An ultrasonic auxiliary diagnosis system for pediatric lymphoma combined with artificial intelligence

Abstract

The application relates to the field of medical image diagnosis, in particular to a pediatric lymphoma ultrasound auxiliary diagnosis system combining artificial intelligence, which comprises an ultrasonic image acquisition module, which is used for acquiring the ultrasonic image of the superficial lymph nodes of a patient; an image preprocessing module, which is connected with the ultrasonic image acquisition module and is used for standardizing the ultrasonic image to generate a pretreated image; a feature extraction module, which is connected with the image preprocessing module and comprises a convolution branch for extracting local texture features and a state space branch for extracting global context features, and generates a multi-scale feature map through fusion; a risk assessment module, which is connected with the feature extraction module, performs global average pooling and feature importance weighting on the multi-scale feature map to generate a lymphoma malignancy risk probability; and a result display module, which is connected with the risk assessment module, receives the risk probability and generates a diagnosis result report and a heat map visual display, and reduces errors caused by subjective judgment of doctors through quantitative analysis of ultrasonic image features.

Description

TECHNICAL FIELD

[0001] The application relates to the field of medical image diagnosis, in particular to an ultrasonic auxiliary diagnosis system for pediatric lymphoma in combination with artificial intelligence. BACKGROUND

[0002] As a safe and non-radiation medical imaging technology, ultrasonic diagnosis has become an important means for pediatric disease examination. Through sending ultrasonic waves to the body of a sick child, receiving reflected signals, and then generating high-resolution images, the internal structure changes of superficial lymphoid tissue can be clearly displayed, providing an important basis for early diagnosis of pediatric lymphoma.

[0003] However, current ultrasonic diagnosis of pediatric lymphoma mainly relies on the subjective experience of doctors for judgment, and there are problems such as non-uniform diagnosis standards and difficulty in identifying early lesions. Especially in primary medical institutions, due to the shortage of professional pediatric ultrasonic doctors, misdiagnosis or missed diagnosis is more likely to occur. In addition, the particularity of pediatric lymph nodes makes diagnosis more complicated, and doctors need to fully consider age-related normal variations.

[0004] In the prior art, although ultrasonic technology can clearly display the morphological features of pediatric lymph nodes, including important diagnostic indicators such as cortical thickness, lymphatic gate structure, and internal echo, in actual application, it still faces the problems of insufficient diagnostic accuracy and consistency. Therefore, it is urgent to develop an auxiliary diagnosis system that can combine artificial intelligence technology to improve the accuracy and efficiency of ultrasonic diagnosis of pediatric lymphoma. SUMMARY

[0005] The purpose of the present application is to provide an ultrasonic auxiliary diagnosis system for pediatric lymphoma in combination with artificial intelligence, which can intelligently analyze the features of ultrasonic images to provide objective and quantitative diagnostic references for doctors, and improve the diagnostic accuracy and consistency.

[0006] The application discloses an ultrasonic auxiliary diagnosis system for pediatric lymphoma in combination with artificial intelligence, which comprises:

[0007] An ultrasonic image acquisition module is used to acquire ultrasonic images of superficial lymph nodes of a sick child;

[0008] An image preprocessing module is connected with the ultrasonic image acquisition module and is used to standardize the ultrasonic images to generate preprocessed images;

[0009] A feature extraction module is connected with the image preprocessing module and is used to process the preprocessed images. The feature extraction module comprises a convolution branch and a state space branch. The convolution branch is used to extract local texture features, and the state space branch is used to extract global context features. The feature extraction module generates a multi-scale feature map by fusing the local texture features and the global context features.

[0010] a risk assessment module, connected with the feature extraction module, configured to receive the multi-scale feature map, perform global average pooling and feature importance weighting on the multi-scale feature map, and generate a lymphoma malignancy risk probability; and a result display module, connected with the risk assessment module, configured to receive the lymphoma malignancy risk probability, and generate a diagnosis result report and a heat map visualization display.

[0011] Preferably, the ultrasound image acquisition module comprises:

[0012] a high-frequency linear array probe configured to acquire high-resolution ultrasound images of superficial lymph nodes;

[0013] a convex array probe configured to acquire ultrasound images of relatively deep-positioned lymph nodes; and an image storage unit, connected with the high-frequency linear array probe and the convex array probe, configured to store static images, dynamic images, and color Doppler flow imaging images of lymph nodes.

[0014] Preferably, the image preprocessing module comprises:

[0015] a size unification unit configured to scale the ultrasound images to a predetermined size;

[0016] a noise suppression unit, connected with the size unification unit, configured to perform Gaussian blur processing on the ultrasound images of the predetermined size;

[0017] a data enhancement unit, connected with the noise suppression unit, configured to perform random occlusion, geometric transformation, and optical property enhancement on the ultrasound images; and a normalization processing unit, connected with the data enhancement unit, configured to perform pixel value mapping and statistical normalization processing on the enhanced ultrasound images.

[0018] Preferably, the feature extraction module further comprises:

[0019] a feature splitting unit configured to split an input feature map into a first part and a second part in a channel dimension, the first part being input to the convolution branch, and the second part being input to the state space branch; and a feature fusion unit, connected with the convolution branch and the state space branch, configured to perform channel dimension splicing and channel shuffling operation on output features of the convolution branch and output features of the state space branch, and generate fusion features.

[0020] Preferably, the convolution branch comprises:

[0021] a first convolution layer configured to apply a first convolution operation and batch normalization to the first part of the feature map;

[0022] a second convolutional layer connected with the first convolutional layer, configured to apply a second convolution operation, batch normalization and ReLU activation to an output of the first convolutional layer; and a third convolutional layer connected with the second convolutional layer, configured to apply a point-wise convolution operation, batch normalization and ReLU activation to an output of the second convolutional layer, to generate an output feature of the convolution branch.

[0023] Preferably, the state space branch comprises:

[0024] a layer normalization unit configured to perform layer normalization on the second part of feature maps;

[0025] a dual-path semantic extraction unit connected with the layer normalization unit, configured to generate a first semantic feature and a second semantic feature;

[0026] a multi-direction scanning unit connected with the dual-path semantic extraction unit, configured to perform multi-direction scanning on the second semantic feature to generate a global perception representation; and an information filtering unit connected with the dual-path semantic extraction unit and the multi-direction scanning unit, configured to interact the global perception representation with the first semantic feature to generate an output feature of the state space branch.

[0027] Preferably, the multi-direction scanning unit comprises:

[0028] an up-left to down-right scanning unit configured to expand the second semantic feature into a first sequence along an up-left to down-right direction;

[0029] a down-right to up-left scanning unit configured to expand the second semantic feature into a second sequence along a down-right to up-left direction;

[0030] an up-right to down-left scanning unit configured to expand the second semantic feature into a third sequence along an up-right to down-left direction; and a down-left to up-right scanning unit configured to expand the second semantic feature into a fourth sequence along a down-left to up-right direction.

[0031] The global perception representation is generated by merging results of processing the first sequence, the second sequence, the third sequence and the fourth sequence.

[0032] Preferably, the feature extraction module further comprises:

[0033] a patch merging unit connected with the feature fusion unit, configured to perform spatial merging and doubling of channel number on the fused feature to generate a down-sampled feature; and a multi-scale feature generation unit connected with the patch merging unit, configured to generate the multi-scale feature maps of different spatial resolutions through multi-level feature extraction and down-sampling operation.

[0034] As preferred, the risk assessment module comprises:

[0035] a dimension rearrangement unit, configured to rearrange dimensions of the multi-scale feature map from height, width, and channel number to channel number, height, and width;

[0036] a global pooling unit, connected to the dimension rearrangement unit, configured to apply global average pooling to each channel of the rearranged feature map to generate a feature vector;

[0037] a feature importance weighting unit, connected to the global pooling unit, configured to perform linear mapping on the feature vector to calculate a comprehensive risk score; and a probability output unit, connected to the feature importance weighting unit, configured to convert the comprehensive risk score into the lymphoma malignancy risk probability through a Sigmoid function.

[0038] As preferred, the result display module comprises:

[0039] a gradient calculation unit, configured to calculate gradients of the prediction score with respect to the last convolutional layer feature map;

[0040] a weight calculation unit, connected to the gradient calculation unit, configured to perform global average pooling on the gradients of each channel to generate channel importance weights;

[0041] a class activation mapping unit, connected to the weight calculation unit, configured to multiply the channel importance weights with the corresponding feature map and sum to generate a class activation map;

[0042] a heatmap generation unit, connected to the class activation mapping unit, configured to upsample the class activation map to the original image size and apply a heatmap color scheme; and a report generation unit, configured to generate a diagnosis result report containing patient information, risk assessment results, and clinical recommendations.

[0043] The present application has the following beneficial effects:

[0044] 1. Improved diagnostic accuracy: By quantitatively analyzing ultrasound image features, the system effectively reduces errors caused by subjective judgments of doctors. In clinical verification, the accuracy of benign and malignant discrimination of the system is improved by about 25% compared with traditional diagnostic methods.

[0045] 2. Improved diagnostic efficiency: The system can complete image analysis within 30 seconds, significantly reducing the diagnosis time of 3-5 minutes of traditional methods and significantly improving the work efficiency.

[0046] 3. Reduced missed diagnosis rate: The system can identify lesion features, and through the innovative dual-path feature extraction architecture of the feature extraction module, it can simultaneously capture local details and global context information, improving the detection rate of early lymphoma.

[0047] 4. Improve diagnostic consistency: The system provides standardized risk assessment, reducing diagnostic differences between different medical institutions and different doctors, and improving diagnostic consistency by about 40%.

[0048] 5. Reduce medical costs: Provide reliable auxiliary diagnostic tools for primary medical institutions, reduce unnecessary referrals by about 45% and repeated examinations by about 50%, while early diagnosis also reduces the cost of later treatment by about 30%. BRIEF DESCRIPTION OF DRAWINGS

[0049] Figure 1 The overall architecture of the system of the present application is shown in the figure;

[0050] Figure 2 The structure of the ultrasound image acquisition module of the present application is shown in the figure;

[0051] Figure 3 The processing flowchart of the image preprocessing module of the present application is shown in the figure;

[0052] Figure 4 The structure of the feature extraction module of the present application is shown in the figure;

[0053] Figure 5 The detailed structure of the convolution branch of the present application is shown in the figure;

[0054] Figure 6 The detailed structure of the state space branch of the present application is shown in the figure;

[0055] Figure 7 The processing flowchart of the risk assessment module of the present application is shown in the figure;

[0056] Figure 8 The heat map generation diagram of the result display module of the present application is shown in the figure. DETAILED DESCRIPTION

[0057] The present application will be further described in detail below in conjunction with the drawings and specific embodiments. It should be understood that the specific embodiments described herein are only for illustration and explanation of the present application, and are not intended to limit the present application.

[0058] As shown in the figure, Figure 1 The present application provides a pediatric lymphoma ultrasound auxiliary diagnosis system combined with artificial intelligence, which mainly includes an ultrasound image acquisition module 10, an image preprocessing module 20, a feature extraction module 30, a risk assessment module 40 and a result display module 50, and the modules are sequentially connected through a data interface to form a complete diagnosis process.

[0059] The present application aims to develop an ultrasound-assisted diagnosis system for pediatric lymphoma combined with artificial intelligence, which analyzes the features of ultrasound images intelligently to provide objective and quantitative diagnostic references for doctors, and improve the accuracy and consistency of diagnosis. In the prior art, the ultrasound diagnosis of pediatric lymphoma mainly relies on the subjective experience of doctors, which has problems such as non-uniform diagnostic criteria and difficulty in identifying early lesions. Especially in primary medical institutions, due to the shortage of professional pediatric ultrasound doctors, misdiagnosis or missed diagnosis is more likely to occur.

[0060] As shown in Figure 2 The ultrasound image acquisition module 10 includes a high-frequency linear array probe 11, a convex array probe 12 and an image storage unit 13. This module is responsible for obtaining high-resolution ultrasound images of superficial lymph nodes of the patient.

[0061] The high-frequency linear array probe 11 is selected in the frequency range of 7-15 MHz, which is suitable for high-resolution imaging of superficial lymph nodes. In a preferred embodiment of the present application, for children under 5 years old, a frequency of 12-15 MHz is usually used to obtain clearer images of superficial lymph nodes; for children aged 5-12 years old, a frequency of 8-12 MHz is selected to balance the resolution and penetration depth requirements.

[0062] The convex array probe 12 is selected in the frequency range of 5-8 MHz, which is used to detect relatively large and deep lymph nodes, taking into account a certain penetration and field of view. Preferably, for intra-abdominal lymph node examination, a frequency of 5-6 MHz is usually selected; for mediastinal lymph nodes with moderate position, a frequency of 6-8 MHz is selected to achieve an ideal detection effect.

[0063] The specific process of collecting images is as follows: the patient lies on his back on the examination bed, removes his neck clothes, accessories, etc., and fully exposes the neck examination area. The doctor selects the appropriate high-frequency probe, applies coupling agent on the patient's skin, and starts the examination, scans the lymph nodes from multiple angles, and fully observes the shape, size, number, position, structure, etc. of the lymph nodes. According to the approximate depth of the cervical lymph nodes, adjust the image depth, generally set at 2-5 cm, to ensure that the entire lymph node can be completely displayed in the image. In addition, the focal point is set at the depth where the lymph node is expected to be, to improve the resolution of the image.

[0064] The image storage unit 13 stores lymph node images with diagnostic value, including: (1) longitudinal section image of lymph node, showing the maximum longitudinal diameter and transverse diameter; (2) dynamic image from lesion appearance to lesion disappearance; (3) color Doppler flow imaging, the sampling box covers the entire longitudinal section lymph node, and the blood supply at the lymphatic gate is displayed as much as possible. Preferably, the image storage unit 13 stores the original image data in DICOM format, maintains 16-bit depth, and ensures image quality and subsequent processing accuracy.

[0065] AsFigure 3 As shown, the image preprocessing module 20 includes a size normalization unit 21, a noise suppression unit 22, a data enhancement unit 23, and a normalization processing unit 24. This module preprocesses the acquired ultrasound images to improve the consistency of the input data quality.

[0066] The size normalization unit 21 scales the acquired RGB ultrasound images of arbitrary size to 224×224 pixels. In a preferred embodiment of the invention, a bicubic interpolation algorithm is used for image scaling, which is superior to nearest neighbor interpolation and bilinear interpolation in preserving image details and edge sharpness. Specifically, the system first calculates the scaling factor between the original image and the target size, and then applies a weighted average of 16 neighboring pixels to each pixel location to generate the pixel value at the target location.

[0067] Noise suppression unit 22 applies a 5×5 Gaussian blur kernel with variable σ to process the image. The mathematical expression for the Gaussian blur kernel is:

[0068] ,

[0069] in: For the Gaussian kernel in coordinates The weight value at the location; The standard deviation of the Gaussian distribution determines the degree of ambiguity; is the base of the natural logarithm; and These represent the horizontal and vertical distances of the current pixel relative to the kernel center, respectively. In this invention, The value is adaptively adjusted based on image quality, typically ranging from 0.5 to 2.0. Specifically, the system dynamically adjusts the value by evaluating the image's signal-to-noise ratio (SNR). Value: Select when SNR < 10dB Perform strong noise reduction; when 10dB≤SNR<20dB, select When SNR≥20dB, select To preserve more details, data augmentation unit 23 implements various image augmentation strategies to improve the model's adaptability to various variations. Specifically, these include:

[0070] 1. Random Occlusion Enhancement: Three 50x50 pixel rectangular occlusions are randomly added to the image with a probability of 0.5. The location of the occlusion area... satisfy and The location is generated through uniform random sampling. This enhancement simulates the acoustic shadowing regions commonly found in ultrasound examinations, improving the model's robustness to local defects. Specifically: and They represent the first The x and y coordinates of the top left corner of the occluded area; The value ranges from 1 to 3, corresponding to three random occlusion blocks.

[0071] 2. Affine Transformation Enhancement: Includes rotation, translation, scaling, and shearing operations. Rotation angle range is ±15°, translation range is ±10% of the image size, scaling range is 0.95-1.05 times, and shearing angle range is ±10°. The affine transformation matrix can be represented as:

[0072] ,

[0073] in: Parameters for controlling rotation, shearing, and scaling; The parameters control the translation; the third row [0,0,1] in the matrix represents the standard form of the affine transformation, ensuring homogeneous coordinate representation of the transformation. In practical applications, the system randomly generates the transformation matrix based on a preset parameter range and applies it to the input image.

[0074] 3. Brightness / Contrast Enhancement: Brightness and contrast are randomly adjusted within a range of ±20%. Brightness adjustment can be expressed as... Contrast adjustment can be expressed as ,in: These are the original pixel values; These are the adjusted pixel values; This is a brightness adjustment parameter, ranging from ±0.2. The contrast adjustment parameter is also within the range of ±0.2; both parameters are generated through uniform random sampling.

[0075] 4. Vertical Flip: The image is vertically flipped with a probability of 0.5 to simulate different scanning directions. The normalization processing unit 24 performs two-stage normalization on the image. First, the pixel values ​​are mapped from [0, 255] to [0, 1], i.e. Then, the image is converted to tensor format and normalized using the mean (0.5, 0.5, 0.5) and standard deviation (0.5, 0.5, 0.5).

[0076]

[0077] in: These are the pixel values ​​after the first stage of normalization. This is the mean vector of the three RGB channels; This is the standard deviation vector of the three RGB channels; For the final normalized output, the values ​​are restricted to the range [-1, 1]. This normalization process accelerates model convergence and improves training stability.

[0078] As Figure 4 shown, the feature extraction module 30 is the core innovative part of the present application, which adopts a SS-Conv-SSM (State Space-Convolution-State Space Model) dual-path architecture, including a feature splitting unit 33, a convolution branch 31, a state space branch 32, a feature fusion unit 34, a Patch merging unit 35 and a multi-scale feature generation unit 36.

[0079] The feature splitting unit 33 first applies a Patch embedding operation to the input image, mapping the 224x224x3 RGB image to an initial feature map of 56x56xC (where C is usually 64). The Patch embedding operation uses a 4x4 convolution kernel with a stride of 4 for spatial down-sampling and channel up-sampling, which can be represented as:

[0080]

[0081] wherein: is the input image, with a dimension of ; EmbedPatch represents the Patch embedding operation, which is implemented through a convolution kernel; is the initial feature map after embedding, with a dimension of ; represents a real tensor space. This operation does not need to flatten the image, effectively preserving the two-dimensional spatial structure, ensuring that the subsequent modules can extract spatial and sequential features in parallel. Then, the feature splitting unit 33 equally divides the feature map into two parts and in the channel dimension:

[0082]

[0083] wherein: is the input feature map, with a dimension of ; and represent the height and width of the feature map, respectively; SplitChannel represents the channel splitting operation, which divides the input feature map into two parts in the channel dimension; and are the two parts of the feature after splitting, each containing channels; represents a real tensor space. This splitting strategy allows one group of channels to focus on local texture and edge extraction, while the other group of channels is dedicated to global context modeling. As Figure 5As shown, the convolution branch 31 is responsible for extracting local texture and edge features, including a first convolution layer 311, a second convolution layer 312, and a third convolution layer 313, forming a three-level series structure.

[0084] The first convolution layer 311 is responsible for extracting local texture and edge features, including a first convolution layer 311, a second convolution layer 312, and a third convolution layer 313, forming a three-level series structure. Application Convolution and batch normalization, forming an intermediate expression:

[0085] ,

[0086] wherein: is the first part of the feature output by the feature splitting unit; represents the convolution operation; is the 3x3 convolution kernel weight of the first convolution layer; represents the first batch normalization operation; is the output feature map of the first convolution layer. Preferably, the weight of the 3x3 convolution kernel is automatically learned through network training, and the batch normalization parameters include the running mean, the running variance, the scaling factor and the bias term, which are all optimized through the training process.

[0087] The second convolution layer 312 is responsible for extracting deeper spatial features through 3x3 convolution, batch normalization and ReLU activation:

[0088] ,

[0089] wherein: is the output of the first convolution layer; is the 3x3 convolution kernel weight of the second convolution layer; is the second batch normalization operation; ReLU represents the rectified linear unit activation function, defined as ReLU(x)=max(0,x); is the output feature map of the second convolution layer. This layer introduces nonlinearity, enhancing the expression ability of the model. The third convolution layer 313 performs channel fusion and compression on through point-wise convolution (PWConv, i.e. 1x1 convolution), and completes the output with batch normalization and ReLU activation:

[0090] ,

[0091] wherein: is the output of the second convolution layer; is the third batch normalization operation; PWConv represents the point-wise convolution operation (1x1 convolution), which only integrates features in the channel dimension without changing the spatial resolution; ReLU is also a rectified linear unit activation function; is the final output feature map of the convolution branch 31. The weight dimension of the point-wise convolution is ​The input channel number and the output channel number are both , and the convolution kernel size is 1x1. Through the above-mentioned series connection of three-level convolution and normalization, the Conv branch realizes multi-level generalized extraction of various local structures (including texture, edge and micro pattern), and is particularly suitable for capturing key diagnostic features such as the boundary morphology of lymph nodes and internal echo texture.

[0092] As shown in Figure 6 , the state space branch 32 focuses on capturing global dependency information, including a layer normalization unit 321, a two-way semantic extraction unit 322, a multi-directional scanning unit 323 and an information filtering unit 324.

[0093] The layer normalization unit 321 performs layer normalization on to obtain

[0094] ,

[0095] wherein Fs is the second part of the features output by the feature shunting unit; LN represents the layer normalization operation; and Z is the output feature map after layer normalization. Layer normalization is performed on the channel dimension of the feature map to eliminate the statistical difference between channels and make the network training more stable. The calculation formula of layer normalization is as follows:

[0096] ,

[0097] wherein x is the input feature; μ is the mean of the feature map in the channel dimension; σ is the standard deviation of the feature map in the channel dimension; γ is a learnable scaling parameter; β is a learnable bias parameter; and ε is a small constant (usually 1e -5 ) for numerical stability to prevent the denominator from being zero. The two-way semantic extraction unit 322 divides the normalized result Z into two paths for processing. The first path generates semantic features through linear mapping and SiLU activation:

[0098]

[0099] wherein Z is the output of layer normalization; W1 is the linear mapping weight matrix of the first path; SiLU represents the Sigmoid linear unit activation function; and Z1 is the semantic features generated by the first path. The SiLU activation function is defined as SiLU(x)=x·σ(x), wherein σ(x)= is the Sigmoid function. The SiLU activation function has a smooth derivative compared with ReLU, which helps gradient propagation.

[0100] The second path first goes through linear mapping, depth separable convolution and SiLU activation to obtain

[0101] ,

[0102] wherein: is the layer normalized output; is the linear mapping weight matrix of the second path; represents a depthwise separable convolution operation; is also a sigmoid linear activation function; is the semantic feature generated by the second path. The depthwise separable convolution first applies a spatial convolution (depthwise convolution) to each input channel independently, and then uses a pointwise convolution (pointwise convolution) for channel fusion, which greatly reduces the amount of calculation compared with the standard convolution. In this embodiment, the depthwise convolution uses a convolution kernel to enhance the local-channel coupling response. The multi-direction scanning unit 323 is a key innovation point of the present application, which will unfold in four different paths as a sequence to realize efficient capture of global information. Specifically, the left-up to right-down scanning unit 3231 will unfold in a row-first order as a first sequence:

[0103] ,

[0104] the right-down to left-up scanning unit 3232 will unfold in a reverse order as a second sequence:

[0105] ,

[0106] the right-up to left-down scanning unit 3233 will unfold along the diagonal direction as a third sequence:

[0107] ,

[0108] the left-down to right-up scanning unit 3234 will unfold along the anti-diagonal direction as a fourth sequence:

[0109] ,

[0110] wherein: is the second path output of the two-path semantic extraction unit; are the sequences obtained by scanning in four different directions, respectively; is the sequence index, ranging from ; is the total number of pixels of the feature map; and are the height and width of the feature map, respectively; represents the remainder of divided by , corresponding to the column index of the feature map; represents the remainder of divided by​​​ The integer quotient corresponds to the row index of the feature map; the two values ​​within square brackets represent the column and row indices of the feature map, respectively. Each sequence is processed by the S6 module and then restored to a two-dimensional image. And sum the elements to generate a globally aware representation:

[0111] ,

[0112] in: Indicates the first Two-dimensional feature maps of a sequence after processing by the S6 module; The value range is 1 to 4, corresponding to four different scanning directions; The element-wise summation of the feature maps in four directions is the global perception representation. This represents a summation operation, adding corresponding elements from the four feature maps. The S6 module is a variant of the state-space model that efficiently captures long-range dependencies through recursive state updates. Its mathematical expression is:

[0113] ,

[0114] ,

[0115] in: It is the first of the input sequence One element; It is a moment The state vector; It is a moment The state vector; It is a moment The output; , , It is a learnable parameter matrix, where Control state transition, Controlling the impact of input Control output mapping. In this invention, a linear complexity variant implemented using diagonalization is used, which greatly improves processing efficiency. The SS2D (four-way scan) method employs a linear complexity cross-scan strategy to ensure efficient capture of long-range dependencies, with a computational complexity of only [missing information]. Far lower than the self-attention mechanism Its complexity makes it particularly suitable for real-time processing needs. Among other things: and It is Big O notation for computational complexity. The number of pixels in the input feature map. The information filtering unit 324 will process the globally perceived representation. With the first semantic features Perform element-wise multiplication, followed by layer normalization and linear mapping, and output the result:

[0116] ,

[0117] wherein: is the global-aware representation; is the first path output of the dual-path semantic extraction unit; denotes the Hadamard product (element-wise multiplication); denotes the layer normalization operation; is the linear mapping weight matrix; is the final output feature map of the state space branch 32. This operation realizes the reinforced filtering of key semantic information, highlighting the features valuable for diagnosis and suppressing irrelevant information. The feature fusion unit 34 fuses the Conv branch output and the SSM branch output After concatenation in the channel dimension, the channel features are redistributed by the channel shuffling operation Shuffle to obtain the fused feature with the same number of channels :

[0118] ,

[0119] wherein: is the output feature map of the convolution branch 31; is the output feature map of the state space branch 32; ShuffleConcat denotes the shuffling operation after channel concatenation; is the fused feature map, with the dimension ; denotes a real tensor space of dimensions. The implementation of channel shuffling is to first reshape the feature map to the shape of ( is the number of groups, usually set to 2), then transpose the middle two dimensions to obtain , and finally reshape back to . This process breaks the channel isolation between branches and maximizes the complementarity of local and global information. The patch merging unit 35 performs spatial merging on using a window and doubles the number of channels to , with the output dimension being:

[0120] ,

[0121] wherein: is the output of the feature fusion unit; PatchMerging denotes the patch merging operation; is the down-sampled feature map, with the dimension ; denotes a A real tensor space of dimension 1. The specific implementation involves dividing adjacent... The region's features are flattened and the number of channels is reduced from [previous value] using a linear mapping. Increase to Simultaneously, the spatial resolution is reduced by half. This operation preserves information about adjacent pixel regions while downsampling, giving the feature map a wider receptive field and allowing for further feature extraction in subsequent stages. The multi-scale feature generation unit 36 ​​repeatedly stacks and merges four levels of SS-Conv-SSM with Patch, and the system sequentially outputs multi-scale feature maps:

[0122]

[0123] in: These represent feature maps at four different scales; These represent four different dimensions of real tensor space. This structure forms a top-down pyramid-like representation, fully integrating local texture and global semantic features, providing stable and rich multi-scale feature support for subsequent risk assessment modules. In practical applications, this multi-scale feature design is particularly suitable for the analysis of lymph nodes of different sizes: smaller lymph nodes (diameter <5mm) mainly utilize high-resolution features. and Medium-sized lymph nodes (5-10 mm) are mainly utilized and Larger lymph nodes (>10mm) should be fully utilized. and A vast and expansive field of experience.

[0124] like Figure 7 As shown, the risk assessment module 40 includes a dimension rearrangement unit 41, a global pooling unit 42, a feature importance weighting unit 43, and a probability output unit 44. This module converts the output of the feature extraction module into the probability of lymphoma malignancy, providing a quantitative risk assessment.

[0125] Dimension rearrangement unit 41 generates the feature map output by the feature extraction module. Dimensional replacement ,in: Input feature map; and These are the height and width of the feature map, respectively. It is the number of channels; This indicates a dimension replacement operation; and These represent the real tensor spaces before and after the permutation, respectively. This operation prepares for subsequent global processing on each channel. Global pooling unit 42 performs global pooling on each channel. Apply global average pooling:

[0126] ,

[0127] wherein: is the global average pooling result of the channel ; and are the height and width of the feature map, respectively; denotes the value at the channel , position in the feature map; denotes the double summation over the spatial dimensions of the feature map; is a normalization factor to ensure the result is an average value. Through this operation, a feature vector that aggregates global information is obtained wherein: is the feature vector; is the global average pooling result of each channel; the superscript denotes the transpose of the vector; denotes a -dimensional real vector space. Since this operation compresses the spatial dimension information into channel statistics, it not only significantly reduces the subsequent computational load, but also ensures the robustness of the model to local translation and deformation. In practical applications, this feature is particularly important for ultrasound images, because the slight movement of the child during the ultrasound examination often leads to slight changes in the position of the lymph nodes. The feature importance weighting unit 43 directly maps the feature vector to the risk score (logit) through a linear mapping:

[0128] ,

[0129] wherein: is the feature vector obtained by global pooling; is the weight vector, which has the same dimension as the feature vector; the superscript denotes the transpose of the vector; is a scalar bias term; is the output risk score, which is a scalar value. This design allows the model to finely adjust the importance of each channel feature through learnable weights while maintaining a simple structure. During the model training process, the system automatically learns the weight vector so that it can accurately identify features related to lymphoma risk. For example, feature channels related to lymphatic gate structure, boundary clarity, and internal echo usually get higher weight values.

[0130] The probability output unit 44 converts the risk score into the probability of lymphoma malignancy risk using the Sigmoid function:

[0131] ,

[0132] wherein: the risk score output by the feature importance weighting unit; the base of the natural logarithm; denotes the exponential function; is the converted lymphoma malignancy risk probability, which ranges from 0 to 1. This probability output not only corresponds directly to the doctor's experience, but also facilitates the setting of different thresholds to balance sensitivity and specificity. In clinical applications, the risk threshold is determined based on ROC curve analysis: generally, a risk probability <0.3 is defined as low risk, 0.3-0.7 as medium risk, and >0.7 as high risk. These thresholds can be adjusted according to the actual needs of different medical institutions to balance the risk of missed diagnosis and over-diagnosis.

[0133] As shown in Figure 8 , the result display module 50 includes a gradient calculation unit 51, a weight calculation unit 52, a class activation mapping unit 53, a heat map generation unit 54, and a report generation unit 55. This module generates intuitive visual results and diagnostic reports, enhancing the explainability of the system.

[0134] The gradient calculation unit 51 implements the first step of the Grad-CAM algorithm, calculating the gradient of the prediction score with respect to the last convolutional layer feature map. Specifically, first, the prediction score of the model for the lymphoma positive class is obtained Then, the gradient of this score with respect to the last convolutional layer feature map is calculated where denotes the prediction score of class c (lymphoma positive); denotes the kth channel feature map of the last convolutional layer; denotes the partial derivative of , i.e., the gradient. This step is implemented through backpropagation and can determine which feature regions contribute most to the prediction result. The weight calculation unit 52 globally averages the gradient of each channel to generate the channel importance weight:

[0135] ,

[0136] where: denotes the importance weight of channel k for class c; denotes the gradient of the prediction score at position (i,j) of the feature map ; denotes the sum over all positions of the feature map; is the normalization factor, equal to the number of pixels of the feature map, i.e. where and height and width of the feature map respectively. In this invention, c denotes the positive class of lymphoma. The class activation mapping unit 53 multiplies the channel importance weight with the corresponding feature map and sums them up to generate the class activation map (CAM):

[0137] ,

[0138] where: is the class activation map of class c; is the importance weight of channel k for class c; is the feature map of the k-th channel; denotes the weighted sum over all channels; the ReLU function is defined as ReLU(x) = max(0, x) to filter out negative values and only keep the positive contributions. This step generates a heat map that visually shows the image regions that the model paid attention to when making the prediction.

[0139] The heat map generation unit 54 up-samples the class activation map to the original image size to generate the heat map:

[0140] ,

[0141] where: is the class activation map; Upsample denotes the up-sampling operation, which usually adopts the bilinear interpolation method; is the up-sampled class activation map, which has the same size as the original ultrasound image (e.g., 224 , a gradient color spectrum from blue (low response) to red (high response) is usually used, and finally the heat map is semi-transparently superimposed on the original ultrasound image to form an intuitive visualization result.

[0142] In practical applications, this heat map visualization is of great value in guiding doctors in diagnosis. For example, if the heat map highlights the irregular boundary area of the lymph node or the area where the lymphatic gate disappears, it usually suggests a high possibility of malignancy; while if the heat map mainly focuses on the lymphatic gate area and the lymphatic gate structure is well preserved, it suggests a high possibility of benignity.

[0143] The report generation unit 55 integrates all the analysis results to generate a comprehensive diagnostic report. The report contains the following main contents:

[0144] 1. Basic information: patient's name, age, gender, examination site, examination date, etc.

[0145] 2. Lymph node measurement data: long diameter, short diameter, long-to-short diameter ratio. The long-to-short diameter ratio is an important indicator, and benign lymph nodes usually have a long-to-short diameter ratio > 2, while malignant lymph nodes often have a long-to-short diameter ratio < 2. The system will automatically calculate this ratio and highlight it in the report.

[0146] 3. Lymph node morphology description: Based on image features, the system automatically generates a morphology description, including shape (ovate, round, irregular, etc.), boundary (clear, blurred), internal echo (uniform, non-uniform), lymphatic gate (present, absent), etc.

[0147] 4. Risk assessment results: Lymphoma malignancy risk probability value, 95% confidence interval, risk level (low / medium / high).

[0148] 5. Visualization results: Original ultrasound images, heat map overlay display, key region labeling and analysis.

[0149] 6. Clinical recommendations: Automatically generate corresponding clinical recommendations based on risk level. For example, low risk (<0.3) is recommended to review after 3-6 months; medium risk (0.3-0.7) is recommended to perform further examination such as elastic ultrasound, ultrasound contrast or PET-CT; high risk (>0.7) is recommended to perform pathological biopsy as soon as possible to confirm the diagnosis.

[0150] Preferably, the report adopts a standardized format, supports multiple export formats such as PDF, HTML, DICOM SR (structured report), and provides interface integration with hospital HIS / RIS / PACS systems.

[0151] The complete workflow of the system is as follows:

[0152] 1. Ultrasound image acquisition: The doctor uses a high-frequency linear array probe or a convex array probe to collect ultrasound images of the child's lymph nodes.

[0153] 2. Image preprocessing: The system performs size unification, noise suppression, data enhancement and normalization processing on the collected ultrasound images.

[0154] 3. Feature extraction: The system captures local texture features and global context features simultaneously through the SS-Conv-SSM dual-channel feature extraction architecture, generating multi-scale feature maps.

[0155] 4. Risk assessment: The system performs global average pooling and feature importance weighting on the multi-scale feature maps to calculate the lymphoma malignancy risk probability.

[0156] 5. Results display: The system generates heat map visualization and diagnostic result report to assist the doctor in making diagnosis decisions.

[0157] In a multi-center clinical validation study, the system showed excellent performance on a test set containing 2500 ultrasound images: for benign and malignant discrimination, the system achieved a sensitivity of 92.7% and a specificity of 90.3%, with an accuracy of 91.5%; for early lesions (diameter <1cm), the system's detection rate was improved by 32.5% compared to traditional methods; and the diagnostic consistency between different medical institutions and different doctors was improved by 42.3%.

[0158] In addition, the average diagnosis time of the system is only 28 seconds, which is significantly shorter than the 3-5 minutes of traditional manual interpretation, significantly improving work efficiency. In the application of primary medical institutions, the system reduces unnecessary referrals by 47.2% and repeated examinations by 52.1%, significantly reducing medical costs.

[0159] In summary, the present application provides an ultrasound auxiliary diagnosis system for pediatric lymphoma combined with artificial intelligence, which realizes the effective fusion of local details and global context information through the innovative SS-Conv-SSM dual-path feature extraction architecture and multi-directional scanning strategy, provides an objective and quantitative auxiliary tool for ultrasound diagnosis of pediatric lymphoma, effectively improves the diagnosis accuracy, efficiency and consistency, reduces the misdiagnosis rate and medical costs, and has a wide clinical application prospect.

[0160] The above description is only the preferred embodiment of the present application and is not intended to limit the present application. For those skilled in the art, the present application can have various modifications and changes. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. An ultrasonic auxiliary diagnosis system for pediatric lymphoma combined with artificial intelligence, characterized in that, The application relates to a lymphoma risk assessment method based on multi-scale feature fusion, comprising the following steps: An ultrasonic image acquisition module is used to acquire an ultrasonic image of a superficial lymph node of a child patient; an image preprocessing module is connected with the ultrasonic image acquisition module and is used to perform standardized processing on the ultrasonic image to generate a pretreated image; A feature extraction module is connected with the image preprocessing module and is used to process the pretreated image, wherein the feature extraction module comprises a convolution branch and a state space branch, the convolution branch is used to extract local texture features, the state space branch is used to extract global context features, and the feature extraction module generates a multi-scale feature map by fusing the local texture features and the global context features; A risk assessment module is connected with the feature extraction module and is used to receive the multi-scale feature map, perform global average pooling and feature importance weighting on the multi-scale feature map, and generate a lymphoma malignancy risk probability; And a result display module is connected with the risk assessment module and is used to receive the lymphoma malignancy risk probability, generate a diagnosis result report and a heat map visual display; The feature extraction module further comprises a feature shunt unit used to divide an input feature map into a first part and a second part in a channel dimension, the first part is input into the convolution branch, and the second part is input into the state space branch; and a feature fusion unit connected with the convolution branch and the state space branch and used to perform channel dimension splicing and channel shuffling operation on output features of the convolution branch and output features of the state space branch to generate fused features; The convolution branch comprises a first convolution layer used to apply a first convolution operation and batch normalization to the first part of the feature map; a second convolution layer connected with the first convolution layer and used to apply a second convolution operation, batch normalization and ReLU activation to output of the first convolution layer; and a third convolution layer connected with the second convolution layer and used to apply a point-wise convolution operation, batch normalization and ReLU activation to output of the second convolution layer to generate output features of the convolution branch; layer normalization unit pair performing layer normalization to obtain: , The state space branch comprises a layer normalization unit used to perform layer normalization processing on the second part of the feature map; a double-path semantic extraction unit connected with the layer normalization unit and used to generate first path semantic features and second path semantic features; a multi-direction scanning unit connected with the double-path semantic extraction unit and used to perform multi-direction scanning processing on the second path semantic features to generate global perception representation; and an information filtering unit connected with the double-path semantic extraction unit and the multi-direction scanning unit and used to interact the global perception representation with the first path semantic features to generate output features of the state space branch; , where: x is the input feature; μ is the mean of the feature map in the channel dimension; σ is the standard deviation of the feature map in the channel dimension; γ is a learnable scaling parameter; β is a learnable bias parameter; ε is a small constant, 1e -5 for numerical stability, to prevent the denominator from being zero; Wherein: Fs is the second part of the feature output by the feature shunt unit; LN represents a layer normalization operation; Z is an output feature map after layer normalization; and a calculation formula of the layer normalization is as follows: , Wherein: Z is the layer normalized output; W1 is the linear mapping weight matrix of the first path; SiLU represents the Sigmoid linear unit activation function; The semantic feature generated for the first path; the SiLU activation function is defined as SiLU(x)=x·σ(x), wherein σ(x)= is the Sigmoid function; the SiLU activation function has a smooth derivative compared with ReLU, which helps gradient propagation; The double-path semantic extraction unit divides the normalization result Z into two paths for processing, the first path generates semantic features through linear mapping and SiLU activation: The second path first obtains through linear mapping, depth separable convolution and SiLU activation: , wherein: is the layer normalized output; is the linear mapping weight matrix of the second path; denotes a depthwise separable convolution operation; is also a sigmoid linear activation function; is the generated semantic feature of the second path; The multi-direction scanning unit comprises: an upper-left-to-lower-right scanning unit for unfolding the second road semantic feature into a first sequence along an upper-left-to-lower-right direction; a lower-right-to-upper-left scanning unit for unfolding the second road semantic feature into a second sequence along a lower-right-to-upper-left direction; an upper-right-to-lower-left scanning unit for unfolding the second road semantic feature into a third sequence along an upper-right-to-lower-left direction; and a lower-left-to-upper-right scanning unit for unfolding the second road semantic feature into a fourth sequence along a lower-left-to-upper-right direction; wherein the global perception representation is generated by merging results processed by the first sequence, the second sequence, the third sequence and the fourth sequence.

2. The system of claim 1, wherein, The ultrasound image acquisition module comprises: a high-frequency linear array probe for acquiring high-resolution ultrasound images of superficial lymph nodes; a convex array probe for acquiring ultrasound images of lymph nodes at a relatively deep position; and an image storage unit connected with the high-frequency linear array probe and the convex array probe, for storing static images, dynamic images and color Doppler flow imaging images of lymph nodes.

3. The system of claim 1, wherein, The image preprocessing module comprises: a size unification unit for scaling the ultrasound images to a predetermined size; a noise suppression unit connected with the size unification unit for performing Gaussian blur processing on the ultrasound images of the predetermined size; a data enhancement unit connected with the noise suppression unit for performing random occlusion, geometric transformation and optical property enhancement on the ultrasound images; and a normalization processing unit connected with the data enhancement unit for performing pixel value mapping and statistical normalization processing on the enhanced ultrasound images.

4. The system of claim 1, wherein, The feature extraction module further comprises: a Patch merging unit connected with the feature fusion unit for spatially merging and doubling the number of channels of the fused features to generate down-sampled features; and a multi-scale feature generation unit connected with the Patch merging unit for generating the multi-scale feature maps of different spatial resolutions through multi-level feature extraction and down-sampling operations.

5. The system of claim 1, wherein, The risk assessment module comprises: a dimension rearrangement unit for rearranging the dimensions of the multi-scale feature maps from height, width and number of channels to number of channels, height and width; a global pooling unit connected with the dimension rearrangement unit for applying global average pooling to each channel of the rearranged feature maps to generate a feature vector; a feature importance weighting unit connected with the global pooling unit for performing linear mapping on the feature vector to calculate a comprehensive risk score; and a probability output unit connected with the feature importance weighting unit for converting the comprehensive risk score into the lymphoma malignancy risk probability through a Sigmoid function.

6. The system of claim 1, wherein, The result display module comprises: a gradient calculation unit configured to calculate gradients of the prediction score with respect to the last convolutional layer feature map; a weight calculation unit connected to the gradient calculation unit, configured to perform global average pooling on the gradients of each channel to generate channel importance weights; a class activation mapping unit connected to the weight calculation unit, configured to multiply the channel importance weights with corresponding feature maps and sum them up to generate a class activation map; a heat map generation unit connected to the class activation mapping unit, configured to upsample the class activation map to the original image size and apply a heat map color scheme; and a report generation unit configured to generate a diagnosis result report containing patient information, risk assessment results and clinical recommendations.

Images