Automatic classification method of children lymphoma types based on imageomics

Abstract

The application relates to a kind of automatic classification methods of children lymphoma type based on imageomics, belong to artificial intelligence technical field.It includes the following steps: collecting children lymphoma CT image, constructs data set;The original three-dimensional CT image is preprocessed and adaptively gray window is adjusted, and preprocessed three-dimensional CT image data is obtained;Imageomics feature extraction and region perception gradient weighted screening are carried out, and the comprehensive screening score of each feature is obtained;A classification model of children lymphoma type is constructed, the enhanced imageomics feature, the clinical prior feature and the deep feature modulated by region perception weight are fused, the adaptive fusion of features is realized by using dynamic gating mechanism, and the category probability vector is output;The model is trained and optimized by a composite loss function, and a trained model is obtained;The CT image of children lymphoma to be classified is input into the trained model, and a classification result is obtained.The application can improve the accuracy of automatic classification of children lymphoma type.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology, specifically relating to an automatic classification method for childhood lymphoma types based on radiomics. Background Technology

[0002] Childhood lymphoma is one of the most common malignant tumors in childhood, and accurate subtyping is crucial for developing individualized treatment plans and improving prognosis. Currently, pathological subtyping of lymphoma mainly relies on tissue biopsy, which is an invasive procedure with limitations in sample collection and a long diagnostic cycle. In terms of imaging diagnosis, CT is a routine examination for evaluating lymphoma, but traditional imaging diagnosis heavily relies on the experience of radiologists for subjective visual assessment. Different subtypes of lymphoma show overlapping appearances on CT images, making differential diagnosis challenging, especially for inexperienced physicians, who are prone to misdiagnosis. In recent years, radiomics and deep learning technologies have provided new avenues for the quantitative analysis and automated classification of tumors.

[0003] Existing technologies suffer from the following drawbacks: conventional fixed-window or global normalization preprocessing methods cannot adapt to intensity differences between different devices and patients, especially for low-contrast tissues in children, and are prone to losing key diagnostic information; conventional radiomics methods heavily rely on precise 3D segmentation of tumor regions, a process that is time-consuming and prone to introducing errors, while the extracted high-dimensional features have high redundancy, which can easily lead to model overfitting; conventional deep learning methods are mostly end-to-end black-box models, lacking explicit utilization of medical prior knowledge (such as typical density, location, and age), and may focus on irrelevant image features, affecting interpretability and generalization performance; conventional training strategies only optimize classification accuracy without considering the model's confidence in its own predictions, making it difficult to handle samples with ambiguous categories, and unable to provide uncertainty references during prediction, which is not conducive to clinical decision support. Summary of the Invention

[0004] To achieve the above objectives, the present invention employs the following technical solution: This invention provides an automatic classification method for childhood lymphoma types based on radiomics, comprising the following steps: S1. Collect CT images of childhood lymphoma and construct a dataset of childhood lymphoma CT images; S2. Preprocess and adaptive grayscale window adjustment are performed on the original 3D CT image to obtain preprocessed 3D CT image data; S3. Perform radiomics feature extraction and gradient-weighted filtering for region perception on the preprocessed 3D CT image data to obtain a comprehensive filtering score for each feature. S4. Construct a classification model for childhood lymphoma types by fusing enhanced radiomics features, clinical prior features, and deep features modulated by region-aware weights, and using a dynamic gating mechanism to achieve adaptive feature fusion and output a category probability vector. S5. During the model training process, the model is trained and optimized using a composite loss function to obtain a well-trained classification model for childhood lymphoma types. S6. Input the CT images of the pediatric lymphoma to be classified into the trained pediatric lymphoma classification model to obtain the classification results.

[0005] Furthermore, Z-fractional normalization is applied to the original 3D CT images to eliminate intensity scale differences caused by the scanning protocol. An adaptive gray-scale window adjustment method driven by image statistics and medical prior knowledge is used to dynamically optimize image contrast and focus visual attention on the density range most likely to contain tumor tissue. Raw 3D CT image data Z-score standardization is performed, using the voxel gray mean and standard deviation of the entire training set to eliminate intensity scale differences caused by different scanning protocols, resulting in standardized 3D CT image data. For standardized 3D CT image data Based on the statistical characteristics of its gray-level histogram and combined with the prior typical CT values ​​of childhood lymphoma, the optimal window level and width of the adaptive window are dynamically calculated. The adaptive window is then used to perform linear contrast stretching and truncation on the image, mapping the gray-level values ​​of key density intervals to the range of 0 to 1, thus obtaining preprocessed 3D CT image data. .

[0006] Furthermore, three-dimensional radiomics features are extracted from the preprocessed three-dimensional CT image data to form an initial radiomics feature vector. The three-dimensional gradient magnitude map of the preprocessed three-dimensional CT image data is calculated based on the gradient magnitude region-aware weighting mechanism to construct the region-aware weight matrix. In order to reduce feature redundancy and select the most discriminative feature subset, a dual-index screening mechanism is adopted to comprehensively evaluate the stability of each feature in the same type of sample and its discriminative power among different types of samples, and calculate the comprehensive screening score of each feature in the initial radiomics feature vector.

[0007] Furthermore, three-dimensional radiomics features are extracted from the preprocessed three-dimensional CT image data to form an initial radiomics feature vector. This includes: selecting basic and computationally efficient 3D radiomics feature categories, including first-order statistical features (such as mean, standard deviation, skewness, kurtosis, energy, entropy, etc.) and texture features (such as gray-level co-occurrence matrix, gray-level run matrix, gray-level size region matrix, neighborhood gray-level difference matrix); and using 3D radiomics feature extraction tools to process preprocessed 3D CT image data. Full-image feature calculation is performed directly on the entire 3D image since precise segmentation is not required. However, pre-setting discretization parameters (such as quantizing grayscale values ​​to a fixed number of bins, e.g., 64 bins) is necessary to calculate texture features. The extraction process is based on the image's 3D voxel grid, calculating all available features for each feature category to generate an initial imageomics feature vector. .

[0008] Furthermore, clinical prior features are introduced, and these features are fused and nonlinearly enhanced with the screened radiomics features to obtain an enhanced feature vector. .

[0009] Furthermore, a 3D convolutional neural network backbone is used to extract depth feature maps from preprocessed 3D CT image data. A spatial attention guidance module is employed, using a region-aware weight matrix as a prior attention template to spatially recalibrate the depth feature maps, making the network focus more on potential lesion areas, thus obtaining attention feature maps. .

[0010] Furthermore, the enhanced feature vector is fused with the global deep feature vector extracted from the deep feature map. The contribution ratio of the two types of features in the final decision is adaptively determined through a dynamic gating mechanism to achieve information complementarity and generate a joint feature vector.

[0011] Furthermore, the joint feature vector is input into the classifier to obtain the original scores for each category. A learnable data-dependent temperature parameter is introduced to scale the scores to quantify the cognitive uncertainty of the model's predictions, and a flexible positive function is used to generate the classification probability.

[0012] Furthermore, the composite loss function includes batch classification loss and region-aware feature consistency regularization loss.

[0013] Furthermore, the batch classification loss uses focused cross-entropy loss as the basic classification loss to address the potential class imbalance problem in the data. The loss of each sample is dynamically weighted using a temperature parameter. When the model's prediction uncertainty for a certain sample is high, the contribution of that sample to the total batch loss is automatically reduced, avoiding the model forcibly fitting samples that are difficult to judge.

[0014] Furthermore, the region-aware feature consistency regularization loss constructs a consistency regularization loss using a region-aware weight matrix, encouraging the local feature vector of each sample's deep feature map in the potential lesion region to be as close as possible to the average feature vector of all samples of the same class in that region, thereby enhancing the intra-class compactness and semantic consistency of feature representation.

[0015] The advantages of this invention are: This invention employs an adaptive gray-scale window adjustment method that combines global image statistics with prior CT values ​​for lymphoma. This method dynamically optimizes image contrast for different patients and scanning protocols, effectively enhancing subtle density differences in key areas. Based on a radiomics feature extraction strategy that does not require precise segmentation, it utilizes gradient magnitude to construct region-aware weights to focus on potential lesions. Features are screened using both stability and discriminative power indicators, reducing computational redundancy and dependence on segmentation accuracy. A multimodal fusion and attention-guided classification model is adopted, innovatively fusing screened radiomics features, clinical priors, and deep features. Gradient priors are used as spatial attention templates, aligning the model's learning process with medical knowledge. A data-dependent temperature parameter is introduced into the model to quantify cognitive uncertainty. Based on this, an uncertainty-weighted focusing loss function and feature consistency regularization loss are designed, improving the model's robustness to difficult samples and the intra-class compactness of feature representations. Attached Figure Description

[0016] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0017] Figure 1 This is a flowchart of the method of the present invention; Figure 2 The original 3D CT image; Figure 3 A binarized image of a 3D segmentation mask; Figure 4 To standardize 3D CT image data; Figure 5 High-density region of interest; Figure 6 The effects of different preprocessing methods on improving image contrast; Figure 7 The performance of different pretreatment methods on three lymphoma subtypes; Figure 8 This is a three-dimensional gradient magnitude plot; Figure 9 A 3D visualization of gradient magnitude; Figure 10 This is the composite loss function of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Example 1 In this embodiment, as Figure 1 As shown, this invention provides an automatic classification method for childhood lymphoma types based on radiomics, the specific steps of which include: S1. Construction of a CT image dataset for childhood lymphoma The pediatric lymphoma CT image dataset covers common clinical subtypes of pediatric lymphoma. Enhanced CT scans of the chest, abdomen, or pelvis of pediatric patients (typically aged 0-18 years) diagnosed with lymphoma by pathological biopsy were selected from hospital image archiving and communication systems and pathology information systems. Acquisition ensured that each patient included at least one CT scan sequence with clear tumor visualization (e.g., venous phase) performed before treatment. Simultaneously, corresponding patient age information and pathology reports were collected. The pathology report serves as the gold standard for determining the specific pathological type of lymphoma (e.g., Burkitt lymphoma, lymphoblastic lymphoma, diffuse large B-cell lymphoma).

[0020] Data annotation is performed manually. Precise 3D regions of interest (ROIs) of the tumor lesions are manually delineated on 3D CT images, generating a 3D segmentation mask. This mask defines the voxel extent of the tumor. Furthermore, based on the pathology report, each sample is assigned a specific pathological type label, which serves as the target category for automatic classification in this method. Additionally, the centroid coordinates in 3D space can be calculated from the delineated tumor regions, providing a rough representation of the tumor's macroscopic location.

[0021] The final dataset was divided into three independent parts: a training set, a validation set, and a test set. The training set was used to learn the model parameters; the validation set was used to monitor model performance and adjust hyperparameters during training to prevent overfitting; and the test set was used to evaluate the model's generalization and classification capabilities.

[0022] All collected data has been anonymized to protect patient privacy and reviewed by the medical ethics committee before use.

[0023] In one embodiment, the original 3D CT image is as follows: Figure 2 As shown, the binarized image of the 3D segmentation mask is as follows: Figure 3 As shown, the binarized image of the 3D segmentation mask is an axial 2D slice in the 3D data, where the tumor area is white and the non-tumor area is black, and the tumor type is lymphoblastic lymphoma.

[0024] S2, CT Image Preprocessing and Adaptive Gray Window Adjustment CT images of childhood lymphoma have a three-dimensional spatial structure and a wide range of original grayscale values. Furthermore, images acquired under different scanning devices and protocols exhibit differences in intensity scale. Additionally, the anatomical structures of pediatric patients are relatively small, resulting in potentially low inter-tissue contrast. The density difference between the tumor area and the surrounding normal tissue may not be significant. Conventional techniques, which use fixed window widths and levels for linear mapping or normalize the grayscale of the entire volume data to a fixed range, are prone to failing to adapt to the tumor density characteristics of different cases and the tissue features of children, and may lose subtle density differences that are crucial for classification.

[0025] This invention first performs Z-score normalization on the original 3D CT image to eliminate intensity scale differences caused by the scanning protocol. Then, it employs an adaptive grayscale window adjustment method driven by image statistics and medical prior knowledge to dynamically optimize image contrast and focus the visual attention on the density range most likely to contain tumor tissue. The specific steps are as follows: 1) Z-score standardization based on global statistics Raw 3D CT image data Z-score standardization is performed, using the voxel gray mean and standard deviation of the entire training set to eliminate intensity scale differences caused by different scanning protocols, resulting in standardized 3D CT image data. , specifically definition This represents the original 3D CT image data, with dimensions of [dimension number missing]. The values ​​are directly acquired by CT scanning equipment and include the original Henle unit values. Represents standardized 3D CT image data, with dimensions of Alignment of grayscale intensity scales between different samples is achieved; It represents depth, that is, the number of voxels in the depth direction of a 3D CT image; Indicates height, i.e., the number of voxels in the height direction of a 3D CT image; The width refers to the number of voxels in the width direction of a 3D CT image.

[0026] In one embodiment, standardized 3D CT image data such as Figure 4 As shown.

[0027] 2) Combining prior adaptive gray-scale window calculation and mapping For standardized 3D CT image data, based on its gray-level histogram statistical characteristics and combined with the prior typical CT values ​​of childhood lymphoma, the optimal window level and width of the adaptive window are dynamically calculated. Then, the adaptive window is used to perform linear contrast stretching and truncation on the image, mapping the gray-level values ​​of key density intervals to the range of 0 to 1, as shown below: , In the formula, This indicates that the preprocessed 3D CT image data has its voxel gray values ​​normalized to the 0 to 1 range through a linear mapping of an adaptive window, thereby enhancing the contrast of key areas. This indicates the adaptive window level, which determines the center value of the grayscale display window. The calculation method is expressed as follows: ; This indicates the adaptive window width, which determines the size of the grayscale display window range. The calculation method is expressed as follows: ; This represents the average gray value of a high-density region of interest in an image. In one embodiment, a high-density region of interest is, for example... Figure 5 As shown, through standardized 3D CT image data The statistical characteristics of potential lesion regions in the current image are obtained by applying fast threshold pre-segmentation. This represents the median CT value typical of childhood lymphoma, serving as a reference anchor introduced from prior medical knowledge. An example value is 40 Henlein units. This represents the image statistical weighting coefficient, used to adjust the contribution ratio of the current image estimate in the window level calculation; an example value is 0.7. This represents the prior knowledge weighting coefficient, used to adjust the contribution ratio of medical prior values ​​in window calculation; an example value is 0.3. Represents standardized 3D CT image data The 95th percentile of all voxel gray values ​​is used to estimate the boundaries of high-density regions in the image. Represents standardized 3D CT image data The 5th percentile of all voxel gray values ​​is used to estimate the boundaries of low-density regions in the image. This represents the window width scaling factor, used to fine-tune the base window width calculated based on percentile differences to optimize contrast. An example value is 1.2. This represents the truncation function, which will... Limited to the range Inside, it is used to saturate the grayscale values ​​outside the window to the window boundary.

[0028] In one embodiment, such as Figure 6 As shown, the superiority of the adaptive gray-scale window adjustment method of this invention is verified by comparing the improvement effects of different preprocessing methods on image contrast. The experiment compares four preprocessing methods: the fixed window width and window level method is a routine clinical technique that uses a fixed window width and window level for linear mapping; the global normalization method normalizes the gray-scale of the entire data to a fixed range; the histogram equalization method enhances contrast by adjusting the image histogram; and the method of this invention is an adaptive gray-scale window adjustment method driven by image statistics and prior medical knowledge.

[0029] The experiment used the same dataset of pediatric lymphoma CT images, applied four different preprocessing methods, and then calculated the contrast improvement index.

[0030] The box plot illustrates the distribution of the contrast improvement index for each method. The boxes represent the interquartile range of the data, the median represents the median, the dashed line represents the data range, and outliers are indicated by circles. The plot shows that the method of this invention has the highest box position and a narrowest distribution range, indicating that its improvement effect is stable and significantly superior to other methods.

[0031] like Figure 7 As shown in the bar chart, the performance of different preprocessing methods on three lymphoma subtypes is illustrated. The method of this invention achieved the highest contrast improvement index in Burkitt lymphoma, lymphoblastic lymphoma, and diffuse large B-cell lymphoma. The contrast improvement index on the coordinate axis is a dimensionless indicator; a higher value indicates a better contrast improvement effect. Experimental results show that the method of this invention can adaptively adjust the window width and window level according to the tumor density characteristics of different cases and the tissue characteristics of children, effectively enhancing the contrast of key areas and laying a good foundation for subsequent feature extraction and classification.

[0032] S3. Radiomics Feature Extraction and Region-Aware Gradient-Weighted Filtering Preprocessed CT images contain potential distinguishing features of lymphoma lesions in terms of morphology and texture. Conventional radiomics techniques usually rely on precise segmentation of the tumor region and then extraction of a large number of features. This process is not only limited by segmentation accuracy, but the extracted features also have high dimensionality and redundancy. Direct fusion with deep features can easily lead to the curse of dimensionality and model overfitting.

[0033] This invention employs a radiomics feature extraction and screening strategy that does not require precise segmentation. This strategy first extracts basic features from the global image, then uses image gradient information to construct region-aware weights, guiding feature calculation to focus on potential lesion regions. Finally, a dual-index screening mechanism is used to obtain a concise and discriminative feature subset. The specific steps are as follows: 1) Preliminary extraction of global radiomics features From preprocessed 3D CT image data Extract a set of basic, computationally efficient 3D image omics features to form the initial image omics feature vector. , specifically definition This represents the initial radiomics feature vector, with dimension 1. It includes first-order statistical features and texture features calculated from the global image; This represents the total number of initial radiomics features.

[0034] In one implementation, the specific steps for preliminary extraction of global radiomics features are as follows: a) Select basic and computationally efficient 3D image omics feature categories, mainly including first-order statistical features (such as mean, standard deviation, skewness, kurtosis, energy, entropy, etc.) and texture features (such as gray-level co-occurrence matrix, gray-level run matrix, gray-level size region matrix, neighborhood gray-level difference matrix). These features are calculated in 3D space and can capture the morphological and texture information of lesions. b) Use 3D radiomics feature extraction tools (such as the PyRadiomics library) to process preprocessed 3D CT image data. Full-image feature calculation is performed directly on the entire 3D image since precise segmentation is not required. However, pre-setting discretization parameters (such as quantizing grayscale values ​​to a fixed number of bins, e.g., 64 bins) is necessary to calculate texture features. The extraction process is based on the image's 3D voxel grid, calculating all available features for each feature category to generate an initial imageomics feature vector. .

[0035] 2) Region-aware weighted feature extraction based on gradient magnitude The gradient magnitude-based region-aware weighting mechanism first calculates the 3D gradient magnitude map of the preprocessed 3D CT image data, and then constructs a region-aware weight matrix based on this map, as follows: , In the formula, This represents the region-aware weight matrix, with size . Each element has a value between 0 and 1, and the sum of all elements is 1. This is used to assign higher weights to high-gradient regions in feature computation. express In The value of the element at that position; This represents a three-dimensional gradient magnitude plot with dimensions of [size missing]. It characterizes the degree of drastic grayscale change at each voxel location in the image. express In The value of the element at position is calculated as follows: ; Indicates the voxel index in the depth direction; Indicates the voxel index in the height direction; Indicates the voxel index in the width direction; Represents preprocessed 3D CT image data The partial derivatives in the depth direction are calculated using the Sobel operator isogradient filter. Represents preprocessed 3D CT image data The partial derivative in the height direction is calculated using the Sobel operator isogradient filter; Represents preprocessed 3D CT image data The partial derivative in the width direction is calculated using the Sobel operator isogradient filter; This represents the focus control parameter, which controls the degree of focus of the weight distribution. The larger the value, the higher the relative weight obtained in the high gradient region. The example value is 5. This represents the natural exponential function.

[0036] In one embodiment, the three-dimensional gradient magnitude map is as follows: Figure 8 As shown, this characterizes the degree of grayscale change at each voxel location in the image; as Figure 9 As shown, the gradient magnitude 3D visualization is a three-dimensional surface representation of the gradient magnitude map, which intuitively presents the spatial distribution of high gradient regions in the image.

[0037] 3) Feature selection based on stability and discriminative power To reduce feature redundancy and select the most discriminative feature subset, a dual-index screening mechanism is used to comprehensively evaluate the stability of each feature among samples of the same class and its discriminative power among samples of different classes. Based on this, the comprehensive screening score of each feature in the initial radiomics feature vector is calculated, expressed as: , In the formula, Indicates the first Features The overall screening score indicates that the feature is more stable in the same type of samples and has a stronger discriminative power between different categories. Indicates feature index; Indicates the first Features The intra-class stability score indicates that the feature exhibits less numerical fluctuation and better robustness among samples of the same class. The calculation method is expressed as follows: ; The total number of pathological categories of lymphoma is determined by the annotations in the dataset; This index represents the pathological category of lymphoma, with values ​​ranging from [value range missing]. ; Indicates the first The first of all training samples of class Features The standard deviation of the values; Indicates the first The first of all training samples of class Features The average of the values; Indicates the first Features The inter-class discriminative score indicates that the feature has a greater discriminative ability among samples of different classes. A higher score indicates a greater difference between samples of different classes. The calculation method is expressed as follows: ; Indicates the first Features The global average across all training samples of all categories; This represents the weighting coefficient for the stability score, with an example value of 0.4, used to balance the importance of the stability score and the discriminative score in the overall score. This represents the weighting coefficient of the discrimination score, with an example value of 0.6, used to balance the importance of the stability score and the discrimination score in the overall score.

[0038] S4. Construction of a classification model for childhood lymphoma types Accurate identification of childhood lymphoma types depends not only on imaging features but also on clinical factors such as patient age and macroscopic tumor location. Conventional deep convolutional neural networks can automatically learn abstract feature representations of images, but their learning process lacks explicit use of prior medical knowledge, which may cause the network to focus on image artifacts or normal anatomical structures that are irrelevant to diagnosis, thus affecting the interpretability and generalization performance of the model.

[0039] This invention constructs a multimodal fusion and attention-guided deep classification model. It fuses enhanced radiomics features, clinical prior features, and deep features modulated by region-aware weights, and utilizes a dynamic gating mechanism to achieve adaptive feature fusion. The model outputs classification probabilities and estimates of the model's cognitive uncertainty regarding the prediction results. The specific steps are as follows: 1) Multimodal feature construction and enhancement We consider introducing readily available clinical prior features and fusing them with screened radiomics features through nonlinear enhancement to construct a more discriminative multimodal feature representation. This is achieved through a fully connected layer containing feature interaction terms, represented as follows: , In the formula, This represents the enhanced feature vector, with dimension . It integrates radiomics information with prior clinical information, and has stronger characterization capabilities; This indicates the dimension of the enhanced feature vector, with an example value of 128. This represents the image omics feature vector obtained through screening, with a dimension of . It is based on the comprehensive screening score. For all Sort the initial features in descending order and select the top-ranked features. A feature vector composed of 1 feature; This indicates the number of radiomics features after screening; The patient's age is represented as a scalar quantity in years and is considered clinical prior information. A rough three-dimensional spatial vector representing the location of the tumor, with dimensions of... By calculating the three-dimensional gradient magnitude map The three-dimensional barycenter coordinates of all voxels with a mid-amplitude value higher than a given threshold are obtained. This belongs to clinical prior information derived from images; This indicates the centroid coordinates of the tumor region along the depth direction of the image. This indicates the centroid coordinates of the tumor region in the image height direction; This indicates the centroid coordinates of the tumor region along the image width direction; This indicates an embedding encoding function, which converts scalars... sum vector Convert and splice into one A dense vector of dimension, specifically implemented as ; This represents the ReLU activation function, used to implement nonlinear mappings; This represents the dimension of the clinical feature embedding vector, with an example value of 16. This represents the first transformation weight matrix, which acts as a linear transformation on the radiomics feature vectors, and has a dimension of . , are trainable parameters; This represents the second transformation weight matrix, which acts on the linear transformation of the embedded clinical feature vectors, and has a dimension of [missing information]. , are trainable parameters; This represents the transformation bias vector, with dimension . , are trainable parameters; The feature interaction strength coefficient is a trainable scalar parameter that controls the contribution of the nonlinear interaction term to the final enhanced feature. Let represent the first projection weight matrix, which is the projection weight matrix of the radiomics feature vectors before interaction, with dimensions of . , are trainable parameters; The second projection weight matrix is ​​the projection weight matrix embedded with the clinical feature vectors before the interaction, and has dimensions of [missing information]. , are trainable parameters; This represents the Hadamard product, which is an element-wise multiplication operation.

[0040] 2) Deep Feature Extraction and Spatial Attention Guidance A 3D convolutional neural network backbone is used to extract depth feature maps from preprocessed 3D CT image data. A spatial attention guidance module is employed, using a region-aware weight matrix as a prior attention template to spatially recalibrate the depth feature maps, making the network more focused on potential lesion regions. This is represented as follows: , In the formula, The attention feature map is a depth feature map after spatial attention modulation, with a size of [size missing]. ; The depth feature map is generated by processing preprocessed 3D CT image data through a 3D convolutional neural network backbone (such as 3D-ResNet, 3D-DenseNet, etc.). The multi-channel feature map obtained by forward propagation has a size of ; This represents a copy function along the channel dimension, which copies the attention weight matrix in two-dimensional space. Next, a four-dimensional attention tensor with the same size as the depth feature map is formed; This represents the size of the depth feature map in the depth direction; This represents the size of the depth feature map in the height direction; This represents the dimension of the depth feature map in the width direction; The number of channels in the depth feature map; This represents the resampling function, which converts samples of size 1 to 2. Region-aware weight matrix Downsampling to the depth feature map spatial size is performed using trilinear interpolation. match, This represents the spatial dimensions of the depth feature map; This represents the attention scaling parameter, which is a trainable scalar parameter used to adaptively adjust the strength of the prior attention template; This represents the attention bias parameter, which is a trainable scalar parameter used to adaptively adjust the attention baseline.

[0041] 3) Fusion of Feature Interaction and Dynamic Gating The enhanced feature vector is fused with the global deep feature vector extracted from the deep feature map. A dynamic gating mechanism adaptively determines the contribution ratio of the two types of features in the final decision, achieving information complementarity, as shown below: , In the formula, Represents the joint feature vector, with dimension . It integrates radiomics features enhanced by screening based on prior medical knowledge and deep abstract features adaptively learned from images; Represents a dynamic gated vector with dimension . Each element has a value between 0 and 1, controlling the proportion of the corresponding dimension of the enhanced feature vector retained in the fusion process. The calculation method is expressed as follows: ; Representing a deep feature vector, obtained by analyzing the deep feature map. The result is obtained by performing three-dimensional global average pooling, in the spatial dimension ( Take the average value over the channel, retaining the channel dimension, with dimension 1. ; express Transpose of; This represents the weight matrix of the first gated linear transformation, which acts on the gated linear transformation of the enhanced feature vectors, and has a dimension of . , are trainable parameters; This represents the weight matrix of the second gated linear transformation, applied to the deep feature vector. The gated linear transformation, with dimension , are trainable parameters; The bias vector representing the gating mechanism has dimension 1. , are trainable parameters; Represents the feature interaction matrix, with dimension . , are trainable parameters used to compute the bilinear interaction between augmented features and deep features; The dimension is A completely one vector; The projection weight matrix represents the depth feature vector, with dimension 1. , are trainable parameters used to project deep feature vectors onto dimensional . The space.

[0042] 4) Classifier construction and uncertainty estimation The joint feature vector is input into the classifier to obtain the raw scores for each category. Simultaneously, a learnable data-dependent temperature parameter is introduced to scale the scores, quantifying the cognitive uncertainty of the model's predictions. A flexible positive function is used to generate the classification probabilities, expressed as: , In the formula, This represents the class probability vector, which is the class probability vector predicted by the model, and has dimensions of 1. And satisfy ; This represents the original score vector, with dimension . The calculation method is expressed as The confidence level, or score, represents the model's "raw confidence level" or score for each category before probability normalization. A higher value indicates that the model believes the sample is more likely to belong to the corresponding category. Represents the original score vector The The element corresponding to the element The original scores for each category; This represents the weight matrix of the classifier, with dimension 1. , are trainable parameters, Joint feature mapping of dimension to A dimensional category score space; Indicates difference from The pathological category index for lymphoma, with values ​​ranging from [value range missing]. ; Indicates difference from and The pathological category index for lymphoma, with values ​​ranging from [value range missing]. ; Represents the Softmax function; This represents the bias vector of the classifier, with dimension 1. , are trainable parameters; This represents the temperature parameter, a scalar greater than 0, used to quantify the uncertainty of the model's prediction for the current sample. A larger value indicates higher uncertainty. The calculation method is expressed as follows: ; This represents the minimum constant value of the temperature parameter, used to ensure numerical stability. An example value is shown below. ; This represents a flexible positive function, ensuring that the output is a positive number. The weight vector representing the temperature regression branch has dimensions of . , are trainable parameters used to map the information of the joint feature vector to a scalar; express Transpose of; The bias scalar representing the temperature regression branch is a trainable parameter.

[0043] S5. Definition and Calculation of Loss Function Conventional cross-entropy loss functions only focus on the degree of matching between predicted probabilities and true labels, without considering the consistency that different feature modalities should have when representing the same pathological entity, nor do they utilize the uncertainty information of the model itself to prevent overfitting to fuzzy or difficult samples.

[0044] like Figure 10 As shown, this invention optimizes classification accuracy, promotes feature representation consistency, and utilizes uncertainty information for regularization through a composite loss function, thereby making the training process more robust. The specific steps are as follows: 1) Focusing classification loss based on temperature scaling and uncertainty weighting Focused cross-entropy loss is used as the basic classification loss to address potential class imbalance in the data. Simultaneously, a temperature parameter is used to dynamically weight the loss for each sample. When the model's prediction uncertainty for a particular sample is high, the contribution of that sample to the total batch loss is automatically reduced, preventing the model from forcibly fitting samples that are difficult to judge. This is expressed as: , In the formula, This represents batch classification loss. By focusing on difficult-to-classify samples and dynamically adjusting sample weights based on model prediction uncertainty, classification accuracy can be optimized and the model can be prevented from overfitting to ambiguous samples. Indicates the number of samples in the training batch; Indicates the sample index within the batch; Indicates the first The one-hot encoding of the true label of each sample, when the sample belongs to the first... The value is 1 if the condition is met, otherwise it is 0. Indicates the first The predicted number of samples The probability of a class is the probability of the model on the first class. The i-th element of the predicted class probability vector for each sample; Indicates the focus parameter, This is used to reduce the loss weight of easily classified samples, making the model pay more attention to samples that are difficult to classify. An example value is 2. This represents an uncertainty-weighted function with a range of . According to the Temperature parameters of a sample The loss weight is dynamically adjusted, and the calculation method is expressed as follows: ; This represents the weighted attenuation coefficient, which controls the rate at which the weight decreases as the temperature parameter increases. An example value is 0.8. Indicates the temperature reference standard, when The time loss weight is 1, and the example value is 1.0.

[0045] 2) Region-aware feature consistency regularization loss To promote consistency between deep feature representations and medical prior knowledge, a consistency regularization loss is constructed using a region-aware weight matrix. This encourages the local feature vector of each sample's deep feature map in the potential lesion region to be as close as possible to the average feature vector of all samples in the same class in that region, thereby enhancing the intra-class compactness and semantic consistency of the feature representation. This is expressed as: , In the formula, The region-aware feature consistency regularization loss utilizes prior weighting of image gradients to force similar samples to converge in the deep feature representation of potential lesion regions, thereby improving the intra-class compactness and semantic consistency of features. This represents the total number of non-overlapping 3D voxel blocks into which the depth feature map is spatially divided; Indicates the voxel block index; Represents the region-aware weight matrix In the The average value of all voxels within a voxel block represents the importance of that region based on prior knowledge; Indicates the first Depth feature map of each sample In the The average value of the eigenvectors at all spatial locations within an individual primitive block; Indicates the first The true class label of each sample is an integer index; Indicates that the current batch belongs to the first... All samples of class in the first The average feature vector of an individual pixel block is updated using a moving average method; This represents the L2 norm of a vector.

[0046] 3) Calculation of total loss function The total loss function for model training is a weighted sum of the focus classification loss and the feature consistency regularization loss. By adjusting the weight coefficients to balance the impact of these two losses on model optimization, it can be expressed as: , In the formula, It represents the total loss during model training, while taking into account both classification accuracy and the quality and consistency of feature representation; This represents the weight coefficient of the regularization loss term, used to balance the importance of classification loss and feature consistency loss in the total loss. An example value is 0.05.

[0047] S6. Training on Classification Models for Childhood Lymphoma Types The training process aims to utilize labeled training data and automatically adjust all trainable parameters in the classification model through optimization algorithms, enabling the model to learn to accurately identify the type of childhood lymphoma from input CT images and related clinical information.

[0048] Training is conducted iteratively using forward and backward propagation. Specifically, in each training iteration, a small batch of samples is randomly selected from the training set. For each sample within the batch, the CT image preprocessing and adaptive grayscale window adjustment described in S2 are performed sequentially to obtain standardized 3D CT image data. Then, the radiomics feature extraction and region-aware gradient-weighted filtering described in S3 are performed to obtain the filtered radiomics features. At the same time, the corresponding clinical prior features (including age and tumor location inferred from the image) are prepared for the sample. Then, the standardized 3D CT image data, the screened radiomics features and clinical prior features are input into the multimodal fusion and attention-guided deep classification model constructed in S4. Forward propagation calculation is performed to obtain the model's predicted probability of the current sample's category and temperature parameters. Then, based on the prediction results and the sample's true pathological label, the total loss function value is calculated according to the composite loss function formula defined in S5 to measure the difference between the current model prediction and the actual situation. Then, a gradient-based optimization algorithm is used to perform backpropagation, calculate the gradient of the total loss with respect to all parameters of the model, and use these gradients to update the parameters to reduce the loss value. This process is repeated continuously, and the model's classification ability on the training set is gradually improved.

[0049] Throughout the training process, model performance is continuously monitored using an independent validation set. After each training cycle or several cycles, the model's classification performance (such as accuracy, macro-average F1 score, etc.) is evaluated on the validation set.

[0050] The decision to stop training is based on an early stopping strategy using validation set performance: when the loss function value on the validation set no longer decreases within several consecutive preset training cycles, it is considered that the model performance has stopped improving and overfitting may have occurred, at which point the training iteration is stopped.

[0051] Select the model parameter state corresponding to the training cycle that has the best performance on the validation set (such as the highest overall index or the lowest loss) as the optimal model after training, and use it for subsequent testing and automatic classification.

[0052] S7. Automatic Classification of Childhood Lymphoma Types Once the childhood lymphoma classification model has been trained according to the procedure described in S6, it can be deployed for the automatic classification of new childhood lymphoma cases. For a new pediatric lymphoma patient with an unknown pathological type, firstly, acquire enhanced CT scan images of the chest, abdomen, or pelvis before treatment and obtain their age information; Then, the patient's original 3D CT image data is input into a pre-trained automated classification system. The system automatically performs a series of standardization processes: first, CT image preprocessing and adaptive gray-scale window adjustment as described in S2 are performed, including Z-score standardization based on global statistics and adaptive gray-scale window calculation and mapping combined with priors, to obtain preprocessed 3D CT image data with enhanced contrast and consistent intensity scale. Then, according to the strategy defined in S3, the system extracts global radiomics features from the preprocessed images and uses the calculated 3D gradient magnitude map to construct region-aware weights, weighting and focusing the features. Finally, a dual-index screening mechanism based on stability and discriminative power is used. The system obtains a concise subset of radiomics features. Simultaneously, it calculates the coarse 3D spatial location of the tumor from the preprocessed image (estimated by the centroid of a high-gradient region) and constructs a clinical prior feature vector together with the patient's age. Then, these processed data are fed into a trained multimodal fusion classification model, which performs forward propagation calculations: the preprocessed image extracts depth features through a 3D convolutional neural network backbone and undergoes spatial attention modulation guided by region-aware weights; the selected radiomics features are fused with clinical features and nonlinearly enhanced, and then the enhanced feature vector is adaptively fused with the modulated depth features through a dynamic gating mechanism.

[0053] The fused joint features are processed by the classifier layer, which outputs the original scores of the model for each preset lymphoma pathology type of the sample. The scores are then scaled and normalized using a data-dependent temperature parameter and a Softmax normalization method to generate a class probability vector. The class with the highest probability in this vector is determined by the system as the model's automatic classification prediction result.

[0054] In addition, the system can output the temperature parameter values ​​calculated by the model as a quantitative reference for the cognitive uncertainty of this prediction, thereby realizing fully automated analysis from the original CT images and basic clinical information to the probability prediction of lymphoma type, providing clinicians with a fast and objective auxiliary diagnostic reference.

[0055] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An automatic classification method for childhood lymphoma types based on radiomics, characterized in that, Includes the following steps: S1. Collect CT images of childhood lymphoma and construct a dataset of childhood lymphoma CT images; S2. Preprocess and adaptive grayscale window adjustment are performed on the original 3D CT image to obtain preprocessed 3D CT image data; S3. Perform radiomics feature extraction and gradient-weighted filtering for region perception on the preprocessed 3D CT image data to obtain a comprehensive filtering score for each feature. S4. Construct a classification model for childhood lymphoma types by fusing enhanced radiomics features, clinical prior features, and deep features modulated by region-aware weights, and using a dynamic gating mechanism to achieve adaptive feature fusion and output a category probability vector. S5. During the model training process, the model is trained and optimized using a composite loss function to obtain a well-trained classification model for childhood lymphoma types. S6. Input the CT images of the pediatric lymphoma to be classified into the trained pediatric lymphoma classification model to obtain the classification results.

2. The automatic classification method for childhood lymphoma types based on radiomics according to claim 1, characterized in that, Z-fraction normalization of the original 3D CT images was performed to eliminate intensity scale differences caused by the scanning protocol. An adaptive gray-scale window adjustment method driven by image statistics and medical prior knowledge was adopted to dynamically optimize image contrast and focus the visual attention on the density range most likely to contain tumor tissue. Raw 3D CT image data Z-score standardization is performed, using the voxel gray mean and standard deviation of the entire training set to eliminate intensity scale differences caused by different scanning protocols, resulting in standardized 3D CT image data. For standardized 3D CT image data Based on the statistical characteristics of its gray-level histogram and combined with the prior typical CT values ​​of childhood lymphoma, the optimal window level and width of the adaptive window are dynamically calculated. The adaptive window is then used to perform linear contrast stretching and truncation on the image, mapping the gray-level values ​​of key density intervals to the range of 0 to 1, thus obtaining preprocessed 3D CT image data. .

3. The automatic classification method for childhood lymphoma types based on radiomics according to claim 2, characterized in that, Three-dimensional radiomics features were extracted from preprocessed three-dimensional CT image data to form an initial radiomics feature vector. The three-dimensional gradient magnitude map of the preprocessed three-dimensional CT image data is calculated based on the gradient magnitude region-aware weighting mechanism to construct the region-aware weight matrix. A dual-index screening mechanism is adopted to comprehensively evaluate the stability of each feature in the same type of sample and its discriminative power among different types of samples, and to calculate the comprehensive screening score of each feature in the initial radiomics feature vector.

4. The automatic classification method for childhood lymphoma types based on radiomics according to claim 3, characterized in that, Three-dimensional radiomics features were extracted from preprocessed three-dimensional CT image data to form an initial radiomics feature vector. This includes: selecting 3D radiomics feature categories, including first-order statistical features and texture features; and using 3D radiomics feature extraction tools to process preprocessed 3D CT image data. Perform full-map feature calculations to generate initial imageomics feature vectors. .

5. The automatic classification method for childhood lymphoma types based on radiomics according to claim 4, characterized in that, By introducing clinical prior features and fusing them with screened radiomics features through nonlinear enhancement, an enhanced feature vector is obtained. .

6. The automatic classification method for childhood lymphoma types based on radiomics according to claim 5, characterized in that, A 3D convolutional neural network backbone is used to extract depth feature maps from preprocessed 3D CT image data. A spatial attention guidance module is then employed, using a region-aware weight matrix as a prior attention template to spatially recalibrate the depth feature maps, resulting in attention feature maps. .

7. The automatic classification method for childhood lymphoma types based on radiomics according to claim 6, characterized in that, The enhanced feature vector is fused with the global deep feature vector extracted from the deep feature map. The contribution ratio of the two types of features in the final decision is adaptively determined by a dynamic gating mechanism to generate a joint feature vector. The joint feature vector is input into the classifier to obtain the original scores of each category. A learnable data-dependent temperature parameter is introduced to scale the scores to quantify the cognitive uncertainty of the model prediction. The classification probability is generated through a flexible positive function.

8. The automatic classification method for childhood lymphoma types based on radiomics according to claim 7, characterized in that, The composite loss function includes batch classification loss and region-aware feature consistency regularization loss.

9. The automatic classification method for childhood lymphoma types based on radiomics according to claim 8, characterized in that, The batch classification loss uses focused cross-entropy loss as the basic classification loss, and dynamically weights the loss of each sample using temperature parameters.

10. The automatic classification method for childhood lymphoma types based on radiomics according to claim 9, characterized in that, The region-aware feature consistency regularization loss is constructed using the region-aware weight matrix.

Images