A fracture image classification method and system fusing feature enhancement

Abstract

The application belongs to the field of computer vision and artificial intelligence, and provides a fracture image classification method and system fusing feature enhancement, which comprises the following steps: after an initial convolution processing based on a pretreated fracture image, the fracture image is sequentially processed by four dense blocks and a feature fusion enhancement module to obtain final fusion features, and then the fusion features are processed by a transition layer and a global pooling layer to obtain modified fusion features; and finally, fracture image classification is performed based on the modified fusion features to obtain a classification result; the feature fusion enhancement module is embedded after each dense block to enhance the feature expression capability at each level of the network, so that the local details and global dependency of the fracture can be captured at different abstraction levels, and the features extracted from each level are enhanced; the output of the feature fusion enhancement module after the last dense block will be used for the final classification task, so that the local detail features and the overall structure information can be extracted, and the accuracy and robustness of fracture classification can be effectively improved.

Description

Technical Field

[0001] This invention belongs to the field of computer vision and artificial intelligence technology, specifically relating to a fracture image classification method and system that integrates feature enhancement. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] Fracture classification in medical image analysis is a key challenge. Traditional methods rely on manual feature extraction techniques, such as edge detection and morphological transformations, but perform poorly on complex fractures and low-quality images, failing to meet clinical needs. In recent years, convolutional neural networks (CNNs) have made significant progress in medical image analysis, automatically learning multi-level features from raw images. However, while standard CNNs excel in local feature extraction, their ability to capture long-range dependencies and global contextual information is limited, whereas fracture diagnosis often requires considering both minute crack details and overall bone structure changes.

[0004] To overcome the limitations of CNNs, researchers have proposed several improvement methods. DenseNet enhances feature propagation and reuse through dense connections between layers, improving model performance and parameter utilization efficiency. However, even advanced CNN architectures still face challenges in processing the complex global structure and minute details of fracture images, especially in scenarios that require integrating multi-scale and multi-modal information.

[0005] Saliency-guided techniques aim to help models identify key regions with diagnostic value. Traditional saliency detection relies on low-level visual features to identify prominent regions, but in medical images, lesions are often not significantly different from surrounding healthy tissue, leading to poor results. Deep learning methods can generate saliency maps that better meet the needs of medical diagnosis, but often require additional labeled data or two-stage processing, increasing computational burden and complexity. Existing saliency-guided methods mostly use spatial weighting, ignoring the relationships between feature channels and failing to fully utilize multidimensional information.

[0006] Attention mechanisms have been introduced into visual models to enhance focus on important regions. Techniques such as self-attention and channel attention allow models to dynamically adjust their focus on different locations and channels, improving the ability to identify key features. However, when processing complex medical images, attention mechanisms alone often need to be combined with other techniques to achieve optimal results.

[0007] In terms of feature fusion, existing multi-branch models extract features at different levels through different network paths, and then fuse them through simple concatenation or weighted summation. Although these methods can integrate features at different levels, they often ignore high-order interaction relationships between features, resulting in information redundancy and wasted computational resources, which is detrimental to resource-constrained medical deployments.

[0008] Feature compression and redundancy removal techniques, such as Tucker decomposition, can decompose high-dimensional feature tensors into compact core tensors and factor matrices, effectively reducing the number of parameters and computational complexity. However, traditional applications often perform this statically, lacking adaptability to the complexity of the input data and failing to dynamically adjust the complexity based on the image content.

[0009] For second-order feature extraction from fracture images, existing methods mainly rely on first-order features extracted through global pooling or convolution operations, ignoring the covariance relationships between features. However, these higher-order statistical information are crucial for identifying subtle fracture features. Although some studies have begun to explore techniques such as second-order pooling and bilinear pooling, the balance between computational efficiency and feature representation capability still needs to be optimized.

[0010] In summary, existing technologies still face challenges such as efficiently integrating local and global information, adaptively adjusting feature strategies, enhancing focus on key regions, and reducing computational redundancy while maintaining high classification accuracy. Summary of the Invention

[0011] To address the aforementioned issues, this invention proposes a fracture image classification method and system that integrates feature enhancement. By employing a deep learning model and introducing multiple feature enhancement strategies, this invention can extract local detailed features while also considering overall structural information, thereby effectively improving the accuracy and robustness of fracture classification and providing reliable technical support for intelligent medical image analysis.

[0012] According to some embodiments, the first aspect of the present invention provides a fracture image classification method with fused feature enhancements, employing the following technical solution:

[0013] A fracture image classification method incorporating enhanced features includes:

[0014] Acquire fracture medical images and perform image preprocessing;

[0015] Based on preprocessed fracture medical images, a pre-trained fracture image classification model is used for classification, specifically:

[0016] After initial convolution processing, the preprocessed fracture images are processed sequentially through four dense blocks and a feature fusion enhancement module to obtain the final fusion features. Then, after processing through a transition layer and a global pooling layer, the corrected fusion features are obtained. Based on the corrected fusion features, fracture images are classified to obtain the classification results.

[0017] The processing steps for each dense block and feature fusion enhancement module are as follows:

[0018] Primary features are extracted based on dense blocks;

[0019] The primary features are enhanced and then split into local features and preserved features along the channel dimension. The local features are enhanced and then fused with the preserved features to obtain the local enhanced features.

[0020] The local enhancement features are subjected to multi-scale dilated convolution and adaptive saliency guidance processing. The processing results are then fused with the local enhancement features to obtain globally relevant features.

[0021] The feature tensor is obtained by concatenating the local enhancement features and the global attention features. The second-order feature optimization tensor decomposition method is used to optimize the feature tensor. The optimized tensor decomposition method is then used to decompose and reconstruct the feature tensor to obtain the fused features.

[0022] According to some embodiments, a second aspect of the present invention provides a fracture image classification system with fused feature enhancements, employing the following technical solution:

[0023] A fracture image classification system incorporating enhanced features, comprising:

[0024] The image preprocessing module is configured to acquire fracture medical images and perform image preprocessing.

[0025] The image classification module is configured to classify pre-processed fracture medical images using a pre-trained fracture image classification model, specifically:

[0026] After initial convolution processing, the preprocessed fracture images are processed sequentially through four dense blocks and a feature fusion enhancement module to obtain the final fusion features. Then, after processing through a transition layer and a global pooling layer, the corrected fusion features are obtained. Based on the corrected fusion features, fracture images are classified to obtain the classification results.

[0027] The processing steps for each dense block and feature fusion enhancement module are as follows:

[0028] Primary features are extracted based on dense blocks;

[0029] The primary features are enhanced and then split into local features and preserved features along the channel dimension. The local features are enhanced and then fused with the preserved features to obtain the local enhanced features.

[0030] The local enhancement features are subjected to multi-scale dilated convolution and adaptive saliency guidance processing. The processing results are then fused with the local enhancement features to obtain globally relevant features.

[0031] The feature tensor is obtained by concatenating the local enhancement features and the global attention features. The second-order feature optimization tensor decomposition method is used to optimize the feature tensor. The optimized tensor decomposition method is then used to decompose and reconstruct the feature tensor to obtain the fused features.

[0032] According to some embodiments, a third aspect of the present invention provides a computer-readable storage medium.

[0033] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in a fracture image classification method with fused feature enhancement as described in the first aspect above.

[0034] According to some embodiments, a fourth aspect of the present invention provides a computer device.

[0035] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of a fracture image classification method with fusion feature enhancement as described in the first aspect above.

[0036] According to some embodiments, a fifth aspect of the present invention provides a computer program product or computer program.

[0037] This invention provides a computer program product or computer program including computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the steps in a fracture image classification method with fused feature enhancement as described in the first aspect above.

[0038] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0039] This invention balances local details and global structural information in fracture image classification. By embedding a Local Context and Dependency Attention (LCAD) module and a low-rank fusion module into deep backbone networks such as DenseNet, it achieves precise localization of fracture areas and efficient fusion of multi-scale features. This module utilizes multi-branch local convolution to enhance fine-grained information such as microcracks and edges, while combining multi-scale dilated convolution and channel attention mechanisms to capture long-distance dependencies, enabling the model to maintain high discriminative power even with complex fracture morphologies. Second-order features are extracted through bilinear descriptor encoding, enhancing the ability to express subtle fracture features. Based on a saliency-guided mechanism, the model can automatically focus on key fracture regions. The low-rank fusion technique using Tucker decomposition and adaptive complexity adjustment significantly reduces computational cost while improving classification performance, meeting the dual requirements of efficiency and accuracy in practical medical scenarios. Attached Figure Description

[0040] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0041] Figure 1 This is a flowchart of a fracture image classification method with fusion feature enhancement according to an embodiment of the present invention;

[0042] Figure 2 This is a flowchart of the algorithm for the fracture image classification model in this embodiment of the invention;

[0043] Figure 3 This is a flowchart of the feature fusion enhancement module in an embodiment of the present invention;

[0044] Figure 4 This is a flowchart of the low-rank fusion and second-order feature enhancement layer in an embodiment of the present invention. Detailed Implementation

[0045] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0046] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0047] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0048] Example 1

[0049] like Figure 1 As shown, this embodiment provides a fracture image classification method based on fused feature enhancement. In this embodiment, the method includes the following steps:

[0050] Acquire fracture medical images and perform image preprocessing;

[0051] Based on preprocessed fracture medical images, a pre-trained fracture image classification model is used for classification, specifically:

[0052] After initial convolution processing, the preprocessed fracture images are processed sequentially through four dense blocks and a feature fusion enhancement module to obtain the final fusion features. Then, after processing through a transition layer and a global pooling layer, the corrected fusion features are obtained. Based on the corrected fusion features, fracture images are classified to obtain the classification results.

[0053] The processing steps for each dense block and feature fusion enhancement module are as follows:

[0054] Primary features are extracted based on dense blocks;

[0055] The primary features are enhanced and then split into local features and preserved features along the channel dimension. The local features are enhanced and then fused with the preserved features to obtain the local enhanced features.

[0056] The local enhancement features are subjected to multi-scale dilated convolution and adaptive saliency guidance processing. The processing results are then fused with the local enhancement features to obtain globally relevant features.

[0057] The feature tensor is obtained by concatenating the local enhancement features and the global attention features. The second-order feature optimization tensor decomposition method is used to optimize the feature tensor. The optimized tensor decomposition method is then used to decompose and reconstruct the feature tensor to obtain the fused features.

[0058] This invention provides a medical image fusion method based on a large kernel attention mechanism, such as... Figure 1 As shown, the process includes the following steps:

[0059] 1. Input fracture medical images

[0060] We acquire publicly available fracture datasets as fracture medical images, which contain images of different types of fractures and non-fracture images, suitable for fracture detection and classification tasks in medical imaging.

[0061] 2. Image Preprocessing

[0062] Image preprocessing mainly includes image resizing, data augmentation, and normalization. First, all images are uniformly scaled down to [size missing]. The size of the dataset is carefully controlled to ensure input consistency. For the training set, data augmentation methods such as random rotation, flipping, and cropping are employed to increase data diversity and improve the model's generalization ability. All images are standardized using the mean and standard deviation of the ImageNet dataset to normalize pixel values, thereby accelerating model convergence and improving training stability. The dataset is divided into training, validation, and test sets to ensure model evaluation and tuning at different stages. This technique uses standard data loading methods, representing the dataset as a set containing image samples and corresponding class labels, and inputs them in batches to improve computational efficiency. The mathematical expression of the data input is as follows:

[0063] The data input in this embodiment can be represented as follows:

[0064] (1);

[0065] in, Represents a dataset consisting of image samples. and their corresponding category labels composition. Represents the input image of a fracture or no fracture, with a size of [size missing]. , , .in, and These represent the image's height and width, respectively, and are fixed at 224. , The number of channels, with a value of 3 ( image). For the category label of the image, Image showing a fracture. 0 indicates an image without fractures. This represents the total number of images in the dataset, including the training, validation, and test sets. This data input method allows for the efficient organization and loading of medical image data for training and inference by deep learning models.

[0066] like Figure 2 As shown in the figure, this embodiment proposes a fracture image classification model including a backbone network module and a feature fusion enhancement module.

[0067] 3. Fracture image classification model and processing workflow, including:

[0068] The fracture image classification model includes a backbone network module, which consists of an initial convolutional layer, four dense blocks, a transition layer, and a global pooling layer; a feature fusion enhancement module is connected after each dense block.

[0069] Specifically, the fracture image classification model uses a deep neural network as the backbone, selecting DenseNet201 as the backbone and optimizing it. DenseNet201 employs a dense connection mechanism, allowing features from each layer to be directly passed to subsequent layers, enhancing information flow and improving the model's feature representation ability. The backbone module mainly consists of an initial convolutional layer, multiple dense blocks, transition layers, and a global pooling layer. Each layer within a dense block uses short connections to improve feature transfer efficiency and reduce the gradient vanishing problem.

[0070] Based on the Dense Block structure, an additional feature fusion module is embedded to enhance the modeling ability for local details and global dependencies. The primary features output by each dense block... The input image is passed to the feature fusion and enhancement module to further optimize the feature representation, enabling the model to more accurately classify fracture images. Through multi-level convolution and pooling operations, multi-scale, high-level feature representations are extracted from the input image.

[0071] The preprocessed fracture image is input into the fracture image classification model. It first passes through a primary convolutional layer for feature extraction, resulting in a densely packed input image. The feature extraction process for dense blocks can be represented as follows:

[0072] (2);

[0073] in, The feature map output by the dense block, i.e., the primary feature. Represents the computational mapping of DenseNet201. These are network parameters.

[0074] The feature fusion enhancement module includes a local feature enhancement submodule and a local-to-global feature perception and fusion submodule;

[0075] In fracture image classification tasks, it is necessary to focus on both fine-grained local features and capture the overall bone structure and its long-range dependencies. Traditional convolutional models, which rely solely on a fixed kernel size, often struggle to address both requirements. To address this, we propose an LCDA (Local Context and Dependency Attention) feature fusion and enhancement mechanism. This mechanism integrates local and global information in a multi-stage, multi-modal manner and leverages Tucker decomposition to reduce redundancy and computational burden during the fusion process.

[0076] The local feature enhancement submodule includes convolutional layers, channel splitting layers, multi-branch local convolutional layers, and fusion layers.

[0077] After enhancing the primary features, they are then split along the channel dimension into local features and preserved features. The local features are then enhanced and fused with the preserved features to obtain the enhanced local features, specifically:

[0078] After performing local context enhancement on the primary features, primary extended features are obtained;

[0079] The primary extended features are split into local features and retained features along the channel dimension according to a set ratio;

[0080] Multi-branch local convolution is used to divide local features into two parts to obtain local features at different scales, and the local features at different scales are concatenated in the channel dimension to obtain the enhanced local features.

[0081] The enhanced local features are then concatenated and fused with the preserved features to obtain the enhanced local features.

[0082] Specifically, in receiving primary features from fracture images Then, local context enhancement is performed first. The core idea is to... Convolution reduces the number of channels from Expand to This allows features to have a greater expressive space.

[0083] Specifically, define the convolution kernel With bias And a non-linear activation function (such as GELU) is used for feature mapping, specifically:

[0084] (3);

[0085] (4);

[0086] in, Split This indicates that the feature tensor is divided into two parts along the channel dimension according to a specified ratio. Occupying the number of channels , used for local feature enhancement, is a local feature; Occupying the number of channels This method preserves most of the information from the original features for subsequent global information processing, thus retaining the features. This unbalanced segmentation ratio ensures that sufficient global contextual information is preserved while enhancing local details.

[0087] This embodiment employs a multi-branch local convolutional structure to incorporate local features. It is further divided into two parts, each using... and The convolution kernel is used to process the data, enabling the model to capture local features at different scales simultaneously.

[0088] (5);

[0089] (6);

[0090] (7);

[0091] in, and These are two branches of local features. and They are and Convolution kernel, This indicates a concatenation operation along the channel dimension. and They are function pairs and Enhanced features after processing yes SigmoidActivation function. This multi-branch structure enables the model to capture the edges, textures, and morphological features of the fracture area more precisely, improving the ability to detect minute cracks and subtle structural changes.

[0092] Then, and Local enhancement features are obtained by re-splicing along the channel dimension. :

[0093] (8);

[0094] This step enhances the network's ability to extract local details while retaining sufficient global information for subsequent processing.

[0095] The local and global feature perception and fusion submodule includes multi-scale dilated convolutional layers, adaptive saliency guidance layers, low-rank fusion and second-order feature enhancement layers.

[0096] Multi-scale dilated convolutions are performed on local enhancement features to obtain global initial features. Then, adaptive saliency-guided processing is applied to these global initial features, and the result is fused with the local enhancement features to obtain globally relevant features. Specifically:

[0097] The local enhancement features are split along the channel dimension to obtain two local enhancement sub-features;

[0098] After performing small-scale regular convolution and large-scale dilated convolution on the two local augmentation features respectively, local augmentation features of different scales are obtained and then concatenated to obtain the global initial features;

[0099] By using adaptive saliency guidance, the initial global features are processed separately in terms of channel and space before being concatenated to obtain the global features of interest.

[0100] Specifically, after local context enhancement, a multi-scale dilated convolutional structure is used to strengthen the perception of the overall skeletal structure and long-range dependencies. In this embodiment, the multi-scale dilated convolutional layer uses both small-scale regular convolutions and large-scale dilated convolutions, significantly expanding the receptive field.

[0101] (9);

[0102] (10);

[0103] (11);

[0104] (12);

[0105] in, Split It is a channel-level splitting operation. and It consists of two parts of a local enhancement feature, namely, a local enhancement feature sub-feature. It is small scale Convolution kernel, It is large scale Convolution kernel, This indicates a dilated convolution operation, with a dilation rate of [value missing]. , yes The corresponding bias term, yes It is the corresponding bias term. Processed through small-scale convolution The obtained small-scale features; Processed by large-scale dilated convolution The obtained large-scale features; yes and The result after concatenation is the global initial feature; yes Sigmoid Activation function. This multi-scale dilated convolutional structure enables the model to effectively capture positional and contextual information within the overall structure of the skeleton, providing a more comprehensive basis for determining the type and severity of fractures and enhancing the ability to model long-range dependencies.

[0106] This embodiment designs multi-branch local convolutional structures (such as the 3×3 and 5×5 convolutional branches in the LCAD module) and multi-scale dilated convolutions (small and large branches) to enable the model to simultaneously focus on local features at different scales, and integrates these features through a low-rank fusion module, thereby improving the ability to recognize various fracture morphologies.

[0107] The adaptive saliency guidance layer processes the initial global features separately in terms of channel and spatial dimensions before concatenation, resulting in globally relevant features. Specifically:

[0108] Global average pooling is used to capture the global context information of each channel of the global initial features. Based on the global context information, the dependencies between all channels are learned through nonlinear transformation to obtain the channel attention weights.

[0109] Channel-guided features are calculated using channel attention weights and global initial features;

[0110] The global initial features are downsampled to obtain downsampled features, and the spatial covariance matrix of the downsampled features is calculated.

[0111] The diagonal elements of the spatial covariance matrix based on the downsampling features are normalized to obtain the spatial saliency weights. Broadcast multiplication is then performed based on the spatial saliency weights and the downsampling features to obtain the spatial modulation features.

[0112] After upsampling the spatial modulation features, they are concatenated with local enhancement features and channel-guided features to obtain the globally relevant features.

[0113] Specifically, this embodiment designs a saliency guidance mechanism based on second-order statistics and adaptive attention, enabling the model to automatically focus on key diagnostic regions in fracture images. This mechanism operates simultaneously in both the channel and spatial dimensions, forming dual saliency guidance, allowing the model to automatically focus on key diagnostic regions in fracture images. Unlike traditional saliency detection methods, this embodiment achieves this through a dual dimension of feature channel importance and spatial criticality.

[0114] Channel salience guidance:

[0115] At the channel level, this embodiment employs a channel attention mechanism based on statistical characteristics, which can learn the importance of different channels for fracture recognition and dynamically adjust channel weights accordingly to highlight the expression of key feature channels. First, global average pooling is used to capture the global contextual information of each channel. :

[0116] (13);

[0117] in, It represents the coordinate index on the feature map. Indicates the channel index. , It is the number of channels. and These represent the spatial location indices of the feature map (in the height and width directions), respectively.

[0118] Then, the dependencies between channels are learned through nonlinear transformation:

[0119] (14);

[0120] (15);

[0121] in, and These are the parameters of the fully connected layers for dimensionality reduction and dimensionality expansion. It's the dimensionality reduction ratio. yes ReLU Activation function yes Sigmoid Activation function This is the channel descriptor vector obtained through global average pooling. The final channel attention weights are... The importance of each channel was modulated to obtain channel guiding features. :

[0122] (16);

[0123] This channel saliency mechanism can adaptively emphasize channels containing key fracture features while suppressing the influence of irrelevant or redundant channels.

[0124] Spatial saliency guidance:

[0125] In the spatial dimension, spatial saliency guidance based on bilinear descriptor coding (BDC) is introduced. The BDC module first downsamples the features to a fixed size to reduce the computational burden:

[0126] (17);

[0127] in, It is a downsampling operation that includes adaptive pooling and convolution, reducing the feature map size to a predefined size (e.g., ...). ), This is a downsampling feature. Then, the spatial covariance matrix of the downsampling feature is calculated to capture the dependencies between spatial locations:

[0128] (18);

[0129] (19);

[0130] (20);

[0131] (twenty one);

[0132] in, and It is the size of the feature map after downsampling. It is Reshape the tensor to the shape [B,C,-1]. It is the number of channels. This is the batch size; -1 indicates that the remaining dimensions will be automatically calculated (i.e., ...). ); It is a tensor shape reconstruction function that reconstructs a three-dimensional feature map into a two-dimensional form that facilitates the calculation of covariance. It is the mean of the downsampling features of all spatial locations; It is the centered feature, that is, the reshaped tensor of the original feature minus the mean of the downsampled feature. This is the covariance matrix of the downsampling features, reflecting the correlation between different channels. By extracting the diagonal elements of the covariance matrix of the downsampling features, the spatial significance distribution of each channel is obtained:

[0133] (twenty two);

[0134] (twenty three);

[0135] in, It is a function that extracts the diagonal elements of the covariance matrix, representing the variance within each channel, reflecting the degree of spatial variability of that channel; Through softmax Normalized channel weights, also known as spatial significance weights, are used to highlight channels with high spatial variability. This indicates that the dimension of the covariance matrix is... , It refers to the number of channels.

[0136] These spatial saliency weights are applied to BDC features to highlight regions with high spatial variability, which typically correspond to fracture locations:

[0137] (twenty four);

[0138] in, This indicates an appropriate broadcast multiplication operation. Through Weighted BDC features, or spatially modulated features, capture spatial saliency information by calculating feature covariance and extracting diagonal weights. This allows the model to focus more on the feature representation of the fracture area. This dual saliency-guided mechanism enables the model to automatically focus on key diagnostic areas when processing complex fracture images, improving classification accuracy and interpretability. This is especially beneficial for cases with indistinct fracture features or those located in unconventional locations, allowing for more accurate localization and identification of the fracture area and reducing the risk of missed or misdiagnosed cases.

[0139] Subsequently, the spatially saliency-enhanced features are upsampled back to their original size and then fused with the local enhancement features:

[0140] (25);

[0141] (26);

[0142] in, It is an upsampling operation. It is a balancing parameter that controls the contribution of BDC characteristics. These are upsampled BDC features, used in conjunction with local enhancement features. Integration to form a globally relevant feature By calculating the covariance matrix of downsampled features and extracting diagonal elements as channel weights, the model can capture the correlations between channels, which often correspond to unique patterns in different types of fractures. This second-order feature representation significantly enhances the model's ability to identify subtle fracture features, distinguishing fracture types that are not spatially obvious but have unique patterns in feature correlation. Through downsampling and covariance calculation, second-order statistical information of the features is further extracted, forming a multi-level second-order representation of fracture features, greatly improving the model's ability to represent complex fracture patterns.

[0143] The low-rank fusion and second-order feature enhancement layer works by using a second-order feature optimization tensor decomposition method based on the feature tensor. The optimized tensor decomposition method is then used to decompose and reconstruct the feature tensor to obtain the fused features. Specifically:

[0144] The covariance matrix is ​​calculated based on the second-order features of the feature tensor, and the second-order covariance matrix is ​​obtained.

[0145] The feature tensor is subjected to adaptive evaluation pooling and flattening to obtain tensor flattened features, and the complexity score of the tensor features is determined based on the tensor flattened features.

[0146] The adaptive rank is determined based on the complexity score of tensor features, and the optimized tensor decomposition method is obtained from the adaptive rank.

[0147] The diagonal elements of the second-order covariance matrix are extracted as channel weights, and the channel weights are mapped and reshaped to obtain the modulation coefficient tensor.

[0148] The core tensor is adjusted using the modulation coefficient tensor, and the modulation core tensor is decomposed based on the optimized tensor decomposition method to obtain the truncated modulation core tensor.

[0149] The fused features are obtained by reconstructing and fusing using the clipped modulation core tensor.

[0150] Specifically, acquiring local enhancement features With global attention features Then, a low-rank fusion module is introduced to complete the final feature fusion and compression based on Tucker decomposition. To do this, the two are first concatenated along the channel dimension to obtain the feature tensor. Then, a three-mode tensor decomposition method is used.

[0151] Second-order feature covariance representation:

[0152] The low-rank fusion module first calculates the second-order statistics of the input features, i.e., the second-order features of the feature tensor, to capture higher-order correlations between features. Specifically, for the input features (feature tensor)... Downsampling is performed using a fixed-size adaptive pooling operation, and then its covariance representation is calculated:

[0153] (27);

[0154] (28);

[0155] (29);

[0156] (30);

[0157] (31);

[0158] in, Through adaptive average pooling operation Processed features It is an adaptive average pooling operation; Through reshape The function will Reshape Two-dimensional tensor, This is the number of feature channels; -1 indicates that the remaining dimensions (all spatial locations) are automatically calculated. It is the spatial size of the feature map after pooling; It is the characteristic mean, and the calculation method is similar to that of formula (19), which is applied to all spatial locations. The eigenvalues ​​are averaged. It accesses the value of all channels at position (i,j) by index; yes The result after centering (minus the mean) ); It is the calculated covariance matrix, representing the correlation between different channels; Is with The corresponding matrix.

[0159] This second-order covariance representation can capture the interrelationships and higher-order dependencies between feature channels, containing rich structural information and discriminative features. It can provide richer structural information and texture description, making it particularly suitable for processing medical images with complex structures and subtle textures, such as fractures. For fracture images, this higher-order information is crucial for distinguishing minute fracture textures and morphological changes.

[0160] For fracture images, the correlation and covariance information between feature channels often contain important diagnostic clues, especially for small or atypical fractures. This embodiment introduces second-order feature processing modules, such as the covariance calculation and diagonal weight extraction function in the low-rank fusion module, and the BDC (Bilinear Descriptor Coding) feature encoding module in the LCAD module, enabling the model to capture the interrelationships between feature channels and enhance the ability to identify subtle fracture features.

[0161] Feature complexity evaluation and adaptive rank adjustment:

[0162] This embodiment innovatively designs a feature complexity estimator that adaptively adjusts the rank of the Tucker decomposition by analyzing the statistical properties of the input features.

[0163] (32);

[0164] (33);

[0165] (34);

[0166] in, It is an adaptive average pooling operation. and These are parameters of the fully connected layer. It is the ReLU activation function. yes Sigmoid Activation function. It is a tensor pooling feature, which is the feature information extracted from the feature tensor X through adaptive average pooling, and is used to evaluate the complexity of the feature. This is an operation that flattens a multidimensional tensor into a one-dimensional vector, making it easier for subsequent fully connected layers to process it. It is a tensor flattening feature. It is the complexity score of tensor features.

[0167] The output of the complexity estimator The value is between 0 and 1, reflecting the complexity of the input features.

[0168] Based on the complexity score, the model dynamically adjusts the rank of the Tucker decomposition:

[0169] (35);

[0170] in, It is the basic rank. It is the minimum rank ratio (e.g., 0.5). The adaptive rank mechanism not only enables efficient feature compression and fusion but also achieves adaptive allocation of computational resources through a complexity estimator. This allows the model to dynamically allocate computational resources based on the complexity of the input image. Lower tensor ranks are used for simple fracture images to improve computational efficiency, while higher tensor ranks are used for complex fracture images to ensure feature representation capabilities, achieving a balance between computational efficiency and classification accuracy. This innovation is particularly important for medical image processing because the complexity of fractures varies greatly among different cases. Introducing a complexity estimator and a complexity-based dynamic rank adjustment mechanism allows the model to adaptively adjust the tensor ranks and computational resources used according to the complexity of the input image, improving computational efficiency while maintaining recognition accuracy.

[0171] Tucker decomposition of covariance modulation:

[0172] This invention combines second-order feature representation with Tucker decomposition, modulating the core tensor with covariance information to enhance the expressive power of key features. First, the diagonal elements of the covariance matrix are extracted as channel weights. :

[0173] (36);

[0174] Then, these channel weights are mapped to the modulation coefficients of the core tensor through a fully connected layer. :

[0175] (37);

[0176] (38);

[0177] in, It is the rank of the three dimensions of the Tucker decomposition. These are parameters of the fully connected layer. yes Sigmoid The activation function, formula (38), reshapes the one-dimensional modulation coefficients into a modulation coefficient tensor with the same shape as the core tensor. This facilitates subsequent element-by-element modulation of the core tensor.

[0178] Based on Tucker decomposition theory, the input feature tensor It can be decomposed into the product of the core tensor and the three mode matrices:

[0179] (39);

[0180] in, It is the core tensor. There are three pattern matrices. Indicates along the first nTensor product of 1 / 2 dimensions.

[0181] Through modulation coefficient tensor Adjusting the core tensor enhances its ability to represent key features, resulting in the modulated core tensor. :

[0182] (40);

[0183] in, This represents element-wise multiplication. It is the modulation intensity coefficient (e.g., 0.1). This is the original core tensor. This covariance modulation mechanism enables the core tensor to adaptively emphasize the part containing key fracture information based on the second-order statistical properties of the input features, thereby improving classification accuracy.

[0184] According to adaptive rank The model's dynamic pruning factor matrix and core tensor:

[0185] (41);

[0186] (42);

[0187] (43);

[0188] in, , , It is based on adaptive rank The actual use of rank in dynamic calculation , , It is the pattern matrix (factor matrix) after the corresponding cropping. It is the trimmed modulation core tensor.

[0189] Finally, the fusion features are reconstructed using Tucker:

[0190] (44);

[0191] (45);

[0192] in, It is the result of projecting the feature tensor X onto a low-dimensional space; It is a reconstructed fusion feature that contains a compressed representation of local and global information. Through second-order feature enhancement and adaptive rank adjustment, it retains the most discriminative feature components and suppresses redundant channels.

[0193] After the first dense block and feature fusion enhancement module is processed, it is input into the next combination of dense block and feature fusion enhancement module, until all four dense blocks and feature fusion enhancement modules are processed in sequence, and the final fused feature is output.

[0194] The modified fusion features are then processed sequentially through a transition layer and a global pooling layer. .

[0195] This embodiment introduces a low-rank fusion mechanism to decompose the high-dimensional feature tensor into a core tensor and a series of factor matrices. Redundancy is reduced through orthogonal initialization constraints while preserving key information. Furthermore, the module modulates the core tensor using feature covariance, making the fusion process focus more on the parts related to second-order features, thereby improving feature representation while reducing computational complexity. This embodiment can focus on capturing the fine-grained structure of the fracture area while also constructing long-range dependencies within the overall skeletal structure. Simultaneously, second-order feature enhancement and saliency guidance improve the perception of key fracture regions. Low-rank fusion and adaptive rank adjustment significantly reduce the computational burden while maintaining feature representation capabilities, enabling the model to have higher inference efficiency and generalization performance in practical applications of medical image analysis.

[0196] In the classification and prediction phase, the model employs a multi-task learning strategy, using both a main classification head and an auxiliary classification head simultaneously:

[0197] (46);

[0198] (47);

[0199] (48);

[0200] (49);

[0201] in, It is the main classification feature vector obtained by correcting the fused features through global average pooling (GAP). and These are the weights and bias parameters of the main classification header. This is the prediction output of the main classification head. Similarly, It is the feature vector of the auxiliary classification head obtained by global average pooling of the local enhanced features in the last feature fusion enhancement module. and These are the weights and bias parameters of the auxiliary classification head. It is the prediction output of the auxiliary classification head. This represents the Global Average Pooling operation. The final loss function is the principal classifier head loss. and auxiliary classification head loss Weighted combination:

[0202] (50);

[0203] in, It is a weighting coefficient (e.g., 0.4) used to balance the contributions of the two classifiers.

[0204] This embodiment introduces a multi-task learning strategy using a main classification head and an auxiliary classification head. Through joint supervision of features at different levels, it enhances the model's feature learning ability and generalization performance. During the training phase, the model simultaneously optimizes both classification heads. The main classification head classifies based on modified fusion features enhanced by complete feature fusion, while the auxiliary classification head directly utilizes locally enhanced features from intermediate processing to provide additional gradient signals. The losses of the two are combined with a weight ratio of 0.4, ensuring that the main classification task receives higher priority while intermediate features are effectively supervised. Through this multi-task learning strategy, the model can simultaneously learn useful information from features at different levels, enhancing its understanding of fracture features and improving classification accuracy and model robustness. This multi-task learning strategy is particularly effective when dealing with medical datasets with limited sample sizes, allowing for more efficient use of limited training data. During the actual inference phase, the model intelligently uses only the prediction results from the main classification head. This "rich training, concise inference" design improves the network's feature learning ability and generalization performance while maintaining efficient and intuitive inference, making it particularly suitable for medical image classification tasks like fracture classification that require simultaneous attention to local details and global structure.

[0205] 4. Learning is achieved by updating network parameters through backpropagation:

[0206] A backpropagation mechanism is employed to iteratively update network parameters, aiming to continuously bring the fracture classification model closer to the optimal solution. Specifically, during network training, the initial learning rate is set to 1e-4, the batch size to 64, and the number of training epochs to 10. In each batch, the model calculates the prediction results through forward propagation and compares them with the true labels to obtain the cross-entropy loss; then, the chain rule is used to solve for the gradient of the loss with respect to the learnable parameters of each layer, and the parameters are updated using the Adam optimizer. After 10 training epochs, the model weights with the best performance on the validation set are selected as the final result, which can effectively improve the accuracy and robustness of fracture classification while maintaining a fast convergence speed. In terms of fracture classification recognition rate, this invention improves by about 11 percentage points compared to the original DenseNet201 (80% accuracy), ultimately reaching a recognition rate of 91%, which is basically close to the human recognition level.

[0207] Fractures are diverse, including transverse fractures, spiral fractures, and compression fractures, each with significantly different appearances in imaging. Existing models often struggle to effectively distinguish between different fracture types when faced with these complex fractures, leading to decreased classification accuracy. This embodiment improves adaptability to different fracture types by combining various feature enhancement techniques, such as channel attention, multi-scale convolution processing, and low-rank fusion. This allows the model to more comprehensively capture the features of different fracture types, improving classification accuracy and model robustness.

[0208] Efficiently integrating feature fusion enhancement modules into existing backbone networks while maintaining training stability and inference efficiency is a significant challenge. This embodiment achieves seamless integration of the feature enhancement module with the backbone network by selectively embedding LCAD modules into the Dense Blocks of DenseNet (by checking the number of features in the block) and combining this with a multi-task learning strategy involving the main and auxiliary classification heads. This improves model performance while maintaining training stability and computational efficiency. This design allows the feature enhancement module to fully utilize the dense connectivity of DenseNet, effectively enhancing feature representation capabilities without significantly increasing model complexity. Compared to simply stacking additional modules or replacing the entire network structure, this embedding strategy retains the knowledge transfer advantages of pre-trained models while enhancing key features, resulting in a design that balances computational efficiency and classification performance.

[0209] This embodiment innovatively combines low-rank fusion, multi-branch local convolution, multi-scale dilated convolution, second-order feature processing, and adaptive complexity adjustment techniques to effectively solve the above-mentioned technical problems and proposes an efficient and accurate fracture image classification method, providing a new technical solution for medical image analysis.

[0210] This invention proposes a deep learning-based fracture classification method, aiming to improve the accuracy and robustness of fracture image classification through innovative feature extraction and fusion mechanisms. Based on the DenseNet201 backbone network, an innovative feature fusion enhancement module is embedded in its DenseBlock, focusing on multi-scale local feature extraction and global information capture. Efficient extraction of fracture features at different scales is achieved through multi-branch local convolution and multi-scale dilated convolution. Tucker decomposition technology is used to decompose high-dimensional feature tensors into core tensors and factor matrices, achieving feature compression and redundancy removal. Simultaneously, covariance calculation and core tensor modulation enhance the second-order feature representation capability, and computational resource allocation is dynamically adjusted according to image complexity. This architectural design allows the model to simultaneously focus on minute local fracture features and the overall skeletal structure when processing fracture images, and can adaptively adjust feature extraction and fusion strategies according to the characteristics of different fracture types, improving classification accuracy. Furthermore, by introducing a multi-task learning mode with an auxiliary classification head and a main classification head, the model's feature learning capability and generalization performance are further enhanced.

[0211] It is understood that this embodiment is not limited to a specific backbone network. Although DenseNet201 was used as an example as the base network in the technical implementation process for preliminary extraction of multi-level features of medical images, this backbone network is replaceable. That is, any deep network with good representation capabilities (such as ResNet, Inception, EfficientNet, etc.) can be combined with the LCDA (Local Context and Dependency Attention) module proposed in this invention while maintaining the core idea, thereby achieving the same or similar effects in local feature enhancement, global dependency capture, and feature compression and fusion. The DenseNet201 network is only an exemplary choice, and this invention is also applicable to any other convolutional neural network or deep network with feature extraction capabilities, all of which are within the protection scope of this invention.

[0212] The fracture image classification technology of this invention can be widely applied to products in fields such as medical imaging diagnosis, surgical navigation and evaluation, telemedicine, and big data education and research. For example, it can be embedded in the software system of digital X-ray machines, CT scanners, or MRI equipment to achieve real-time assisted diagnosis of fracture detection and classification; it can also be integrated into telemedicine platforms or hospital information systems (PACS) to provide cloud-based fracture identification and classification services for remote areas or inter-hospital collaboration; it can also be used in orthopedic surgical planning systems for preoperative image evaluation and intraoperative assistance to improve the accuracy of fracture reduction and internal fixation surgery; and by combining with medical image management software or big data research platforms, it can provide rapid batch annotation and automated analysis support for pathological research and medical teaching, thereby effectively promoting innovation and development in the field of intelligent medical imaging.

[0213] Example 2

[0214] This embodiment provides a fracture image classification system with fused feature enhancement, including:

[0215] The image preprocessing module is configured to acquire fracture medical images and perform image preprocessing.

[0216] The image classification module is configured to classify pre-processed fracture medical images using a pre-trained fracture image classification model, specifically:

[0217] After initial convolution processing, the preprocessed fracture images are processed sequentially through four dense blocks and a feature fusion enhancement module to obtain the final fusion features. Then, after processing through a transition layer and a global pooling layer, the corrected fusion features are obtained. Based on the corrected fusion features, fracture images are classified to obtain the classification results.

[0218] The processing steps for each dense block and feature fusion enhancement module are as follows:

[0219] Primary features are extracted based on dense blocks;

[0220] The primary features are enhanced and then split into local features and preserved features along the channel dimension. The local features are enhanced and then fused with the preserved features to obtain the local enhanced features.

[0221] The local enhancement features are subjected to multi-scale dilated convolution and adaptive saliency guidance processing. The processing results are then fused with the local enhancement features to obtain globally relevant features.

[0222] The feature tensor is obtained by concatenating the local enhancement features and the global attention features. The second-order feature optimization tensor decomposition method is used to optimize the feature tensor. The optimized tensor decomposition method is then used to decompose and reconstruct the feature tensor to obtain the fused features.

[0223] The examples and application scenarios implemented by the above modules and corresponding steps are the same, but are not limited to the content disclosed in Embodiment 1 above. It should be noted that the above modules, as part of the system, can be executed in a computer system such as a set of computer-executable instructions.

[0224] The descriptions of each embodiment in the above embodiments have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0225] The proposed system can be implemented in other ways. For example, the system embodiments described above are merely illustrative, and the division of modules described above is only a logical functional division. In actual implementation, there may be other division methods. For example, multiple modules may be combined or integrated into another system, or some features may be ignored or not executed.

[0226] Example 3

[0227] This embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps in the fracture image classification method with fusion feature enhancement as described in Embodiment 1 above.

[0228] Example 4

[0229] This embodiment provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps in the fracture image classification method with fusion feature enhancement as described in Embodiment 1 above.

[0230] Example 5

[0231] This embodiment provides a computer program product or computer program that includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the steps in the fracture image classification method with fused feature enhancement described in Embodiment 1.

[0232] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of hardware embodiments, software embodiments, or embodiments combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage and optical storage) containing computer-usable program code.

[0233] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A fracture image classification method incorporating enhanced features, characterized in that, include: Acquire fracture medical images and perform image preprocessing; Based on pre-processed fracture medical images, a pre-trained fracture image classification model is used for classification. The fracture image classification model includes a backbone network module, which comprises an initial convolutional layer, four dense blocks, a transition layer, and a global pooling layer. Each dense block is followed by a feature fusion enhancement module, specifically: After initial convolution processing, the preprocessed fracture image is processed sequentially through four dense blocks and a feature fusion enhancement module to obtain the final fusion feature. Then, it is processed through a transition layer and a global pooling layer to obtain the corrected fusion feature. Based on the corrected fusion feature, the fracture image is classified to obtain the classification result. The processing steps for each dense block and feature fusion enhancement module are as follows: Primary features are extracted based on dense blocks; After enhancing the primary features, they are further split along the channel dimension into local features and preserved features. The local features are then enhanced and fused with the preserved features to obtain the locally enhanced features. Multi-scale dilated convolution and adaptive saliency guidance are then applied to the locally enhanced features, and the results are fused with the locally enhanced features to obtain the globally relevant features. Specifically: The local enhancement features are split along the channel dimension to obtain two local enhancement sub-features; After performing small-scale regular convolution and large-scale dilated convolution on the two local augmentation features respectively, local augmentation features of different scales are obtained and then concatenated to obtain the global initial features; By using adaptive saliency guidance, the initial global features are processed separately in terms of channel and space before being concatenated to obtain the global features of interest. The process of using adaptive saliency guidance to process the initial global features separately in terms of channel and spatial dimensions before concatenation yields globally relevant features. Specifically: Global average pooling is used to capture the global context information of each channel of the global initial features. Based on the global context information, the dependencies between all channels are learned through nonlinear transformation to obtain the channel attention weights. Channel-guided features are calculated using channel attention weights and global initial features; The global initial features are downsampled to obtain downsampled features. The spatial covariance matrix of the downsampled features is calculated to reflect the correlation between different channels. The diagonal elements of the spatial covariance matrix based on downsampling features are normalized to obtain spatial significance weights, i.e., channel weights, which are used to highlight channels with high spatial variability. Broadcast multiplication is performed based on spatial saliency weights and downsampling features to obtain spatial modulation features, which highlight regions with high spatial variability, corresponding to fracture locations. After upsampling the spatial modulation features, they are concatenated with local enhancement features and channel-guided features to obtain the global interest features; The feature tensor is obtained by concatenating the local enhancement features and the global attention features. The second-order feature optimization tensor decomposition method is then used to decompose and reconstruct the feature tensor to obtain the fused features.

2. The fracture image classification method with fusion feature enhancement as described in claim 1, characterized in that, The process involves enhancing the primary features, then splitting them along the channel dimension into local features and retained features. The local features are then enhanced and fused with the retained features to obtain the enhanced local features. Specifically: After performing local context enhancement on the primary features, primary extended features are obtained; The primary extended features are split into local features and retained features along the channel dimension according to a set ratio; Multi-branch local convolution is used to divide local features into two parts to obtain local features at different scales, and the local features at different scales are concatenated in the channel dimension to obtain the enhanced local features. The enhanced local features are then concatenated and fused with the preserved features to obtain the enhanced local features.

3. The fracture image classification method with fusion feature enhancement as described in claim 1, characterized in that, A second-order feature optimization tensor decomposition method based on feature tensors is used to decompose and reconstruct the feature tensors to obtain fused features. Specifically: The covariance matrix is ​​calculated based on the second-order features of the feature tensor, and the second-order covariance matrix is ​​obtained. The feature tensor is subjected to adaptive evaluation pooling and flattening to obtain tensor flattened features, and the complexity score of the tensor features is determined based on the tensor flattened features. The adaptive rank is determined based on the complexity score of tensor features, and the optimized tensor decomposition method is obtained from the adaptive rank. The diagonal elements of the second-order covariance matrix are extracted as channel weights, and the channel weights are mapped and reshaped to obtain the modulation coefficient tensor. The core tensor is adjusted using the modulation coefficient tensor, and the modulation core tensor is decomposed based on the optimized tensor decomposition method to obtain the truncated modulation core tensor. The fused features are obtained by reconstructing and fusing using the clipped modulation core tensor.

4. The fracture image classification method with fusion feature enhancement as described in claim 1, characterized in that, The feature fusion enhancement module includes a local feature enhancement submodule and a local-global feature perception and fusion submodule; The local feature enhancement submodule includes convolutional layers, channel splitting layers, multi-branch local convolutional layers, and fusion layers; The local and global feature perception and fusion submodule includes a multi-scale dilated convolutional layer, an adaptive saliency guidance layer, and a low-rank fusion and second-order feature enhancement layer.

5. A fracture image classification system with fusion feature enhancement, characterized in that, include: The image preprocessing module is configured to acquire fracture medical images and perform image preprocessing. The image classification module is configured to classify pre-processed fracture medical images using a pre-trained fracture image classification model. The fracture image classification model includes a backbone network module, which comprises an initial convolutional layer, four dense blocks, a transition layer, and a global pooling layer. Each dense block is followed by a feature fusion enhancement module, specifically: After initial convolution processing, the preprocessed fracture image is processed sequentially through four dense blocks and a feature fusion enhancement module to obtain the final fusion feature. Then, it is processed through a transition layer and a global pooling layer to obtain the corrected fusion feature. Based on the corrected fusion feature, the fracture image is classified to obtain the classification result. The processing steps for each dense block and feature fusion enhancement module are as follows: Primary features are extracted based on dense blocks; After enhancing the primary features, they are further split along the channel dimension into local features and preserved features. The local features are then enhanced and fused with the preserved features to obtain the locally enhanced features. Multi-scale dilated convolution and adaptive saliency guidance are then applied to the locally enhanced features, and the results are fused with the locally enhanced features to obtain the globally relevant features. Specifically: The local enhancement features are split along the channel dimension to obtain two local enhancement sub-features; After performing small-scale regular convolution and large-scale dilated convolution on the two local augmentation features respectively, local augmentation features of different scales are obtained and then concatenated to obtain the global initial features; By using adaptive saliency guidance, the initial global features are processed separately in terms of channel and space before being concatenated to obtain the global features of interest. The process of using adaptive saliency guidance to process the initial global features separately in terms of channel and spatial dimensions before concatenation yields globally relevant features. Specifically: Global average pooling is used to capture the global context information of each channel of the global initial features. Based on the global context information, the dependencies between all channels are learned through nonlinear transformation to obtain the channel attention weights. Channel-guided features are calculated using channel attention weights and global initial features; The global initial features are downsampled to obtain downsampled features. The spatial covariance matrix of the downsampled features is calculated to reflect the correlation between different channels. The diagonal elements of the spatial covariance matrix based on downsampling features are normalized to obtain spatial significance weights, i.e., channel weights, which are used to highlight channels with high spatial variability. Broadcast multiplication is performed based on spatial saliency weights and downsampling features to obtain spatial modulation features, which highlight regions with high spatial variability, corresponding to fracture locations. After upsampling the spatial modulation features, they are concatenated with local enhancement features and channel-guided features to obtain the global interest features; The feature tensor is obtained by concatenating the local enhancement features and the global attention features. The second-order feature optimization tensor decomposition method is then used to decompose and reconstruct the feature tensor to obtain the fused features.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps in the fracture image classification method with fusion feature enhancement as described in any one of claims 1-4.

7. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the fracture image classification method with fusion feature enhancement as described in any one of claims 1-4.

8. A computer program product, characterized in that, The computer program product includes a computer program that, when executed by a processor, implements the steps in a fracture image classification method with fusion feature enhancement as described in any one of claims 1-4.

Images