments, the present invention provides a computer-readable storage medium storing computer instructions that are executed by a processor as described in any of the above embodiments, namely, a generative artificial intelligence method for detecting cerebral artery lesions based on plain brain CT scans.

[0210] Computer-readable storage media may take the form of any combination of one or more readable media. A readable medium may be a readable signal medium or a readable storage medium. A readable storage medium may, for example, include, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatuses, or devices, or any combination thereof. More specific examples (not an exhaustive list) of readable storage media may include: electrical connections having one or more wires, portable disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0211] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0212] Embodiments of the present invention may also be computer program products, which include computer program instructions that, when executed by a processor, cause the processor to perform the steps in a generative artificial intelligence method for detecting cerebral artery lesions based on plain brain CT scan according to various embodiments of the present invention, as described in the "Exemplary Methods" section above.

[0213] The steps of the method of the present invention are not limited to the specific order described above, unless otherwise specifically stated. Furthermore, in some embodiments, the invention may also be implemented as a program recorded on a recording medium, the program comprising machine-readable instructions for implementing the method according to the invention. Therefore, the invention also covers recording media storing programs for performing the method according to the invention.

[0214] Although the invention has been described with reference to preferred embodiments, various modifications can be made and components can be replaced with equivalents without departing from the scope of the invention. In particular, the technical features mentioned in the various embodiments can be combined in any manner as long as there is no structural conflict. The invention is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the claims.

Claims

1. A generative artificial intelligence method for detecting cerebral artery lesions based on plain brain CT scan, characterized in that, include: Step 1: Acquire plain CT images of the brain to be processed; Step 2: Convert the plain CT brain images into CTA images using an adaptive noise cancellation network. The specific process includes: The input plain brain CT images are dynamically normalized. High-dimensional semantic features of plain brain CT images are extracted using latent feature mapping generation units. During the forward noise diffusion stage, Gaussian noise is gradually added through adaptive cosine scheduling to transform the plain brain CT image into a pure noise state. The cosine scheduling is defined as follows: Where d is the resolution of the input image, t is the number of diffusion time steps, ∈ t Here is the Gaussian noise vector sampled at diffusion time step t, and sec is the secant function; In the reverse denoising stage, noise is gradually removed through a multi-scale feature fusion engine to generate a high-quality CTA image; The multi-scale feature fusion engine includes a cross-layer attention module for extracting vascular wall, lumen morphology, and hemodynamic features; A real-time lesion feature analysis engine is embedded in the denoising process. The real-time lesion feature analysis engine is used to predict the type of cerebral artery lesion in real time based on the feature information of the current decoding stage. Step 3: Based on the CTA images generated in Step 2, perform multimodal lesion detection, including: Extract morphological-dynamic combined features of vascular regions, including vessel diameter, wall thickness, and blood flow velocity gradient; The features are input into a dynamic weighted decision model, which outputs detection results through a two-stage classification strategy. The first stage of the two-stage classification strategy determines the presence of lesions, and the second stage determines the type and severity of the lesions.

2. The generative artificial intelligence method for detecting cerebral artery lesions based on plain brain CT scan according to claim 1, characterized in that, The adaptive noise cancellation network includes: The latent feature map generation unit includes a multi-level convolutional-downsampling structure for mapping the input image to the latent space; The adaptive cosine scheduling module is used to dynamically adjust the noise scheduling curve based on the resolution of the input image to ensure the consistency of image generation at different scales. A multi-scale feature fusion engine, including a cross-layer attention module, is used to extract local and global vascular features through a self-attention mechanism; Integrated residual skip connections are used to preserve anatomical information from the original plain CT brain images; The real-time lesion feature analysis engine, embedded in the intermediate layer of the denoising stage, is used to extract lesion-related features through a lightweight Transformer structure, output the probability distribution of lesion types, and determine the final prediction result through a dynamic threshold determination mechanism.

3. The generative artificial intelligence method for detecting cerebral artery lesions based on plain brain CT scan according to claim 2, characterized in that, The dynamic weight decision model includes: First-stage classifier: Determines the presence of lesions based on morphological features; Second-stage classifier: If a lesion is present, a multi-task learning framework is used to distinguish between stenosis and aneurysm types and quantify the severity of the lesion; The multi-task learning framework includes a shared feature extraction layer and a task-specific classification head; A hybrid activation function combining Softmax and Sigmoid is used to process the class probability and continuous score output, respectively.

4. The generative artificial intelligence method for detecting cerebral artery lesions based on plain brain CT scan according to claim 3, characterized in that, The training strategy for the generative model includes: vPrediction parameterization: Define the noise prediction target as Where, σ t Let x be the standard deviation of the noise at diffusion time step t. t This refers to the intermediate latent image features or states after noise perturbation; Min-SNR weighted loss function: through Reducing the weights of low-noise step sizes improves model robustness; where γ is a preset threshold to accelerate model convergence and reduce noise sensitivity. SNR(t) is the signal-to-noise ratio at diffusion time step t, used as a weighting coefficient to balance the loss contribution at different time steps. Multi-resolution distillation training: Distillation loss of low-resolution images is introduced in the pre-training stage to accelerate the model’s learning of the global structure.

5. The generative artificial intelligence method for detecting cerebral artery lesions based on plain brain CT scan according to claim 4, characterized in that, The adaptive noise cancellation network includes: Structured State Space Diffusion Engine: A structured state space model is introduced into the forward diffusion and backward denoising stages of the diffusion model, and the temporal-spatial dynamic features of blood vessels are modeled through matrix decomposition. The structured state-space model includes a state transition matrix A, an input matrix B, an output matrix C, and a jump connection matrix D. The parameters are optimized using gradient descent, as shown in the following formula: Where x t For the potential state, u t For input noise, y t For output features; The structured state-space model and the diffusion model are jointly trained, and the generation process is accelerated through a non-autoregressive SaShiMi architecture.

6. The generative artificial intelligence method for detecting cerebral artery lesions based on plain brain CT scan according to claim 5, characterized in that, The inverse denoising stage embeds a conditional guidance module: The lesion type label is injected into the diffusion model as conditional information through a cross-attention mechanism; Conditional information is extracted by the CLIP encoder to obtain text embeddings, which are then fused with image features through attention, using the following formula: Where Q, K, and V are the query, key, and value matrices, respectively, and d k For feature dimensions; The condition guidance module supports multimodal condition input, enabling dynamic control of the generation process.

7. The generative artificial intelligence method for detecting cerebral artery lesions based on plain brain CT scan according to claim 6, characterized in that, The generative model supports lightweight fine-tuning strategies: The LoRA module is used to fine-tune the network parameters, updating only the low-rank matrix instead of all parameters, thus reducing computational resource consumption. The LoRA module is embedded in the Unet structure of the diffusion model, using the rank decomposition formula: Where W0 is the original parameter, W is the new weight actually used in this layer after fine-tuning, and α and β are low-rank matrices.

8. A generative artificial intelligence cerebral artery lesion detection system based on plain brain CT scan, characterized in that, The system for implementing the method as described in any one of claims 1 to 7 comprises: The image preprocessing module is used to perform dynamic normalization and resolution adaptation. The adaptive noise cancellation module includes a latent feature map generation unit, an adaptive cosine scheduling module, and a multi-scale feature fusion engine. A structured state-space diffusion engine is used to convert plain CT brain images into CTA images; The real-time lesion analysis module, embedded in the generation process, is used to provide immediate lesion prediction; The multimodal detection module includes a dynamic weighted decision model for outputting the final detection results. The full-process automation controller coordinates the operation of various modules to achieve end-to-end processing from raw image input to detection output.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions that are executed by a processor according to any one of claims 1 to 7.

10. A computer program product, characterized in that, The computer program product stores computer instructions, which are executed by a processor according to any one of claims 1 to 7.

Images