Craniocerebral ultrasonic image information processing method, system and program product

Abstract

The invention provides a craniocerebral ultrasonic image information processing method, which is executed by a computer or a special processing device and comprises the following steps of: receiving craniocerebral ultrasonic image data; performing image enhancement processing on the received craniocerebral ultrasonic image data; analyzing the data after image enhancement processing by using a deep learning model based on a convolutional neural network so as to automatically detect an area indicating a lesion in the image data and identify information indicating a lesion type; according to the analysis result, analysis result data containing information indicating the position, size, form and the like of the lesion area is generated; according to the analysis result data, generating suggestion data used for assisting in obtaining the diagnosis information; and outputting the analysis result data and the suggestion data. The problems of low diagnosis accuracy, low automation degree, insufficient data processing capability and the like in the prior art are effectively solved, auxiliary diagnosis information is provided for clinical doctors, and the doctors can make diagnosis decisions more quickly and more accurately.

Description

Technical Field

[0001] The present invention relates to the field of medical imaging technology, and in particular to a real-time intelligent analysis and diagnosis system based on cranial ultrasound. Background Art

[0002] Cranial ultrasound is a noninvasive, safe, and economical medical imaging technique that has been widely used to detect brain diseases such as cerebral hemorrhage, cerebral edema, and brain tumors. However, existing cranial ultrasound image analysis technologies have the following limitations. First, diagnostic accuracy depends on the operator's experience, and image quality and diagnostic results are easily affected by subjective factors. The lack of standardized analysis tools leads to variability in diagnostic results. Second, operability and standardization are low, and the quality of image data collected by different operators varies, leading to discrepancies in clinical interpretation. The lack of standardized assessment methods hinders accurate diagnosis. Furthermore, traditional ultrasound image analysis requires lengthy manual operation and analysis, resulting in low diagnostic efficiency, especially in complex cases. Existing technologies have a low degree of automation, making it difficult to achieve real-time intelligent image analysis. The diagnostic process relies on manual operation and cannot quickly obtain information about the lesion area. Existing technologies lack data processing capabilities, making it difficult to effectively process complex two-dimensional images. In particular, they face difficulties in detecting and distinguishing subtle lesions, which affects the accuracy of early disease detection and diagnosis. Existing equipment lacks intelligent algorithms, cannot automatically mark lesions, and lacks decision-making support, requiring extensive manual intervention by physicians.

[0003] Therefore, there is an urgent need for a cranial ultrasound image information processing method that can improve diagnostic accuracy, automation and data processing capabilities to overcome the shortcomings of existing technologies.

[0004] It should be noted that the information disclosed in the above background technology section is only used to enhance the understanding of the background of the present invention, and therefore may include information that does not constitute prior art known to ordinary technicians in this field. Summary of the Invention

[0005] In view of the problems of low diagnostic accuracy, low degree of automation, and insufficient data processing capabilities in existing cranial ultrasound image analysis technology, the present invention provides a cranial ultrasound image information processing method, which aims to improve diagnostic accuracy, enhance diagnostic efficiency, and enhance the accessibility of medical resources to assist doctors in making faster and more accurate diagnostic decisions.

[0006] An embodiment of the present invention provides a method for processing cranial ultrasound image information, which is executed by a computer or a dedicated processing device and includes the following steps:

[0007] receiving cranial ultrasound image data obtained by performing a cranial ultrasound scan on a living human body;

[0008] Performing image enhancement processing on the received cranial ultrasound image data;

[0009] Analyze the enhanced image data using a deep learning model based on a convolutional neural network to automatically detect areas in the image data indicating lesions and identify information indicating the type of lesion.

[0010] Based on the analysis results, analysis result data including information such as the location, size, and morphology of the lesion area is generated;

[0011] Generate a recommendation report to assist in obtaining diagnostic information based on the analysis result data;

[0012] Output analysis result data and recommendation report.

[0013] In some optional embodiments, the image enhancement processing includes at least one of denoising, adaptive filtering and edge detection.

[0014] In some optional embodiments, identifying information indicating the type of lesion includes identifying information indicating the type of lesion including cerebral hemorrhage, cerebral edema, or brain tumor.

[0015] In some optional embodiments, analysis result data including information such as the location, size, and morphology of the indicated lesion area is generated, including calculating the area or volume of the indicated lesion area.

[0016] In some optional embodiments, the area indicating the lesion region is calculated by counting the number of pixels in the pixel set indicating the lesion region.

[0017] In some optional embodiments, the volume of the indicated lesion region is calculated by calculating the area of ​​the indicated lesion region in each slice image and summing the areas.

[0018] In some optional embodiments, generating a recommendation report for assisting in obtaining diagnostic information includes generating a recommendation report based on indicative data in combination with preset rules or a pre-trained machine learning model.

[0019] In some optional embodiments, the following steps are further included:

[0020] Upload the cranial ultrasound image data and / or analysis result data to the cloud platform.

[0021] In some optional embodiments, the following steps are also included: receiving feedback information provided by the cloud platform regarding the data uploaded to the cloud platform; and adjusting the deep learning model according to the feedback information.

[0022] The present invention also provides a cranial ultrasound image information processing system, comprising:

[0023] An input interface for receiving cranial ultrasound image data;

[0024] An image processing module, coupled to the input interface, for performing image enhancement processing on cranial ultrasound image data;

[0025] an artificial intelligence analysis module, coupled to the image processing module and configured with a deep learning model based on a convolutional neural network, for receiving the image enhancement processed data, performing analysis to automatically detect areas indicative of lesions in the image data, and identifying information indicative of the type of lesion;

[0026] an analysis result generation module, coupled to the artificial intelligence analysis module, for generating analysis result data including information indicating the location, size, morphology, etc. of the lesion area based on the analysis results of the artificial intelligence analysis module;

[0027] an auxiliary information generation module, coupled to the analysis result generation module, for generating, based on the analysis result data, suggestion data for assisting in obtaining diagnostic information;

[0028] The output module is coupled to the analysis result generation module and the auxiliary information generation module, and is used to output the analysis result data and the suggestion data, and the output includes displaying through a display screen.

[0029] The present invention also provides a computer program product comprising a non-transitory computer-readable storage medium having computer program instructions stored thereon. When the computer program instructions are executed by a processor, the processor executes the above-mentioned method for processing cranial ultrasound image information.

[0030] It is to be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the disclosure.

[0031] The present invention provides a method, system and program product for processing cranial ultrasound image information, which have the following features:

[0032] Beneficial effects:

[0033] The method for processing cranial ultrasound image information provided by the present invention performs image enhancement processing on cranial ultrasound image data and analyzes it using a deep learning model based on a convolutional neural network. It can automatically detect areas in the image data indicating lesions, identify the type of lesion, and generate analysis result data containing information such as the location, size, and morphology of the lesion area. Based on the analysis result data, recommendation data for assisting in obtaining diagnostic information is further generated. This method can reduce reliance on operator experience, improve diagnostic accuracy and efficiency, provide auxiliary diagnostic information to clinicians, and help doctors make diagnostic decisions more quickly and accurately, thereby potentially improving patient diagnosis and treatment outcomes. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Other features, objects and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings.

[0035] Figure 1 is a flow chart of a method for processing cranial ultrasound image information according to an embodiment of the present invention;

[0036] Figure 2 It is a structural diagram of a cranial ultrasound image information processing system according to an embodiment of the present invention. DETAILED DESCRIPTION

[0037] Example embodiments will now be described more fully with reference to the accompanying drawings. However, example embodiments can be embodied in many forms and should not be construed as limited to the examples set forth herein; rather, these embodiments are provided so that this disclosure will be thorough and complete and will fully convey the concepts of the example embodiments to those skilled in the art. The described features, structures, or characteristics may be combined in any suitable manner in one or more embodiments.

[0038] In addition, the accompanying drawings are merely schematic illustrations of the present disclosure and are not necessarily drawn to scale. Identical reference numerals in the figures denote identical or similar parts, and thus repetitive descriptions thereof will be omitted. Some of the block diagrams shown in the accompanying drawings are functional entities that do not necessarily correspond to physically or logically separate entities. These functional entities may be implemented in software, in one or more hardware modules or integrated circuits, or in different networks and / or processor devices and / or microcontroller devices.

[0039] The flowcharts shown in the accompanying drawings are merely exemplary and do not necessarily include all steps. For example, some steps may be decomposed, while some steps may be combined or partially combined. Therefore, the actual execution order may change according to actual circumstances.

[0040] The brain ultrasound imaging method of the present invention utilizes ultrasound waves to penetrate brain tissue and reconstruct images by measuring the reflected ultrasound signals. Different tissue structures have different reflective properties for ultrasound waves, resulting in different grayscale values ​​on the ultrasound image, reflecting the density and structural information of the tissue. Image enhancement technology aims to improve the visual quality of images, suppress noise, highlight edges, and increase image contrast, making them easier to observe and analyze. Convolutional neural networks are deep learning models that automatically learn features in images by simulating the connections between neurons in the human brain. They then extract local and global image information through operations such as convolution and pooling. These features can be used for tasks such as image classification, object detection, and image segmentation. By training on a large number of annotated brain ultrasound images, convolutional neural networks can learn the characteristics of different types of lesions and automatically detect and classify them. Based on the detected lesion areas, combined with preset rules or trained machine learning models, auxiliary diagnostic recommendations can be generated to provide decision support for doctors. The combined application of these technical principles can effectively improve the automation and accuracy of brain ultrasound image analysis, providing clinicians with faster and more reliable diagnostic evidence.

[0041] like Figure 1 As shown, the present invention provides a method for processing cranial ultrasound image information, which is executed by a computer or a dedicated processing device. The method includes the following steps:

[0042] S100. Receive brain ultrasound image data. This embodiment of the present invention uses an ultrasound probe to scan the human brain to obtain ultrasound image data reflecting the brain tissue structure. This data can be two-dimensional slice images or three-dimensional volume data. Specifically, a high-resolution ultrasound probe is used to capture two-dimensional or three-dimensional brain ultrasound images. This process uses the principle of ultrasonic reflection to obtain tissue image information.

[0043] I(x,y)=f(U(x,y),G(x,y))

[0044] Where I(x,y) represents the pixel value of the ultrasound image; U(x,y) is the original acquisition data of the ultrasound signal; and G(x,y) is the mapping function related to the reflected wave and tissue density.

[0045] The embodiment of the present invention adopts a portable ultrasound device, integrates a high-performance processing chip, supports a flexible probe replacement design, and is suitable for different groups of people, such as newborns, adults, etc., which enhances the applicability and scalability of the embodiment of the present invention.

[0046] S200, perform image enhancement processing on the received cranial ultrasound image data. In the embodiment of the present invention, the original ultrasound image is preprocessed to improve the image quality, for example, to reduce noise, enhance contrast, highlight edges, etc. Common image enhancement methods include denoising, adaptive filtering and edge detection. These processes are not isolated from each other, and one or more combinations can be selected according to actual conditions. For example, denoising can be performed first, then contrast enhancement, and finally edge detection. Among them, denoising can be performed by methods such as median filtering, Gaussian filtering, wavelet transform, etc., or a denoising network based on deep learning, such as U-Net or a convolutional neural network with a similar architecture can be used. Adaptive filtering can be performed by methods such as Wiener filtering, Kalman filtering, etc., or adaptive filtering based on learning can be used. Edge detection can be performed by methods such as Sobel operator, Canny operator, Laplacian operator, etc. In the specific implementation process, for the speckle noise and low contrast problems unique to cranial ultrasound images, a speckle noise suppression network based on deep learning and contrast-limited adaptive histogram equalization (CLAHE) technology can be used to improve image quality.

[0047] S300. Analyzing the image-enhanced data using a deep learning model based on a convolutional neural network refers to automatically analyzing the image using a pre-trained convolutional neural network (CNN) model to automatically detect lesion areas and automatically identify lesion types, as well as identify information indicating lesion types. CNN models can employ a variety of architectures, such as instance segmentation networks like Mask R-CNN, YOLACT, and SOLOv2, or multi-task learning architectures like ResNet and EfficientNet, which use a shared feature extraction backbone. Model training requires a large amount of labeled data, including raw cranial ultrasound images and lesion information pre-processed by experienced radiologists or sonographers, such as lesion bounding boxes, pixel-accurate lesion masks, and lesion type labels. During training, appropriate data augmentation techniques, such as random rotation, scaling, flipping, brightness / contrast adjustment, and noise overlay, can be used to increase data diversity and improve model robustness. After model training is complete, it can be deployed on a computer or dedicated processing device for analyzing new cranial ultrasound images.

[0048] S400. Based on the analysis results, generate indication data including information indicating the lesion area. The indication data includes analysis result data indicating information such as the location, size, and morphology of the lesion area. This embodiment of the present invention performs quantitative analysis on the detected lesion area, for example, calculating the area, volume, location coordinates, shape compactness, circularity, eccentricity, smoothness, etc. of the lesion area. When calculating the area, the number of pixels in the lesion area pixel set can be calculated and multiplied by the actual physical area of ​​the pixels. When calculating the volume, the area of ​​the lesion area in each slice image can be calculated and summed.

[0049] S500. Generate a recommendation report based on the indicated data. In an embodiment of the present invention, generating a recommendation report including recommendation data for assisting in obtaining diagnostic information refers to generating auxiliary diagnosis or treatment recommendations based on quantitative analysis results and AI confidence, combined with preset rules or trained machine learning models, for example, recommending CT examination, MRI examination, follow-up, referral to a specialist, indicating certain possible clinical risks, etc. Outputting analysis result data and recommendation data refers to presenting the quantitative analysis results and auxiliary diagnosis recommendations to the doctor in the form of a structured report, so that the doctor can quickly review and adopt them.

[0050] Through the above steps, the automated analysis and intelligent diagnosis of cranial ultrasound images are realized, which improves the efficiency and accuracy of diagnosis.

[0051] In some embodiments, the image enhancement processing includes at least one of denoising, adaptive filtering and edge detection. In view of the speckle noise and low contrast problems of cranial ultrasound images, as well as the need to highlight the edges and structures of lesions, the image enhancement processing method of an embodiment of the present invention aims to improve the visual quality of ultrasound images, thereby providing clearer and easier-to-analyze images for subsequent lesion detection and identification. Denoising refers to reducing various noises present in ultrasound images, such as speckle noise, electronic noise, etc., to improve the signal-to-noise ratio of the image. Denoising can be performed using a variety of methods, including but not limited to: median filtering, Gaussian filtering, wavelet transform, non-local mean filtering, etc. For example, when using median filtering to remove noise, the specific formula is expressed as:

[0052] I denoised (x,y)=Median(I(x+i,y+j)

[0053] Where (i, j) is the pixel position within the filter window.

[0054] Adaptive filtering refers to automatically adjusting the parameters of the filter according to the local features of the image to achieve better denoising effects while retaining the detailed information of the image. Adaptive filtering can adopt a variety of methods, including but not limited to: Wiener filtering, Kalman filtering, minimum mean square error filtering, etc. Preferably, an adaptive Wiener filter based on the local mean and variance can be used to adjust the filter strength according to the texture information of the local image, so as to better balance the denoising effect and detail retention. Edge detection refers to extracting edge information in the image, such as tissue boundaries, lesion boundaries, etc., to highlight the structural features of the image. Adaptive filtering is used to reduce random noise in the image, and the formula is expressed as;

[0055]

[0056] Where G(i,j) is a weighted filter kernel that can adjust the filter strength according to the local characteristics of the signal.

[0057] Edge detection can be performed using a variety of methods, including but not limited to: Sobel operator, Canny operator, Laplacian operator, etc. Preferably, the Canny operator can be used for edge detection, which can effectively suppress noise and extract clear edges. The Sobel operator is used to detect the edge of an image, and the formula is:

[0058]

[0059] Among them, Gx and Gy are the Sobel kernels in the horizontal and vertical directions respectively.

[0060] Denoising, adaptive filtering, and edge detection are not mutually exclusive. They can be combined based on practical needs. For example, denoising can be performed first, followed by adaptive filtering, and finally edge detection to achieve optimal image enhancement. In practice, appropriate image enhancement methods can be selected based on the characteristics of different cranial ultrasound images.

[0061] In some embodiments, image enhancement also uses a denoising network based on deep learning, such as a convolutional neural network with a U-Net or similar structure. In view of the speckle noise and low contrast problems of cranial ultrasound images, and the need to highlight the edges and structures of lesions, a speckle noise suppression network based on deep learning is adopted. Compared with filters such as median filtering and BM3D, deep learning methods, for example, using U-Net or a convolutional neural network with a similar architecture, can be trained on a large amount of paired data such as noisy / noise-free, simulated or actual data to more effectively remove speckle while maximally retaining the edge and detail information of the image. The specific implementation can be an end-to-end convolutional network that inputs the original ultrasound image and outputs the denoised image. The following is the convolution operation formula of CNN and its parameter description. By learning a large number of noisy images and noise-free images, more effective noise suppression is achieved. Given an input image I and a convolution kernel K, the output O of the convolution operation can be expressed as:

[0062]

[0063] Among them, O(i,j) is the output of position (i,j) after the convolution operation; I(i+m,j+n) is the pixel value of position (i+m,j+n) in the input image I; K(m,n) is the mth and nth values ​​in the convolution kernel K.

[0064] Specifically, the input image I is a two-dimensional or three-dimensional ultrasound image, representing the original image to be convolved. The convolution kernel K is a small matrix used to slide on the image and extract specific features, such as edges, textures, etc. The output image O is the image after the convolution operation, which usually contains the feature information of the input image. The stride is the number of pixel steps that controls the convolution kernel each time it slides, usually 1 or 2. The stride affects the size of the output image. Padding is to add zero values ​​to the image boundary to ensure that the convolution kernel can also process at the boundary. In order to increase the nonlinearity of the network, convolutional neural networks usually add an activation function after the convolution layer. The activation function is ReLU (Rectified Linear Unit):

[0065] f(x)=max(0,x)

[0066] The pooling operation is used to reduce the size of the feature map, usually using max pooling or average pooling. For max pooling, the formula is as follows:

[0067]

[0068] The output layer classifies each lesion area through the Softmax function and outputs its probability distribution:

[0069]

[0070] Among them, Zc is the score of the corresponding category, P(y=c|I) is the probability of the classification result, and k is all categories.

[0071] In some embodiments, after denoising, adaptive contrast enhancement techniques, such as Contrast-Constrained Adaptive Histogram Equalization (CLAHE), are employed to enhance overall contrast, particularly detail in dark areas. In specific implementations, this can be further combined with gradient-based edge enhancement algorithms, such as anisotropic diffusion filtering or more advanced learning-based edge enhancement methods, to highlight the boundaries and internal structures of lesions.

[0072] Through the above-mentioned image enhancement processing, the noise in cranial ultrasound images can be effectively reduced, the contrast of the image can be enhanced, and the edges of the image can be highlighted, thereby improving the visual quality of the image, providing a more reliable basis for subsequent lesion detection and identification, thereby improving the accuracy of subsequent analysis using deep learning models.

[0073] In some embodiments, identifying information indicating the type of lesion includes identifying information indicating the type of lesion, including cerebral hemorrhage, cerebral edema, or brain tumor. Identifying information indicating the type of lesion refers to classifying the lesion area detected in the cranial ultrasound image through a deep learning model to determine the type of lesion to which it belongs. This step can help doctors quickly determine the nature of the lesion and provide a basis for subsequent diagnosis and treatment. Lesion types include, but are not limited to, cerebral hemorrhage, cerebral edema, and brain tumors. These lesions are common types of lesions seen in cranial ultrasound examinations and have different acoustic characteristics and clinical significance. Cerebral hemorrhage refers to bleeding in the brain tissue, which usually appears as a high-echo or mixed-echo area on an ultrasound image. Cerebral edema refers to an increase in water content in the brain tissue, which usually appears as a low-echo area on an ultrasound image. Brain tumors refer to tumors in the brain tissue, which appear as masses with uneven echoes on an ultrasound image. By learning from a large amount of labeled data, the deep learning model can automatically identify the characteristics of these lesions and accurately classify them. Preferably, instance segmentation models or multi-task learning architectures, such as Mask R-CNN, YOLACT, SOLOv2, ResNet, and EfficientNet, can be used to accurately segment and classify lesions. In practice, the number of lesion types identified can be increased or decreased based on actual clinical needs. For example, lesions such as cerebral infarction, brain abscess, and parasitic brain diseases can also be identified.

[0074] This invention uses a deep learning model to automatically identify lesion types in cranial ultrasound images, reducing reliance on physician experience and improving diagnostic accuracy and efficiency. Furthermore, automatic lesion classification helps physicians quickly determine the nature of the lesion, providing a basis for subsequent treatment planning, potentially improving patient outcomes.

[0075] In some embodiments, generating the indicative data includes calculating the area or volume of the lesion region. The embodiments of the present invention are intended to perform quantitative analysis on the detected lesion region in order to more comprehensively understand the characteristics of the lesion. Specifically, the lesion region information extraction is performed by extracting the location information of the lesion region (such as a bounding box) through the output of the last layer of the convolutional neural network.

[0076] Bounding Box=(min(x),min(y),max(x),max(y))

[0077] Where (x, y) is the pixel coordinate of the lesion area. The morphology of the lesion can be further accurately extracted through image segmentation methods (such as U-Net network).

[0078] Position refers to the coordinates of the lesion area in the cranial ultrasound image, which can be expressed as pixel coordinates or relative to a specific anatomical landmark. Size refers to the area or volume of the lesion area, reflecting the extent of the lesion. Area calculation can be achieved by counting the number of pixels contained in the lesion area. Volume calculation can be achieved for a three-dimensional ultrasound image, or by calculating and integrating the area of ​​multiple two-dimensional slice images. Morphology refers to the shape characteristics of the lesion area, such as circularity, irregularity, boundary clarity, etc. The area of ​​the lesion area can be calculated by calculating the number of pixels in the pixel set indicating the lesion area. Specifically, the pixel set of the lesion area can be accurately extracted by an image segmentation method such as a U-Net network, and then the number of pixels in the pixel set can be counted and the number of pixels multiplied by the actual area represented by a single pixel to obtain the area of ​​the lesion area. The volume of the lesion area can be calculated by calculating the area of ​​the lesion area in each slice image and summing them up. Specifically, the three-dimensional cranial ultrasound image can be first segmented into multiple two-dimensional slice images, and then the area of ​​the lesion area in each slice image is calculated. Finally, the areas of all slice images are summed to obtain the volume of the lesion area. Those skilled in the art will appreciate that the specific algorithm for calculating area or volume may be selected according to actual circumstances, and the present invention does not impose any limitation thereto.

[0079] By calculating the location, size, and morphology of the lesion, quantitative analysis of the lesion can be achieved, providing doctors with a more objective and accurate basis for diagnosis. Quantifying the size of the lesion can be used to determine its severity and development trend, while analyzing its morphology can assist in determining its nature.

[0080] In some embodiments, the area of ​​the area indicating the lesion region is calculated and determined by calculating the number of pixels in the pixel set indicating the lesion region. An embodiment of the present invention provides a method for calculating the area of ​​the lesion region. The pixel set indicating the lesion region refers to a set of pixel points representing the lesion region extracted from a cranial ultrasound image by methods such as image segmentation. Calculating the number of pixels in the pixel set refers to counting the number of pixel points contained in the set. Since each pixel point represents a certain physical area in the actual image, the area of ​​the lesion region can be obtained by multiplying the number of pixels by the actual area represented by a single pixel. In some embodiments, a U-Net network or a convolutional neural network with a similar structure can be used to perform pixel-level segmentation on the cranial ultrasound image, thereby accurately extracting the pixel set of the lesion region. In other embodiments, traditional image segmentation methods such as threshold segmentation and region growing can be used to extract the pixel set of the lesion region. The statistics of the number of pixels can be implemented by programming languages ​​such as Python, C++, etc., or by dedicated image processing software. Specifically, the area calculation formula is as follows:

[0081]

[0082] Wherein, B is the pixel set of the lesion area.

[0083] Determining lesion area based on pixel count is simple, computationally efficient, and meets the needs of real-time diagnosis. By precisely extracting the pixel set of the lesion area, a more accurate lesion area can be obtained, providing a more reliable basis for subsequent diagnosis.

[0084] In some embodiments, the volume of the indicative lesion region is calculated by calculating the area of ​​the indicative lesion region in each slice image and summing the areas. Embodiments of the present invention provide a method for calculating the volume of a lesion region, particularly suitable for three-dimensional cranial ultrasound images or three-dimensional images reconstructed from multiple two-dimensional slices. This step enables a more accurate assessment of the overall size of the lesion, thereby assisting physicians in making more accurate diagnoses and formulating treatment plans. This method discretizes the three-dimensional volume into multiple two-dimensional slices, calculates the lesion area on each slice, and sums these areas to approximate the lesion volume. This method is based on the concept of integration; when the slices are sufficiently thin, the approximation approaches the true volume. The area of ​​the indicative lesion region in each slice image is determined by calculating the number of pixels in the set of pixels indicative of the lesion region. Specifically, for each slice image, an image segmentation algorithm, such as a U-Net network, threshold segmentation, or region growing, is first used to extract a set of pixels representing the lesion region. The number of pixels in this set is then counted and the number of pixels is multiplied by the actual area represented by each pixel to obtain the area of ​​the lesion region in that slice image. After obtaining the lesion area for each slice image, the summation can be performed using a variety of methods. A simple method is to simply sum the areas of all slice images. Another more accurate method is to average the areas of two adjacent slice images, multiply them by the slice spacing, and then add all the results to estimate the volume more accurately. The formula can be expressed as:

[0085]

[0086] Where Area_z is the lesion area of ​​the z-th slice, and Δz is the slice spacing.

[0087] By calculating and summing the areas of the indicative lesion region across each slice image, a precise assessment of lesion volume can be achieved, overcoming the limitations of traditional methods that rely solely on subjective estimation based on two-dimensional images. Accurately quantifying lesion volume helps physicians determine lesion severity and progression, providing more reliable data support for developing personalized treatment plans.

[0088] In some embodiments, an instance segmentation model or a multi-task learning architecture is used to automatically detect the area indicating the lesion in the image data to detect the lesion location, classify the lesion type and obtain the precise morphology of the lesion area. Specifically, an instance segmentation network such as Mask R-CNN, YOLACT or SOLOv2 is used. This type of network can simultaneously complete the target detection output bounding box, target classification output category and pixel-level segmentation output Mask. For three-dimensional cranial ultrasound, it is extended to a three-dimensional instance segmentation network such as a 3D U-Net variant combined with a detection head or layer-by-layer processing of two-dimensional slices and then three-dimensional reconstruction. A shared feature extraction backbone such as ResNet or EfficientNet is used, and branches are connected to different heads: a target detection head such as RetinaNet or YOLOvX is used to output bounding boxes and category confidences, and a semantic or instance segmentation head such as a U-Net or DeepLab variant is used to output pixel-level areas. This method is highly flexible and can be optimized for different tasks. The specific training and optimization methods are as follows:

[0089] The training is performed using supervised learning.

[0090] Use data augmentation techniques such as random rotation, scaling, flipping, brightness / contrast adjustment, noise overlay, etc. to increase data diversity and improve model robustness.

[0091] Choose an appropriate loss function. For example, use Focal Loss + Smooth L1 Loss for detection tasks, Dice Loss or IoU Loss + Binary Cross-Entropy Loss for segmentation tasks, and Cross-Entropy Loss for classification tasks. Multi-task models require joint optimization of multiple loss functions.

[0092] Use effective optimizers such as Adam, SGD with Momentum, and learning rate scheduling strategies.

[0093] Leveraging transfer learning to fine-tune a backbone network pre-trained on large natural image or general medical image datasets can accelerate training and improve performance.

[0094] Monitor performance indicators on the validation set, such as average precision AP for detection / instance segmentation, Dice Coefficient for segmentation, and perform model selection and hyperparameter tuning.

[0095] Through the above detailed architecture and training scheme, we can build an intelligent analysis core that can automatically detect, classify and segment cranial brain lesions with high accuracy.

[0096] In some embodiments, generating a recommendation report for assisting in obtaining diagnostic information includes combining preset rules or trained machine learning models to generate a recommendation report based on information indicating the location, size, morphology, and other information of the lesion area. This step is intended to generate recommendations for assisting doctors in making diagnostic decisions based on the location, size, morphology, and other information of the lesion area obtained in the aforementioned steps, thereby improving the efficiency and accuracy of diagnosis. The construction of a decision model for generating a recommendation data report can be implemented in the following two ways: an expert system based on preset rules, or a trained machine learning model. An expert system based on preset rules refers to a series of diagnostic rules formulated in advance by clinical experts, which are based on medical knowledge and clinical experience, for example: "If a suspected brain tumor is detected, the size is >X cm 3 , and the morphology is irregular, an enhanced MRI scan is recommended. Based on these rules, the system determines the location, size, morphology, and other information of the lesion area to generate corresponding diagnostic recommendations. Specifically, the method steps for generating an interpretable quantitative analysis report and providing intelligent decision-making assistance recommendations are as follows;

[0097] Receive information indicating the type of lesion, including bounding box, mask, category, confidence level, etc.

[0098] For each detected lesion instance, a quantitative calculation is performed:

[0099] Area: For 2D images, count the number of pixels within the mask and multiply by the actual physical area of ​​the pixel (depending on the ultrasound probe parameters and depth settings). For 3D data, the area of ​​each slice can be calculated and then integrated.

[0100] Volume: For 3D volume data or 3D data reconstructed from multiple 2D slices, calculate the number of voxels within the lesion mask and multiply it by the actual physical volume of the voxel. If processing the 3D mask directly, calculate the number of voxels multiplied by the unit voxel volume.

[0101] Position: Output the pixel / voxel coordinates of the lesion center or convert them into anatomical relative positions, as well as the orientation description relative to important anatomical structures.

[0102] Morphology: Extract the morphological features of the lesion mask, such as compactness, circularity, eccentricity, smoothness of boundary, etc.

[0103] Integrate the above quantitative information, lesion category, AI confidence, etc. to automatically generate a structured report.

[0104] A trained machine learning model refers to a model that is trained on a large amount of labeled data through machine learning algorithms, such as support vector machines (SVM), random forests, gradient boosting trees, or small neural networks, to learn the mapping relationship between lesion characteristics and diagnostic recommendations. The input features of the model include information such as the location, size, and morphology of the lesion area, and the output of the model is a specific auxiliary diagnosis or treatment recommendation category. The recommendation data can include a variety of content, such as: recommendations for further examinations such as CT and MRI, recommendations for follow-up observations, recommendations for referral to a specialist, and indications of certain possible clinical risks. When generating data for a recommendation report, the confidence of the AI ​​model can be used as a reference. For example, for lesion areas with low confidence, a review or further examination can be recommended. Specifically, the decision function for the decision recommendation is:

[0105] Decision Suggestion=f(Area,Shape,Growth Rate)

[0106] Here, f is the decision function, which determines the subsequent examination or treatment recommendations through a rule engine or deep learning model.

[0107] Specifically, the steps for generating a decision-support recommendation report are as follows:

[0108] Based on the analysis results and AI confidence, the decision logic is triggered.

[0109] Decision model: The decision model can adopt one or a combination of the following:

[0110] Rule-based expert system: A set of pre-set clinical rules (developed by experts). For example: "If a suspected tumor is detected and the size is > X cm 3 If the brain is edematous and has an irregular shape, an enhanced MRI scan is recommended. If extensive cerebral edema is detected, it may indicate intracranial hypertension, and ICP monitoring or further evaluation is recommended.

[0111] Machine Learning Classifier: Train a classification model, such as a support vector machine (SVM), random forest, gradient boosting tree, or small neural network. The model's input features include quantitative AI output metrics such as area, volume, morphological characteristics, location, lesion category, and confidence level. The model's output is a specific auxiliary diagnosis or treatment recommendation, such as a CT scan recommendation, an MRI scan recommendation, a follow-up recommendation, a specialist referral recommendation, or an indication of a potential clinical risk.

[0112] The decision-making basis in the recommendation report can be further refined as follows:

[0113] The quantitative indicators of morphological features such as area, shape, and location (such as whether it is close to important blood vessels or ventricles) are used as input.

[0114] The Growth Rate indicator is configured as follows:

[0115] For a single scan: the current morphological characteristics of the lesion indicate signs of invasiveness or rapid growth, such as blurred and irregular boundaries, heterogeneous internal echoes, etc.

[0116] For multiple scans: We perform a longitudinal comparison of the patient's historical scan data, comparing lesion size and morphology, and calculating the true growth rate. Growth Rate then becomes an input feature based on historical data.

[0117] Recommendation output: Output clear, actionable recommendations with key findings that trigger the recommendations, such as lesion type, size, location, or morphological characteristics. Furthermore, a list of recommendations is output, sorted by priority.

[0118] By combining preset rules or trained machine learning models to generate relevant data for auxiliary diagnosis recommendation reports, doctors can be provided with more comprehensive and objective diagnostic information, thereby assisting doctors in making more accurate diagnostic decisions and optimizing treatment plans.

[0119] In some embodiments, the method for processing cranial ultrasound image information further includes the step of uploading the cranial ultrasound image data and / or analysis result data to a cloud platform. This step is intended to enable data sharing and remote collaboration, improving access to medical resources, particularly in resource-scarce or remote areas. Cranial ultrasound image data refers to raw image data obtained by scanning the living human brain with an ultrasound probe. Analysis result data refers to quantitative analysis results generated after processing using the method provided herein, including information such as the location, size, and morphology of the lesion area, as well as diagnostic assistance recommendations. A cloud platform refers to a remote server cluster built based on cloud computing technology with data storage and processing capabilities, providing functions such as data sharing, remote collaboration, and model updates. Uploading the cranial ultrasound image data and / or analysis result data to the cloud platform can be achieved in various ways, such as uploading the data to the cloud server via wireless networks such as Wi-Fi and 5G, or uploading the data to the cloud server via wired networks such as Ethernet. When uploading data, secure protocols such as HTTPS can be used to ensure data security and confidentiality. After uploading, the data can be stored in a cloud database such as MySQL or MongoDB. Preferably, the patient data can be processed in an anonymized or pseudonymized manner to protect the patient's privacy.

[0120] By uploading cranial ultrasound image data and / or analysis results to the cloud platform, data sharing and remote collaboration are achieved, allowing remote medical experts to view patients' scan data and diagnosis results in real time, providing timely diagnostic support to primary care doctors, thereby potentially improving the accuracy and efficiency of diagnosis and increasing the accessibility of medical resources.

[0121] In some embodiments, the method for processing cranial ultrasound image information further includes the steps of: receiving feedback from a cloud platform regarding data uploaded to the cloud platform; and adjusting the deep learning model based on the feedback. This step is intended to leverage the expertise of telemedicine experts to continuously optimize the deep learning model, thereby improving the model's diagnostic accuracy. Receiving feedback from telemedicine experts on the cloud platform refers to the telemedicine experts reviewing the patient's cranial ultrasound image data and / or analysis results through the cloud platform and providing feedback such as diagnostic opinions, annotations, or corrections. This feedback may include annotation corrections to lesion areas, reclassification of lesion types, and adoption or modification of auxiliary diagnostic recommendations. Feedback information can be transmitted via a secure network connection, such as the HTTPS protocol, and can be stored and managed in structured data formats such as JSON or XML. Adjusting the deep learning model based on the feedback refers to retraining or fine-tuning the deep learning model using the feedback provided by the telemedicine experts to better adapt the model to the needs of actual clinical applications. Model adjustment can be achieved using a variety of methods, such as fine-tuning the output layer or using incremental learning techniques. Model adjustments can be performed on cloud servers, leveraging the cloud's powerful computing resources to accelerate model training and improve performance. After the model is updated, the updated version can be deployed and updated on portable ultrasound devices. Expert feedback requires prior cleaning and verification to ensure data quality.

[0122] Specifically, through the expert feedback algorithm, combined with the opinions of remote experts, the system's decision model is automatically adjusted to improve diagnostic accuracy. For example:

[0123] Updated Model=f(Expert Feedback,Current Model)

[0124] Among them, Expert Feedback is used to fine-tune the deep learning model Current Model and improve the model's application effect in actual clinical practice.

[0125] By receiving feedback from remote medical experts and adjusting the deep learning model accordingly, the model is continuously optimized, improving its diagnostic accuracy and generalization capabilities, and promoting the deep integration of artificial intelligence and clinical practice. By incorporating the knowledge of remote experts into model training, the model's diagnostic capabilities in complex or rare cases are effectively improved, thereby providing patients with better medical services.

[0126] like Figure 2As shown, an embodiment of the present invention further provides a cranial ultrasound image information processing system for implementing the cranial ultrasound image information processing method of any of the above embodiments, comprising:

[0127] Input interface M100 is used to receive cranial ultrasound image data. This module is responsible for receiving cranial ultrasound image data collected by the ultrasound probe. Input interface M100 can be any type of interface, such as a USB interface, Ethernet interface, or wireless network interface, as long as it can achieve stable and high-speed data transmission.

[0128] Image processing module M200, coupled to input interface M100, is used to perform image enhancement processing on cranial ultrasound image data. This module is responsible for preprocessing the received cranial ultrasound image data to improve image quality and reduce noise interference, providing better input for subsequent lesion detection and identification. Image enhancement processing can adopt various methods, such as denoising, adaptive filtering, and edge detection.

[0129] The artificial intelligence analysis module M300, coupled to the image processing module M200 and equipped with a deep learning model based on a convolutional neural network, is used to receive image enhancement data and perform analysis to automatically detect areas in the image data indicating lesions and identify information indicating the type of lesion. This module is responsible for automatically analyzing the image enhancement data to achieve automatic detection of lesion areas and automatic identification of lesion types. This module is equipped with a deep learning model based on a convolutional neural network. By learning from a large amount of annotated data, this model can automatically extract features from images and accurately detect and classify lesion areas.

[0130] The analysis result generation module M400 is coupled to the artificial intelligence analysis module M300 and is used to generate analysis result data containing information indicating the location, size, morphology, and other information of the lesion area based on the analysis results of the artificial intelligence analysis module M300. This module is responsible for generating analysis result data containing information indicating the location, size, morphology, and other information of the lesion area based on the analysis results of the artificial intelligence analysis module M300, providing doctors with a more comprehensive and objective diagnosis basis.

[0131] The auxiliary information generation module M500, coupled to the analysis result generation module M400, is used to generate recommendation data to assist in obtaining diagnostic information based on the analysis result data. This module is responsible for generating recommendation data to assist in obtaining diagnostic information based on the analysis result data, such as recommendations for further examinations such as CT and MRI, follow-up observation, referral to a specialist, and possible clinical risks, providing reference for doctors' diagnostic decisions.

[0132] The output module M600, coupled to the analysis result generation module M400 and the auxiliary information generation module M500, is used to output the analysis result data and recommended data, including displaying them on a display screen. This module is responsible for outputting the analysis result data and recommended data output by the auxiliary information generation module M500, including displaying them on a display screen for easy viewing and use by the physician. Alternatively, the data can be printed out via a printer or transmitted to other devices via a network interface.

[0133] Through the collaborative work of the above modules, automated analysis and intelligent diagnosis of cranial ultrasound images are achieved, which improves the efficiency and accuracy of diagnosis, reduces the workload of doctors, and increases the accessibility of medical resources.

[0134] An embodiment of the present invention also provides a computer program product, which includes computer instructions. When the computer program instructions are executed by a processor, the processor executes the above-mentioned method for processing cranial ultrasound image information. It is intended to implement the cranial ultrasound image information processing method proposed by the present invention by software, thereby facilitating deployment and application. Computer program instructions refer to a sequence of instructions that can be executed by a computer or other special processing device to implement specific functions. In the present invention, computer program instructions are used to implement the above-mentioned method for processing cranial ultrasound image information, including steps such as image enhancement processing, lesion area detection, lesion type identification, analysis result generation, and auxiliary information generation. Computer program instructions can be written in a variety of programming languages, such as Python, C++, Java, etc. When the computer program instructions are executed by the processor, the processor can execute each step according to a predetermined process, thereby realizing the technical solution of the present invention.

[0135] By implementing the cranial ultrasound image information processing method proposed in this invention as a computer program product, it can be flexibly deployed and applied on various computers or dedicated processing devices, reducing implementation costs and facilitating the widespread application of the invention. This computer program product can be used to upgrade existing ultrasound equipment, providing it with intelligent analysis capabilities, extending its service life, and increasing its added value.

[0136] The above is a further detailed description of the present invention in conjunction with specific preferred embodiments, and the specific implementation of the present invention should not be considered to be limited to these descriptions. For those skilled in the art to which the present invention belongs, several simple deductions or substitutions can be made without departing from the concept of the present invention, and all of these should be considered to fall within the scope of protection of the present invention.

Claims

1. A method for processing cranial ultrasound image information, characterized in that: The method is executed by a computer or a dedicated processing device and includes the following steps: receiving cranial ultrasound image data; performing image enhancement processing on the cranial ultrasound image data; Analyzing the image-enhanced data using a deep learning model based on a convolutional neural network to automatically detect areas in the image data indicating lesions and identify information indicating the type of lesions; generating, based on the analysis result, indication data containing the indicated lesion area, wherein the indication data includes position, size, and morphological information of the indicated lesion area; A recommendation report is generated based on the indication data.

2. The method for processing cranial ultrasound image information according to claim 1, characterized in that: The image enhancement processing includes at least one of denoising processing, adaptive filtering and edge detection.

3. The method for processing cranial ultrasound image information according to claim 1, characterized in that: The identifying information indicating the type of lesion includes identifying information indicating the type of lesion including cerebral hemorrhage, cerebral edema or brain tumor.

4. The method for processing cranial ultrasound image information according to claim 1, wherein: The generating of the indication data includes calculating the area or volume of the indicated lesion region.

5. The method for processing cranial ultrasound image information according to claim 4, characterized in that: The area of ​​the indicated lesion region is calculated by calculating the number of pixels in the pixel set of the indicated lesion region.

6. The method for processing cranial ultrasound image information according to claim 4, characterized in that: The volume of the indicated lesion region is calculated by calculating the area of ​​the indicated lesion region in each slice image and summing the areas.

7. The method for processing cranial ultrasound image information according to claim 1, characterized in that: The generating of the recommendation report includes generating the recommendation report based on the indication data in combination with preset rules or a pre-trained machine learning model.

8. The method for processing cranial ultrasound image information according to claim 1, characterized in that: The following steps are also included: The brain ultrasound image data and / or the analysis result data are uploaded to a cloud platform.

9. A cranial ultrasound image information processing system, characterized in that: include: An input interface for receiving cranial ultrasound image data; an image processing module, coupled to the input interface, for performing image enhancement processing on the cranial ultrasound image data; an artificial intelligence analysis module, coupled to the image processing module and configured with a deep learning model based on a convolutional neural network, for receiving the image enhancement processed data, performing analysis to automatically detect areas indicative of lesions in the image data, and identifying information indicative of the type of lesion; an analysis result generating module, coupled to the artificial intelligence analysis module, for generating analysis result data including information such as the location, size, and morphology of the indicated lesion area based on the analysis result of the artificial intelligence analysis module; an auxiliary information generation module, coupled to the analysis result generation module, for generating, based on the analysis result data, suggestion data for assisting in obtaining diagnostic information; An output module is coupled to the analysis result generation module and the auxiliary information generation module, and is used to output the analysis result data and the suggestion data, wherein the output includes displaying through a display screen.

10. A computer program product, characterized in that The program product includes computer instructions, which, when executed by a processor, implement the steps of the method for processing cranial ultrasound image information according to any one of claims 1 to 8.

Images