An AI medical image intelligent diagnosis method and system

Abstract

The application relates to the technical field of computer vision, in particular to an AI medical image intelligent diagnosis method and system, which comprises the following steps: acquiring medical image data to be diagnosed and extracting multi-scale image features through a pre-trained image feature extraction network; adjusting the feature distribution based on device type identification through a device perception feature normalization module to generate normalized multi-scale fusion features; adaptively enhancing local texture features through a fine-grained texture alignment network to generate aligned fine-grained diagnosis features; and outputting a disease diagnosis result and a confidence value based on a diagnosis reasoning network. The device perception feature normalization module and the fine-grained texture alignment network are introduced, so that the problem of bottom feature distribution deviation caused by differences in image device brands, models and parameters is effectively solved.

Description

Technical Field

[0001] This invention relates to the field of computer vision technology, and in particular to an AI-based intelligent diagnostic method and system for medical images. Background Technology

[0002] Due to differences in the brand, model, and parameter settings of imaging equipment, different medical institutions acquire medical images with significant distribution shifts in underlying features such as grayscale distribution, texture characteristics, and noise levels. How to achieve stable and reliable diagnosis under such distribution shift conditions is a core problem that needs to be overcome.

[0003] To address the aforementioned distribution shift problem, common solutions are mainly based on domain adaptation techniques, using deep learning methods such as adversarial training, feature alignment, or image style transfer to map and align the feature distributions of the source and target domains. However, the differences in images from different devices are not only reflected at the high-level semantic feature level, but also more significantly in the low-level texture and contrast details.

[0004] Existing domain adaptation methods often focus on capturing global statistical characteristics while neglecting to effectively model these fine-grained underlying differences. This results in a significant weakening of the model's ability to extract and identify key lesion information when deployed across devices, making it difficult to provide stable and accurate medical diagnoses. Summary of the Invention

[0005] The purpose of this invention is to provide an AI-based intelligent diagnostic method and system for medical images to solve the technical problems mentioned in the background.

[0006] One aspect of the present invention provides an AI-based intelligent diagnostic method and system for medical images, comprising: Acquire medical image data to be diagnosed, and input the medical image data to be diagnosed into a pre-trained image feature extraction network to obtain multi-scale image features corresponding to the medical image data to be diagnosed; the multi-scale image features include global feature representation and local texture feature representation; The multi-scale image features are input into the device perception feature normalization module. Based on the device perception feature normalization module, the feature distribution of the device type identifiers corresponding to the imaging devices of different medical institutions is adjusted to generate normalized multi-scale fusion features. The normalized multi-scale fusion features are input into a fine-grained texture alignment network. Based on the fine-grained texture alignment network, adaptive texture enhancement processing is performed on the normalized local texture feature representation to generate aligned fine-grained diagnostic features. The aligned fine-grained diagnostic features are input into the diagnostic inference network, and the disease diagnosis result corresponding to the medical image data to be diagnosed and the confidence value corresponding to the disease diagnosis result are output based on the diagnostic inference network.

[0007] In some embodiments, acquiring medical image data to be diagnosed and inputting the medical image data to be diagnosed into a pre-trained image feature extraction network to obtain multi-scale image features corresponding to the medical image data to be diagnosed includes: The medical image data to be diagnosed is subjected to image preprocessing operations to obtain standardized medical images; the image preprocessing operations include grayscale normalization, size normalization, and noise suppression. The standardized medical images are input into a preset multi-scale convolutional neural network, and feature maps of different receptive field sizes are extracted through the multi-layer convolutional structure of the multi-scale convolutional neural network. A feature aggregation operation is performed on each of the feature maps to generate a global feature representation including global semantic information and a local texture feature representation including local texture details.

[0008] In some embodiments, the device-aware feature normalization module includes a device embedding vector generation unit, a feature distribution estimation unit, and a feature mapping unit; The step of inputting the multi-scale image features into the device perception feature normalization module, and adjusting the feature distribution of the device type identifiers corresponding to imaging devices in different medical institutions based on the device perception feature normalization module to generate normalized multi-scale fusion features includes: Based on the device type identifier, a device embedding vector corresponding to the device type identifier is generated by the device embedding vector generation unit; The multi-scale image features are concatenated with the device embedding vector to obtain an enhanced feature representation for device perception. Based on the feature distribution estimation unit, the distribution parameters of the enhanced feature representation perceived by the device are estimated to generate target distribution parameters; Based on the feature mapping unit, the enhanced feature representation perceived by the device is mapped in feature space according to the target distribution parameters to generate normalized multi-scale fusion features.

[0009] In some embodiments, the fine-grained texture alignment network includes a multi-level texture perception module, an attention weight generation module, and a feature enhancement module; The process involves inputting the normalized multi-scale fusion features into a fine-grained texture alignment network, and then performing adaptive texture enhancement processing on the normalized local texture feature representation based on the fine-grained texture alignment network to generate aligned fine-grained diagnostic features, including: The normalized local texture feature representation is input into the multi-level texture perception module, and multi-scale texture pattern features are extracted through the multi-level texture perception module. The attention weight generation module analyzes the multi-scale texture pattern features and generates attention weight vectors corresponding to each texture pattern. Based on the feature enhancement module, the multi-scale texture pattern features are weighted and fused according to the attention weight vector to generate aligned fine-grained diagnostic features. In some embodiments, the diagnostic reasoning network includes a disease classification unit and a confidence assessment unit; The step of inputting the aligned fine-grained diagnostic features into a diagnostic inference network, and outputting the disease diagnosis result corresponding to the medical image data to be diagnosed and the confidence value corresponding to the disease diagnosis result based on the diagnostic inference network, includes: The aligned fine-grained diagnostic features are input into the disease classification unit, and the classification scores of each preset disease category are output through the fully connected layer structure of the disease classification unit. Based on the classification scores of each preset disease category, the predicted probability distribution of each preset disease category is calculated using the softmax function; The disease category corresponding to the maximum probability value in the predicted probability distribution is determined as the disease diagnosis result; The confidence assessment unit performs uncertainty quantification on the predicted probability distribution to generate a confidence value corresponding to the disease diagnosis result.

[0010] In some embodiments, the training steps for the image feature extraction network, the device-aware feature normalization module, the fine-grained texture alignment network, and the diagnostic inference network include: Acquire medical imaging data of multiple samples, along with corresponding sample device type identifiers and sample diagnostic labels; The medical image data of each sample is input into the initial image feature extraction network to obtain the multi-scale image features of each sample medical image data; The multi-scale image features of each sample are input into the initial device perception feature normalization module, and the feature distribution is adjusted based on the device type identifier of each sample to generate the multi-scale fusion features after sample normalization. The normalized multi-scale fusion features of each sample are input into the initial fine-grained texture alignment network to generate fine-grained diagnostic features after sample alignment. The fine-grained diagnostic features of each sample after alignment are input into the initial diagnostic inference network, and the predicted diagnostic results corresponding to the medical image data of each sample are output. Based on the classification error between the predicted diagnostic results corresponding to the medical image data of each sample and the diagnostic label of the sample, a first loss value is calculated. The second loss value is calculated based on the feature distribution differences between the device type identifiers corresponding to the medical image data of each sample. Based on the first loss value and the second loss value, the model parameters of each network module are adjusted, and iterative training continues until the training stopping condition is met, resulting in the trained network model.

[0011] In some embodiments, the second loss value is calculated based on the following: Calculate the maximum mean difference measure among the normalized features corresponding to different device type identifiers; The second loss value is determined based on the maximum mean difference metric, so that medical image data from different devices can be aligned in the feature space during training.

[0012] In some embodiments, the medical imaging data to be diagnosed includes: computed tomography (CT) images, magnetic resonance imaging (MRI) images, ultrasound images, X-ray images, and positron emission tomography (PET) images.

[0013] In some embodiments, the image feature extraction network is a deep convolutional neural network; the device perception feature normalization module is a feature normalization network based on an attention mechanism; the fine-grained texture alignment network is a multi-scale feature fusion network; and the diagnostic inference network is a fully connected neural network.

[0014] Another aspect of the present invention provides an AI-powered intelligent diagnostic system for medical images, comprising: The image acquisition module is used to acquire medical image data to be diagnosed and input the medical image data to be diagnosed into a pre-trained image feature extraction network to obtain multi-scale image features corresponding to the medical image data to be diagnosed; the multi-scale image features include global feature representation and local texture feature representation; The feature normalization module is used to input the multi-scale image features into the device perception feature normalization module, and adjust the feature distribution of the device type identifiers corresponding to the imaging devices of different medical institutions based on the device perception feature normalization module to generate normalized multi-scale fusion features; the device type identifier is used to characterize the device model information corresponding to the original acquisition device of the medical image data to be diagnosed. The texture alignment module is used to input the normalized multi-scale fusion features into the fine-grained texture alignment network, and perform adaptive texture enhancement processing on the normalized local texture feature representation based on the fine-grained texture alignment network to generate aligned fine-grained diagnostic features. The diagnostic output module is used to input the aligned fine-grained diagnostic features into the diagnostic inference network, and output the disease diagnosis result corresponding to the medical image data to be diagnosed and the confidence value corresponding to the disease diagnosis result based on the diagnostic inference network.

[0015] Compared with the prior art, the present invention has the following beneficial effects: 1. By introducing a device perception feature normalization module, the feature distribution of the device type identifiers corresponding to imaging equipment in different medical institutions is adjusted, which effectively solves the problem of underlying feature distribution offset caused by differences in imaging equipment brand, model and parameter settings; 2. By using a fine-grained texture alignment network to adaptively enhance local texture features, the complete extraction and accurate identification of key lesion information are ensured when deployed across devices. This significantly improves the diagnostic stability and reliability of the model under distribution offset conditions, enabling stable and reliable intelligent diagnosis across different imaging devices. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art 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 these drawings without creative effort.

[0017] Figure 1 This is a flowchart of the AI-based intelligent medical image diagnosis method of the present invention; Figure 2 This is a framework diagram of the AI-powered intelligent medical image diagnosis system of the present invention; Figure 3 This is a diagram illustrating an application scenario of the present invention. Detailed Implementation

[0018] The following will be based on embodiments of the present invention. Figures 1-3 The technical solutions in the embodiments of the present invention will be clearly and completely described together. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0019] Example 1 The process involves acquiring diagnostic medical image data, which may include at least one of computed tomography (CT) images, magnetic resonance imaging (MRI) images, ultrasound images, X-ray images, and positron emission tomography (PET) images. Upon receiving the diagnostic medical image data, image preprocessing is performed to obtain standardized medical images. The image preprocessing specifically includes three steps: grayscale normalization, size normalization, and noise suppression.

[0020] In the grayscale normalization process, the pixel value distribution range of the input raw medical image is statistically analyzed and mapped to ensure that image data from different imaging devices have consistent distribution characteristics in the grayscale space. This step calculates the grayscale histogram distribution parameters of the input image and performs a mapping transformation based on preset target grayscale distribution parameters to eliminate grayscale distribution deviations caused by differences in exposure parameters and contrast settings of imaging devices.

[0021] In the size standardization process, the image data, after grayscale normalization, is uniformly adjusted to a preset standard resolution. This step first detects the actual spatial resolution of the input image, and then scales it to the target resolution using bilinear or bicubic interpolation algorithms. During scaling, the spatial proportions of the image are maintained, and edge padding or cropping operations are used to ensure that the aspect ratio of the output image meets the preset standard format requirements.

[0022] In the noise suppression process, the standardized images are filtered to eliminate various noise interferences introduced during imaging. This step employs an adaptive anisotropic diffusion filtering algorithm, which dynamically adjusts the filtering intensity based on the texture complexity of local image regions. This effectively suppresses Gaussian noise and salt-and-pepper noise while preserving the edge details of lesion areas in the image.

[0023] After image preprocessing, standardized medical images are input into a pre-defined multi-scale convolutional neural network (CNN). The CNN's multi-layered convolutional structure extracts feature maps of different receptive field sizes. This multi-scale CNN employs a deep residual network architecture, comprising multiple residual blocks. Each residual block contains multiple convolutional layers, batch normalization layers, and ReLU activation layers. Dilated convolution operations are introduced into the network design. By setting different dilation rates, the convolutional kernels can cover receptive field ranges of varying sizes, thereby achieving effective capture of targets at different scales in the images.

[0024] During feature extraction, the first convolutional layer of the network uses small-sized kernels to extract fine texture features and local edge information from the image. The middle convolutional layers use medium-sized kernels to extract intermediate semantic features and regional structural information. Subsequent convolutional layers use larger kernels or dilated convolutions to extract global contextual information and overall layout features. This multi-layered, multi-scale feature extraction mechanism enables the acquisition of rich, hierarchically expressed image features.

[0025] The network performs feature aggregation on the feature maps output from each convolutional layer, generating a global feature representation including global semantic information and a local texture feature representation including local texture details. The feature aggregation operation combines spatial attention and channel attention mechanisms. Spatially, global average pooling and global max pooling operations capture the spatial statistics of the feature maps, which are then concatenated and passed through convolutional layers to generate a spatial attention weight map. In the channel dimension, global average pooling compresses the feature maps along the spatial dimension to obtain a channel-dimensional global description vector, which is then passed through fully connected layers to generate channel attention weight vectors.

[0026] Finally, the original feature map is multiplied element-wise with the spatial attention weights and channel attention weights to obtain the attention-weighted aggregated feature representation. In this aggregated feature representation, global features represent the overall semantic category information and spatial layout context of the encoded image, while local texture features represent the fine-grained texture patterns and microstructural features of the encoded lesion region.

[0027] In the device perception feature normalization stage, the acquired multi-scale image features are input into the device perception feature normalization module. This module adjusts the feature distribution of the device type identifiers corresponding to imaging equipment in different medical institutions to address the underlying feature distribution offset caused by differences in imaging equipment brand, model, and parameter settings. The device type identifier is used to characterize the device model information corresponding to the original acquisition device of the medical image data to be diagnosed. This identifier can be a predefined device category number or a device serial number extracted from the metadata information of the image file.

[0028] The device-aware feature normalization module comprises three functional components: a device embedding vector generation unit, a feature distribution estimation unit, and a feature mapping unit. The device embedding vector generation unit is responsible for generating a corresponding device embedding vector based on the device type identifier. This embedding vector is a high-dimensional dense vector that is learned and mapped from the device type identifier space to the feature space. The dimension of the embedding vector is typically set to match the channel dimension of the image features to ensure dimensionality consistency in subsequent feature stitching operations.

[0029] In the feature stitching process, multi-scale image features are stitched together with the device embedding vector. Specifically, the device embedding vector is first broadcast and copied to a shape that matches the spatial dimension of the multi-scale image features, and then concatenated with the multi-scale image features along the channel dimension. The stitched feature representation includes both the content information of the original image and the device origin information, forming an enhanced feature representation for device perception.

[0030] The feature distribution estimation unit receives the enhanced feature representation perceived by the device as input and estimates the distribution parameters of this feature representation. This unit first performs global average pooling along the spatial dimension to obtain the overall statistical vector of the features; then, it performs a nonlinear transformation on the statistical vector through a two-layer fully connected network to output the target distribution parameters. These target distribution parameters include a mean vector and a standard deviation vector, used to describe the target distribution characteristics that the normalized features should possess.

[0031] The feature mapping unit performs feature space mapping on the enhanced feature representation perceived by the device based on the target distribution parameters. This unit first calculates the deviation between the enhanced feature representation and the target mean vector, then standardizes it by dividing by the target standard deviation vector to obtain the normalized feature representation. To avoid excessively limiting the feature expressive power of the normalization operation, the feature mapping unit also introduces learnable scaling and offset parameters to perform an affine transformation on the normalized feature representation. The parameters of this affine transformation are adaptively learned through the network training process. The final output normalized multi-scale fused features have uniform distribution characteristics and no longer significantly depend on the device source of the original image data.

[0032] In the fine-grained texture alignment stage, the normalized multi-scale fused features are input into the fine-grained texture alignment network. Based on this network, adaptive texture enhancement processing is performed on the normalized local texture feature representation to generate aligned fine-grained diagnostic features. The internal structure of the fine-grained texture alignment network includes three functional components: a multi-level texture perception module, an attention weight generation module, and a feature enhancement module.

[0033] The multi-level texture perception module receives normalized local texture feature representations as input and extracts multi-scale texture pattern features through a multi-level convolutional structure. This module adopts a pyramid-shaped feature extraction architecture, setting multiple branch paths of different depths. Each branch path uses convolutional kernels of different sizes and different dilation rates, thereby capturing texture pattern features at different scales. Specifically, the first branch path uses small-sized convolutional kernels and a small dilation rate to focus on extracting fine-grained micro-texture patterns; the second branch path uses medium-sized convolutional kernels and a medium dilation rate to extract medium-scale texture structure features; and the third branch path uses large-sized convolutional kernels and a large dilation rate to extract coarse-grained macro-texture information.

[0034] The attention weight generation module analyzes the multi-scale texture pattern features output by the multi-level texture perception module and generates attention weight vectors corresponding to each texture pattern. This module first upsamples the texture features output from different branch paths to a uniform resolution, and then concatenates them along the channel dimension to form a multi-scale fused texture feature representation. Subsequently, the attention weight distribution along the channel and spatial dimensions is calculated using a channel attention sub-network and a spatial attention sub-network, respectively. The channel attention sub-network generates importance weights for each channel through global pooling and fully connected operations; the spatial attention sub-network generates importance weights for each spatial location through spatial convolution operations.

[0035] Ultimately, the attention weight vector integrates attention information from both the channel dimension and the spatial dimension.

[0036] The feature enhancement module performs weighted fusion processing on multi-scale texture pattern features based on the attention weight vector to generate aligned fine-grained diagnostic features. This module first multiplies the texture features of each branch path element-wise with the corresponding attention weight vector to enhance important texture features and weaken secondary texture features. Then, it adds the weighted features of each branch pixel-wise to achieve the organic fusion of multi-scale texture information.

[0037] During the fusion process, a learnable fusion weight parameter is used to dynamically adjust the contribution ratio of features at different scales. This fusion weight parameter is obtained through adaptive learning during the training process and can automatically determine the optimal combination of texture features at each scale according to the specific diagnostic task requirements.

[0038] In the diagnostic reasoning stage, the aligned fine-grained diagnostic features are input into the diagnostic reasoning network. Based on this network, the disease diagnosis results corresponding to the medical image data to be diagnosed and the confidence values ​​corresponding to the disease diagnosis results are output. The internal structure of the diagnostic reasoning network includes two functional components: a disease classification unit and a confidence assessment unit.

[0039] The disease classification unit receives fine-grained diagnostic features as input and outputs classification scores for each preset disease category through a fully connected layer structure. This unit first performs global average pooling on the input features, compressing spatial dimensional information into a feature vector. Then, it performs a non-linear transformation on the feature vector through a multi-layer fully connected network, ultimately outputting a classification score vector that matches the number of preset disease categories. Each classification score represents the confidence level that the input image belongs to the corresponding disease category.

[0040] Based on the classification scores of each preset disease category, the predicted probability distribution of each preset disease category is calculated using the softmax function. The softmax function performs exponential normalization on the classification score vector, mapping any real value to a probability value between zero and one, and the sum of all probability values ​​equals one. In other words, this predicted probability distribution intuitively expresses the likelihood of the input image belonging to each disease category.

[0041] The disease category corresponding to the maximum probability value in the predicted probability distribution is determined as the disease diagnosis result. This diagnosis result is the final disease judgment conclusion of the input medical image and can be directly used to assist medical diagnosis decision-making.

[0042] The confidence assessment unit performs uncertainty quantification on the predicted probability distribution to generate a confidence value corresponding to the disease diagnosis result. The uncertainty quantification adopts the Bayesian approximation method, which measures the degree of uncertainty of the diagnosis result by calculating the entropy value of the predicted probability distribution.

[0043] Specifically, the information entropy of the predicted probability distribution is first calculated; then, an exponential transformation is used to map the entropy value to a confidence interval between zero and one. A lower entropy value indicates a more concentrated predicted probability distribution, higher certainty of the diagnostic result, and a correspondingly higher confidence value; a higher entropy value indicates a flatter predicted probability distribution, greater uncertainty of the diagnostic result, and a correspondingly lower confidence value.

[0044] During the model training phase, the image feature extraction network, device perception feature normalization module, fine-grained texture alignment network, and diagnostic inference network optimize parameters through joint training. During training, multiple sample medical image data, along with corresponding sample device type identifiers and sample diagnostic labels, are first acquired.

[0045] Each sample of medical image data undergoes preprocessing and feature extraction to generate corresponding multi-scale image features. These features are then input into an initial device perception feature normalization module, where the feature distribution is adjusted based on the device type identifier of each sample, generating normalized multi-scale fusion features. These fusion features are then input into an initial fine-grained texture alignment network to generate aligned fine-grained diagnostic features. Finally, these diagnostic features are input into an initial diagnostic inference network, outputting the predicted diagnostic results for each sample of medical image data. Based on the classification error between the predicted diagnostic results and the sample diagnostic labels, a first loss value is calculated. This first loss value uses the cross-entropy loss function to measure the degree of difference between the predicted probability distribution and the true label distribution.

[0046] Simultaneously, based on the feature distribution differences among the device type identifiers corresponding to each sample of medical image data, a second loss value is calculated using the maximum mean difference metric. First, the normalized features corresponding to different device type identifiers are mapped to a regenerating kernel Hilbert space. Then, the Euclidean distance between the mean embedding vectors of each device type feature distribution in this space is calculated; this distance is the maximum mean difference metric. During training, this metric is used as the second loss value, and minimizing this loss value helps medical image data from different device sources achieve distribution alignment in the feature space.

[0047] The total loss is calculated based on a weighted combination of the first and second loss values, and the model parameters of each network module are adjusted using the backpropagation algorithm. The weight allocation ratio is controlled through hyperparameter settings to balance the two training objectives of improving classification accuracy and enhancing device independence. The training process continues iteratively until a preset training stopping condition is reached, such as when the classification accuracy on the validation set no longer improves or when a preset training epoch limit is reached. The network model obtained after training is the trained network model, which can be used to diagnose diseases from new medical image data.

[0048] The following is a complete execution flow of this method in a specific application scenario.

[0049] In this scenario, a lung computed tomography (CT) scan image of the lungs, collected by a top-tier hospital, is received.

[0050] The image data is transmitted to the diagnostic system of the present invention via a DICOM format file. First, the device type identification information is extracted from the metadata of the DICOM file. This identification indicates that the image was acquired by a certain type of multi-slice spiral CT device.

[0051] The lung CT image was preprocessed. The grayscale normalization process standardized the grayscale range of the image to a preset normalization interval. The size normalization process adjusted the image to a standard slice size. The noise suppression process used an adaptive filtering algorithm to eliminate electronic noise and artifact interference in the image.

[0052] The preprocessed standardized images are input into an image feature extraction network based on a deep residual network architecture. This network extracts multi-scale features of the images through a multi-layer convolutional structure, including shallow edge texture features, mid-layer lung lobe structure features, and deep semantic features. The network aggregates these features and outputs multi-scale image features containing global feature representations and local texture feature representations.

[0053] Multi-scale image features are concatenated with device embedding vectors to form a device-aware enhanced feature representation. The device embedding vectors are generated by the device embedding vector generation unit based on the CT device model identifier. The feature distribution estimation unit analyzes the distribution characteristics of the enhanced feature representation and outputs target distribution parameters. The feature mapping unit normalizes the enhanced features based on the target distribution parameters to generate normalized multi-scale fused features.

[0054] The local texture features in the normalized fusion features are input to a fine-grained texture alignment network. A multi-level texture perception module extracts multi-scale texture patterns from the images, including lung texture, vascular texture, and possible nodule texture. An attention weight generation module evaluates the importance of each texture pattern to the diagnostic task and generates an attention weight distribution. The feature enhancement module performs weighted fusion of texture patterns according to the attention weights, outputting fine-grained diagnostic features.

[0055] Fine-grained diagnostic features are input into the diagnostic inference network. The disease classification unit outputs classification scores indicating whether the image belongs to any of the preset disease categories, including normal, pulmonary nodules, pneumonia, and lung tumors. The predicted probability distribution is calculated using the softmax function, and the pulmonary nodule category shows the highest predicted probability; therefore, the disease diagnosis is determined to be a pulmonary nodule. The confidence assessment unit calculates the entropy of the predicted probability distribution and outputs the confidence value. A high confidence value indicates that the system has a high degree of certainty regarding the diagnostic result.

[0056] Ultimately, the system outputs the diagnostic results of lung nodules and their corresponding confidence scores to assist doctors in making clinical diagnostic decisions.

[0057] Example 2 This embodiment describes the application of this method in ultrasound imaging diagnostic scenarios.

[0058] The system acquires abdominal ultrasound image data collected by a primary healthcare institution. Ultrasound images are characterized by their real-time nature, high noise levels, and image quality significantly affected by the operator's skill level. The system obtains the raw ultrasound data frame stream from the digital interface of the ultrasound equipment and extracts the equipment type identification information.

[0059] The ultrasound image preprocessing workflow is specifically designed for its imaging characteristics. The grayscale normalization process employs an adaptive gain control algorithm to adjust the echo intensity distribution of the ultrasound images, ensuring consistent contrast. The size normalization process uniformly adjusts images acquired under different probe emission parameters to a standard display size. The noise suppression process uses a speckle removal algorithm based on coherence factors to effectively suppress speckle noise unique to ultrasound images while preserving echo information of tissue boundaries.

[0060] The ultrasound image feature extraction network employs a lightweight deep convolutional neural network architecture to meet the application requirements of large ultrasound image data volumes and high real-time processing demands. The network extracts tissue texture features, vascular structure features, and potential mass echo features from ultrasound images through multi-scale convolutional structures. The feature aggregation module integrates these multi-scale features, outputting multi-scale image features that include global feature representations and local texture feature representations.

[0061] The device perception feature normalization module generates device embedding vectors based on the probe type and device model of the ultrasound equipment. Compared with large tertiary hospitals, the ultrasound equipment used in primary healthcare institutions may have significant differences in imaging parameters, probe frequency, and signal processing algorithms. The module effectively normalizes the feature distribution deviation caused by the differences in equipment by estimating the distribution parameters and spatial mapping the features, so that ultrasound images from different devices can be represented in a unified feature space.

[0062] A fine-grained texture alignment network adaptively enhances normalized local texture features. The network's multi-level texture perception module extracts tissue texture patterns from ultrasound images, including liver parenchyma texture, gallbladder wall texture, and blood vessel wall texture. The attention weight generation module focuses on texture regions relevant to lesion detection, such as areas of increased echogenicity, decreased echogenicity, and irregular boundaries. The feature enhancement module generates fine-grained diagnostic features that effectively distinguish between normal and diseased tissues through weighted fusion of these key texture patterns.

[0063] The diagnostic reasoning network classifies diseases based on fine-grained diagnostic features. Preset abdominal ultrasound diagnostic categories include normal, diffuse liver lesions, gallstones, gallbladder polyps, and abdominal masses. It outputs disease diagnosis results and corresponding confidence scores, providing AI-assisted support for ultrasound diagnosis in primary healthcare institutions and effectively addressing the shortage of professional ultrasound physicians in these institutions.

[0064] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered illustrative and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the scope of the invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0065] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment includes only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. An AI-powered intelligent diagnostic method for medical images, characterized in that, include: Acquire medical image data to be diagnosed, and input the medical image data to be diagnosed into a pre-trained image feature extraction network to obtain multi-scale image features corresponding to the medical image data to be diagnosed; the multi-scale image features include global feature representation and local texture feature representation; The multi-scale image features are input into the device perception feature normalization module. Based on the device perception feature normalization module, the feature distribution of the device type identifiers corresponding to the imaging devices of different medical institutions is adjusted to generate normalized multi-scale fusion features. The normalized multi-scale fusion features are input into a fine-grained texture alignment network. Based on the fine-grained texture alignment network, adaptive texture enhancement processing is performed on the normalized local texture feature representation to generate aligned fine-grained diagnostic features. The aligned fine-grained diagnostic features are input into the diagnostic inference network, and the disease diagnosis result corresponding to the medical image data to be diagnosed and the confidence value corresponding to the disease diagnosis result are output based on the diagnostic inference network.

2. The method according to claim 1, characterized in that, The process of acquiring medical image data to be diagnosed and inputting the medical image data to be diagnosed into a pre-trained image feature extraction network to obtain multi-scale image features corresponding to the medical image data to be diagnosed includes: The medical image data to be diagnosed is subjected to image preprocessing operations to obtain standardized medical images; the image preprocessing operations include grayscale normalization, size normalization, and noise suppression. The standardized medical images are input into a preset multi-scale convolutional neural network, and feature maps of different receptive field sizes are extracted through the multi-layer convolutional structure of the multi-scale convolutional neural network. A feature aggregation operation is performed on each of the feature maps to generate a global feature representation including global semantic information and a local texture feature representation including local texture details.

3. The method according to claim 1, characterized in that, The device perception feature normalization module includes a device embedding vector generation unit, a feature distribution estimation unit, and a feature mapping unit. The step of inputting the multi-scale image features into the device perception feature normalization module, and adjusting the feature distribution of the device type identifiers corresponding to imaging devices in different medical institutions based on the device perception feature normalization module to generate normalized multi-scale fusion features includes: Based on the device type identifier, a device embedding vector corresponding to the device type identifier is generated by the device embedding vector generation unit; The multi-scale image features are concatenated with the device embedding vector to obtain an enhanced feature representation for device perception. Based on the feature distribution estimation unit, the distribution parameters of the enhanced feature representation perceived by the device are estimated to generate target distribution parameters; Based on the feature mapping unit, the enhanced feature representation perceived by the device is mapped in feature space according to the target distribution parameters to generate normalized multi-scale fusion features.

4. The method according to claim 1, characterized in that, The fine-grained texture alignment network includes a multi-level texture perception module, an attention weight generation module, and a feature enhancement module; The process involves inputting the normalized multi-scale fusion features into a fine-grained texture alignment network, and then performing adaptive texture enhancement processing on the normalized local texture feature representation based on the fine-grained texture alignment network to generate aligned fine-grained diagnostic features, including: The normalized local texture feature representation is input into the multi-level texture perception module, and multi-scale texture pattern features are extracted through the multi-level texture perception module. The attention weight generation module analyzes the multi-scale texture pattern features and generates attention weight vectors corresponding to each texture pattern. Based on the feature enhancement module, the multi-scale texture pattern features are weighted and fused according to the attention weight vector to generate aligned fine-grained diagnostic features.

5. The method according to claim 1, characterized in that, The diagnostic reasoning network includes disease classification units and confidence assessment units; The step of inputting the aligned fine-grained diagnostic features into a diagnostic inference network, and outputting the disease diagnosis result corresponding to the medical image data to be diagnosed and the confidence value corresponding to the disease diagnosis result based on the diagnostic inference network, includes: The aligned fine-grained diagnostic features are input into the disease classification unit, and the classification scores of each preset disease category are output through the fully connected layer structure of the disease classification unit. Based on the classification scores of each preset disease category, the predicted probability distribution of each preset disease category is calculated using the softmax function; The disease category corresponding to the maximum probability value in the predicted probability distribution is determined as the disease diagnosis result; The confidence assessment unit performs uncertainty quantification on the predicted probability distribution to generate a confidence value corresponding to the disease diagnosis result.

6. The method according to claim 1, characterized in that, The training steps for the image feature extraction network, the device-aware feature normalization module, the fine-grained texture alignment network, and the diagnostic inference network include: Acquire medical imaging data of multiple samples, along with corresponding sample device type identifiers and sample diagnostic labels; The medical image data of each sample is input into the initial image feature extraction network to obtain the multi-scale image features of each sample medical image data; The multi-scale image features of each sample are input into the initial device perception feature normalization module, and the feature distribution is adjusted based on the device type identifier of each sample to generate the multi-scale fusion features after sample normalization. The normalized multi-scale fusion features of each sample are input into the initial fine-grained texture alignment network to generate fine-grained diagnostic features after sample alignment. The fine-grained diagnostic features of each sample after alignment are input into the initial diagnostic inference network, and the predicted diagnostic results corresponding to the medical image data of each sample are output. Based on the classification error between the predicted diagnostic results corresponding to the medical image data of each sample and the diagnostic label of the sample, a first loss value is calculated. The second loss value is calculated based on the feature distribution differences between the device type identifiers corresponding to the medical image data of each sample. Based on the first loss value and the second loss value, the model parameters of each network module are adjusted, and iterative training continues until the training stopping condition is met, resulting in the trained network model.

7. The method according to claim 6, characterized in that, The second loss value is calculated based on the following method: Calculate the maximum mean difference measure among the normalized features corresponding to different device type identifiers; The second loss value is determined based on the maximum mean difference metric, so that medical image data from different devices can be aligned in the feature space during training.

8. The method according to claim 1, characterized in that, The medical imaging data to be diagnosed includes: computed tomography (CT) images, magnetic resonance imaging (MRI) images, ultrasound images, X-ray images, and positron emission tomography (PET) images.

9. The method according to claim 1, characterized in that, The image feature extraction network is a deep convolutional neural network; the device perception feature normalization module is an attention-based feature normalization network; the fine-grained texture alignment network is a multi-scale feature fusion network; and the diagnostic inference network is a fully connected neural network.

10. An AI-powered intelligent diagnostic system for medical images, characterized in that, include: The image acquisition module is used to acquire medical image data to be diagnosed and input the medical image data to be diagnosed into a pre-trained image feature extraction network to obtain multi-scale image features corresponding to the medical image data to be diagnosed; the multi-scale image features include global feature representation and local texture feature representation; The feature normalization module is used to input the multi-scale image features into the device perception feature normalization module, and adjust the feature distribution of the device type identifiers corresponding to the imaging devices of different medical institutions based on the device perception feature normalization module to generate normalized multi-scale fusion features; the device type identifier is used to characterize the device model information corresponding to the original acquisition device of the medical image data to be diagnosed. The texture alignment module is used to input the normalized multi-scale fusion features into the fine-grained texture alignment network, and perform adaptive texture enhancement processing on the normalized local texture feature representation based on the fine-grained texture alignment network to generate aligned fine-grained diagnostic features. The diagnostic output module is used to input the aligned fine-grained diagnostic features into the diagnostic inference network, and output the disease diagnosis result corresponding to the medical image data to be diagnosed and the confidence value corresponding to the disease diagnosis result based on the diagnostic inference network.

Images