An intracranial hematoma intelligent identification system based on deep learning
By using a deep learning-based intelligent intracranial hematoma recognition system, which combines contour and semantic information to segment and classify CT images, the system solves the problems of accuracy and real-time performance in existing intracranial hematoma recognition technologies, achieving a more efficient hematoma recognition effect.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- THE THIRD MEDICAL CENT OF THE CHINESE PEOPLES LIBERATION ARMY GENERAL HOSPITAL
- Filing Date
- 2026-05-12
- Publication Date
- 2026-06-09
Smart Images
Figure CN122176428A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of artificial intelligence and medical monitoring integration, and in particular to an intelligent intracranial hematoma recognition system based on deep learning. Background Technology
[0002] Intracranial hematoma is a common and serious complication following traumatic brain injury, and in severe cases, it can even be life-threatening. Therefore, accurate identification of intracranial hematoma is crucial for patient treatment and prognostic assessment. With the continuous development of medical imaging technologies, such as computed tomography (CT), doctors are provided with clear and detailed images of intracranial structures, making the detection of intracranial hematoma possible. However, due to the complexity of intracranial structures and the diversity of hematoma morphology, manual identification of intracranial hematoma is labor-intensive, inefficient, and prone to missed diagnoses and misdiagnoses.
[0003] Compared to manual identification methods, machine learning technology offers advantages such as high automation, strong adaptability, and high efficiency, providing technical support for the accurate identification of intracranial hematomas. Although some machine learning-based methods for intracranial hematoma identification have been proposed, existing methods still suffer from confusion and misjudgment for hematomas with irregular shapes and indistinct boundaries, failing to meet the accuracy and real-time requirements of clinical diagnosis. Therefore, it is necessary to develop an efficient and accurate automatic method for intracranial hematoma identification. Summary of the Invention
[0004] To address the shortcomings of existing technologies, this invention provides an intelligent intracranial hematoma recognition system based on deep learning.
[0005] To achieve the above objectives, this invention provides a deep learning-based intelligent intracranial hematoma identification system. The system includes: a data acquisition module, an image block segmentation module, an intracranial hematoma classification and identification module, a data storage module, and a data visualization module. The data acquisition module acquires brain CT images of a patient and preprocesses them to obtain standardized brain CT images. The image block segmentation module divides the standardized brain CT images into multiple image blocks based on contour and semantic information, thereby achieving preliminary hematoma identification. The intracranial hematoma classification and identification module acquires image block features and then classifies and identifies intracranial hematomas based on the standardized brain CT images. The data storage module stores the brain CT images, the standardized brain CT images, the image blocks, the image block features, and the classification and identification results. The data visualization module outputs and displays the classification and identification results of the intracranial hematoma classification and identification module. This invention improves the accuracy of intracranial hematoma identification by dividing CT images into image blocks based on contour and semantic information and then identifying intracranial hematomas based on image block features.
[0006] Optionally, the image block division module includes an association map generation submodule and a pixel aggregation submodule; the association map generation submodule uses a pixel association recognition model and a contour extraction model to obtain an affinity map; the pixel aggregation submodule aggregates pixels according to the affinity map to obtain the image block.
[0007] Optionally, the pixel association recognition model includes an encoder and a decoder, and the contour extraction model includes a contour feature extractor and a contour generator; The encoder includes four encoding layers, and there is a skip connection between the encoder and the decoder; The decoder is used to output the affinity map; The contour feature extractor includes four contour feature extraction layers, and there is a one-to-one correspondence between the encoding layer and the contour feature extraction layer; The contour generator includes four contour generation layers and one fusion output layer, and the contour feature extraction layer corresponds one-to-one with the contour generation layer.
[0008] Optionally, the encoder initially extracts the semantic features of the standardized brain CT image through a convolutional layer as the input of the first coding layer. The input of the first coding layer also includes the contour features output by the corresponding contour feature extraction layer. The input of the second to fourth coding layers is the semantic features output by the previous coding layer and the contour features output by the corresponding contour feature extraction layer. The coding layer uses a feature fusion network to fuse input features to obtain fused features, then uses a feature enhancement network to further enhance the fused features, and downsamples the enhanced fused features to obtain the output features of the coding layer. The feature fusion network performs channel number adjustment and depthwise separable convolution on the semantic features and the contour features respectively to obtain high-dimensional semantic features and high-dimensional contour features. The high-dimensional semantic features and the high-dimensional contour features are then enhanced by a first feature enhancement sub-network. Subsequently, a cross-attention mechanism is used to fuse the enhanced high-dimensional semantic features and the high-dimensional contour features to obtain an initial fused feature. This initial fused feature is then residually connected to the original input of the feature fusion network to obtain the fused feature. The feature enhancement network performs layer normalization and channel number adjustment on the fused feature, dividing it into a first feature part and a second feature part along the channel dimension. Then, the second feature enhancement sub-network is used to obtain the enhanced feature of the first feature part, and the enhanced feature is fused with the second feature part to obtain the global feature. Finally, the global feature is linearly transformed and residually connected with the fused feature to obtain the enhanced fused feature.
[0009] Optionally, the first feature enhancement subnetwork enhances the high-dimensional semantic features and the high-dimensional contour features through the following steps: The high-dimensional semantic features and the high-dimensional contour features are flattened and concatenated in the length dimension to form a feature sequence. Then, the feature sequence is linearly transformed to obtain a high-dimensional feature sequence. The high-dimensional feature sequence is mapped to time step coefficients, input modulation matrix and output modulation matrix at each position through three parallel linear layers. Initialize the global state transition matrix and the hidden states of the high-dimensional feature sequence; Discretize the input modulation matrix and the global state transition matrix to obtain a discretized input modulation matrix and a discretized state transition matrix; The discretized input modulation matrix, the discretized state transition matrix, and the hidden state are used to map the feature sequence into a hidden space representation; The positive enhancement feature sequence is obtained using the output modulation matrix and the hidden space representation; Obtain the inverse sequence of the feature sequence, and obtain the inverse enhancement feature sequence corresponding to the inverse sequence; The elements in the reverse enhancement feature sequence are arranged in reverse order and then fused with the forward enhancement feature sequence to obtain the enhancement feature sequence. The enhanced feature sequence is used to reconstruct the high-dimensional semantic features and the high-dimensional contour features to obtain the enhanced high-dimensional semantic features and the high-dimensional contour features.
[0010] Optionally, the contour feature extractor initially extracts the contour features of the standardized brain CT image through a convolutional layer and uses it as the input of the first contour feature extraction layer. The input of the second to fourth contour feature extraction layers is the output of the previous contour feature extraction layer. The input to the contour generation layer is the output of the corresponding contour feature extraction layer, and the output of each contour generation layer is simultaneously used as the input to the fusion output layer to obtain the contour of the standardized brain CT image.
[0011] Optionally, the intracranial hematoma classification and recognition module includes a feature extraction submodule and a classification and recognition submodule; the feature extraction submodule is used to acquire image block features; the classification and recognition submodule includes a classification and recognition model, and the classification and recognition submodule uses the classification and recognition model to classify and recognize intracranial hematomas based on the image block features and the standardized brain CT images.
[0012] Optionally, the feature extraction submodule performs the following steps: The image patch features are obtained, including statistical features, texture features, and spatial features; The normalized image block features are mapped to the pixel locations of the standardized brain CT image to obtain a classification and recognition feature map.
[0013] Optionally, the statistical features include at least the average CT value, the standard deviation of the CT value, the maximum CT value, the minimum CT value, and the entropy of the image block; the texture features include at least the contrast, correlation, and energy of the gray-level co-occurrence matrix; and the spatial features include at least the area and shape factor of the image block.
[0014] Optionally, the classification and recognition model performs the following steps: The classification and recognition model fuses the classification and recognition feature map with the standardized brain CT image to obtain the classification and recognition fused feature; The classification and recognition model achieves classification and recognition of intracranial hematomas based on the classification and recognition fusion features.
[0015] In summary, the present invention has at least the following beneficial effects: 1. This invention divides standardized brain CT images into multiple image blocks based on contour information and semantic information. This contour-constrained image block division method can capture feature information in the image more meticulously, which helps to more accurately identify the location, shape and other details of the hematoma, providing richer and more accurate feature basis for subsequent classification and recognition, and improving the accuracy of intracranial hematoma identification.
[0016] 2. Based on the acquisition of image blocks, this invention generates classification and recognition feature maps by acquiring the statistical, texture, and spatial features of the image blocks, providing more comprehensive information for the classification and recognition model. This helps the model better understand and distinguish different types of intracranial hematomas and improves the performance of classification and recognition.
[0017] 3. This invention integrates classification and recognition feature maps with standardized brain CT images to identify hematomas. This allows the classification and recognition model to utilize abstract features while referencing the original spatial layout information, thereby more accurately locating and identifying hematomas. It can also, to some extent, avoid the problem that the classification and recognition model relies too much on these specifically extracted features, which limits its generalization ability when encountering new data that differs significantly from the training data. Attached Figure Description
[0018] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.
[0019] Figure 1 This is a schematic diagram of the framework of an intelligent intracranial hematoma recognition system based on deep learning, according to an embodiment of the present invention. Figure 2 This is a flowchart illustrating the workflow of a deep learning-based intelligent intracranial hematoma recognition system according to an embodiment of the present invention. Detailed Implementation
[0020] Specific embodiments of the present invention will now be described in detail. It should be noted that the embodiments described herein are for illustrative purposes only and are not intended to limit the invention. In the following description, numerous specific details are set forth in order to provide a thorough understanding of the invention. However, it will be apparent to those skilled in the art that these specific details are not necessary to practice the invention. In other instances, well-known circuits, software, or methods have not been specifically described to avoid obscuring the invention.
[0021] Throughout this specification, references to "an embodiment," "an embodiment," "an example," or "an example" mean that a particular feature, structure, or characteristic described in connection with that embodiment or example is included in at least one embodiment of the invention. Therefore, the phrases "in an embodiment," "in an embodiment," "an example," or "an example" appearing in various places throughout the specification do not necessarily refer to the same embodiment or example. Furthermore, specific features, structures, or characteristics can be combined in one or more embodiments or examples in any suitable combination and / or sub-combination. Moreover, those skilled in the art will understand that the illustrations provided herein are for illustrative purposes and are not necessarily drawn to scale.
[0022] It should be noted in advance that, in one alternative embodiment, except for independent descriptions, the same symbols or letters appearing in all formulas have the same meaning.
[0023] In one optional embodiment, please refer to Figure 1 This invention provides a deep learning-based intelligent identification system for intracranial hematomas, the system comprising: The system comprises a data acquisition module, an image block segmentation module, an intracranial hematoma classification and recognition module, a data storage module, and a data visualization module. The data acquisition module acquires and preprocesses the patient's brain CT images to obtain standardized brain CT images. The image block segmentation module divides the standardized brain CT images into multiple image blocks based on contour and semantic information, thereby achieving preliminary hematoma identification. The intracranial hematoma classification and recognition module acquires image block features and then classifies and identifies intracranial hematomas in conjunction with the standardized brain CT images. The data storage module stores brain CT images, standardized brain CT images, image blocks, image block features, and classification and recognition results. The data visualization module outputs and displays the classification and recognition results from the intracranial hematoma classification and recognition module.
[0024] Specifically, in this embodiment, for the communication connection between modules that need to interact with data, please refer to [link to documentation]. Figure 2 The overall workflow of the system is as follows: 01. The data acquisition module acquires the patient's brain CT images and preprocesses them to obtain standardized brain CT images. Then, the standardized brain CT images are transmitted to the image block division module and the intracranial hematoma classification and recognition module.
[0025] 02. The image block segmentation module receives standardized brain CT images and divides them into multiple image blocks, and transmits the image block segmentation results to the intracranial hematoma classification and recognition module.
[0026] 03. The intracranial hematoma classification and recognition module receives standardized brain CT images and image block division results, and obtains classification and recognition feature maps based on the received image block division results.
[0027] 04. The intracranial hematoma classification and identification module uses classification and identification feature maps and standardized brain CT images to classify and identify intracranial hematomas, obtain the classification and identification results, and transmit them to the data visualization module.
[0028] 05. The data visualization module receives the classification and recognition results and outputs and displays them.
[0029] While the above process is running, the data storage module is responsible for storing all types of data generated throughout the process, including data used to train the pixel association recognition model, contour extraction model, and classification recognition model. The data visualization module can output the classification recognition results via a digital display screen, and relevant personnel can retrieve and output other data from the data storage module according to actual needs.
[0030] Specifically, in this embodiment, the data acquisition module establishes a secure connection with the medical institution's image archiving and communication system or local storage server. First, it directly reads the archived DICOM format brain CT images from the storage device and resamples them to a preset standard size (512×512) using a bilinear interpolation algorithm. During resampling, the aspect ratio is maintained and edge areas are filled (e.g., a black background) to ensure that the spatial proportions of the cranial structure are not affected by distortion. Then, based on the window width and window level in the DICOM metadata, the data acquisition module linearly maps the pixel values of the brain CT images to the [0, 255] interval. The brain CT images are then converted to lossless PNG format, and the pixel values are normalized to the [0, 1] interval, ultimately obtaining a standardized brain CT image. This process of acquiring standardized brain CT images by the data acquisition module is based on existing technology and will not be described in detail here.
[0031] Specifically, in this embodiment, the image block division module includes an affinity map generation submodule and a pixel aggregation submodule. The affinity map generation submodule uses a pixel affinity recognition model and a contour extraction model to obtain an affinity map, and the pixel aggregation submodule aggregates pixels according to the affinity map, thereby dividing the standardized brain CT image into multiple image blocks.
[0032] Specifically, in this embodiment, the pixel association recognition model includes an encoder and a decoder, and the contour extraction model includes a contour feature extractor and a contour generator. The encoder includes four encoding layers (or encoding blocks), and the decoder includes four decoding layers (or decoding blocks) and one softmax layer, with skip connections between the encoder and decoder. The contour feature extractor includes four contour feature extraction layers, with a one-to-one correspondence between the encoding layers and the contour feature extraction layers. The contour generator includes four contour generation layers and one fusion output layer, with a one-to-one correspondence between the contour feature extraction layers and the contour generation layers.
[0033] Specifically, in this embodiment, the encoder first extracts semantic features from standardized brain CT images using a convolutional layer with a kernel size of 3×3 and a stride and padding value of 1. , 64 represents the number of channels. Simultaneously, the first contour feature extraction layer is used to extract contour features from standardized brain CT images. , and contour features With semantic features Simultaneously, it serves as the input to the first encoding layer. The inputs to the second through fourth encoding layers are the semantic features output from the previous encoding layer and the contour features output from the corresponding contour feature extraction layer. Specifically, the input to the second encoding layer is the semantic features output from the first encoding layer. Contour features extracted by the second contour feature extraction layer The input to the third encoding layer is the semantic features output from the second encoding layer. Contour features extracted by the third contour feature extraction layer The input to the fourth encoding layer is the semantic features output from the third encoding layer. Contour features extracted by the fourth contour feature extraction layer , , , , , , .
[0034] In the encoder, each coding layer fuses semantic features and contour features through a feature fusion network to obtain fused features. Then, a feature enhancement network is used to further enhance the fused features, and a convolutional layer with a kernel size of 3×3, a stride of 2, and a padding value of 1 is used to downsample the enhanced fused features, finally obtaining the output features of the coding layer.
[0035] In the feature fusion network, the number of channels of the semantic features is first doubled (denoted as C) through a 1×1 convolution to obtain the feature map. Then feature map A depthwise separable convolution with a kernel size of 3×3, stride, and padding of 1 is input to extract high-dimensional semantic features. Simultaneously, high-dimensional contour features are obtained using the same method. After obtaining the high-dimensional semantic and contour features, the feature fusion network further enhances them through a first feature enhancement subnetwork, resulting in enhanced high-dimensional semantic and contour features. The first feature enhancement subnetwork enhances the high-dimensional semantic and contour features through the following steps: S1. Flatten the high-dimensional semantic features and the high-dimensional contour features and concatenate them in the length dimension to form a feature sequence. Then, perform a linear transformation on the feature sequence to obtain a high-dimensional feature sequence.
[0036] First, the high-dimensional semantic features are flattened from left to right into a high-dimensional semantic feature sequence, and the high-dimensional contour features are also flattened from left to right into a high-dimensional contour feature sequence. The lengths of the high-dimensional semantic feature sequence and the high-dimensional contour feature sequence are W×H (concatenated row by row, where W is the number of pixel columns and H is the number of pixel rows). During the flattening process, the channel dimension needs to be preserved, meaning that each position in the sequence includes C features. Then, the high-dimensional semantic feature sequence and the high-dimensional contour feature sequence are concatenated along the length dimension to obtain a concatenated feature sequence, which is denoted as the feature sequence in this embodiment. Finally, a fully connected layer is used to perform a linear transformation on the obtained feature sequence, completing the dimensionality increase of the feature sequence in the channel dimension, specifically doubling the number of channels C to 2C, resulting in the high-dimensional feature sequence.
[0037] S2. The high-dimensional feature sequence is mapped to the time step coefficient, input modulation matrix and output modulation matrix at each position through three parallel linear layers.
[0038] After obtaining the high-dimensional feature sequence in step S1, it is input into three parallel linear layers. The first linear layer maps the features of the high-dimensional feature sequence back to C dimensions and outputs the time step coefficients at each position in the feature sequence using the softplus function. The second linear layer maps the features of the high-dimensional feature sequence to the state dimension N (N=16) and outputs the input modulation matrix at each position in the feature sequence. , The third linear layer also maps the high-dimensional feature sequence to the state dimension and outputs the modulation matrix at each position in the feature sequence. , It should be noted that both the input modulation matrix and the output modulation matrix are trainable.
[0039] S3. Initialize the global state transition matrix and the hidden state of the high-dimensional feature sequence.
[0040] The global state transition matrix Z is initialized using the HiPPO initialization method. Specifically, the HiPPO initialization is performed using the Legendre polynomial basis (a state matrix initialization method based on the theory of higher-order polynomial projection operators), while simultaneously initializing the hidden states at each position of the high-dimensional feature sequence. , and It should be noted that the global state transition matrix is trainable.
[0041] S4. Discretize the input modulation matrix and the global state transition matrix to obtain a discretized input modulation matrix and a discretized state transition matrix.
[0042] Given the global state transition matrix, the time step coefficients at each position in the feature sequence, and the input modulation matrix, the zero-order preserved discretization method can be used to obtain the discretized input modulation matrix and the discretized state transition matrix at each position in the feature sequence. The discretized input modulation matrix and the discretized state transition matrix satisfy the following relationships: in, Let be the discretized input modulation matrix at the i-th position in the feature sequence. The time step coefficient at the i-th position in the feature sequence. Let I be the input modulation matrix at the i-th position in the feature sequence, and let I be the identity matrix. Let be the discretized state transition matrix at the i-th position in the feature sequence, and t be a continuous time variable.
[0043] S5. The feature sequence is mapped to a hidden space representation using the discretized input modulation matrix, the discretized state transition matrix, and the hidden state.
[0044] After obtaining the discretized input modulation matrix and the discretized state transition matrix, the feature sequence can be mapped to a hidden space representation using the following relationship: in, Let be the hidden space representation of the i-th position in the feature sequence. , This represents the hidden state at position i-1 in the feature sequence. , Let i be the feature vector at the i-th position in the feature sequence. By obtaining the hidden space representation of each position in the feature sequence, the feature sequence can be mapped to the hidden space representation.
[0045] S6. Obtain the positive enhancement feature sequence using the output modulation matrix and the hidden space representation.
[0046] After obtaining the hidden space representation of each position in the feature sequence, the product of the output modulation matrix and the hidden space representation of each position is used as the enhancement feature of that position. By obtaining the enhancement features of each position in the feature sequence, the backup positive enhancement feature sequence can be obtained. Then, a linear layer is used to map the features of the backup positive enhancement feature sequence back to the C dimension, and finally the positive enhancement feature sequence is obtained.
[0047] S7. Obtain the inverse sequence of the feature sequence, and obtain the inverse enhancement feature sequence corresponding to the inverse sequence.
[0048] The elements in the feature sequence are reversed to obtain its inverse sequence. Then, the inverse enhancement feature sequence of this inverse sequence is obtained according to steps S1 to S6. It should be noted that the forward enhancement feature sequence and the inverse enhancement feature sequence are obtained simultaneously. In addition, the inversion only affects the sequence length dimension and does not change the channel dimension.
[0049] S8. After reversing the order of the elements in the reverse enhancement feature sequence, merge it with the positive enhancement feature sequence to obtain the enhancement feature sequence.
[0050] The elements in the inversely enhanced feature sequence are reversed and added to the forward enhanced feature sequence to mitigate the loss of spatial structural information caused by flattening the feature map, resulting in an enhanced feature sequence. In this step, the reversal only affects the sequence length dimension and does not change the channel dimension.
[0051] S9. Reconstruct the high-dimensional semantic features and the high-dimensional contour features using the enhanced feature sequence to obtain the enhanced high-dimensional semantic features and the high-dimensional contour features.
[0052] Because the feature sequence is obtained by flattening and concatenating high-dimensional semantic features and high-dimensional contour features along the length dimension, after obtaining the enhanced feature sequence, it can be split into two sequences of degree H×W. The first H×W elements of the enhanced feature sequence form the first sequence, which can be reshaped into high-dimensional semantic features, i.e., the enhanced high-dimensional semantic features; the last H×W elements of the enhanced feature sequence form the second sequence, which can be reshaped into high-dimensional contour features, i.e., the enhanced high-dimensional contour features.
[0053] After obtaining the enhanced high-dimensional semantic features and high-dimensional contour features, a cross-attention mechanism is used to fuse the enhanced high-dimensional semantic features and high-dimensional contour features to obtain the initial fused features. This process can be represented by the following relation: in, For the initial fusion features, Concat represents concatenation along the channel dimension. For average pooling, This is the feature map obtained by performing the first linear transformation on the semantic features in the feature fusion network. This is the feature map obtained by performing the first linear transformation on the contour features in the feature fusion network. The acquisition method and The acquisition method is the same. Multiply corresponding elements. For enhanced high-dimensional contour features, For enhanced high-dimensional semantic features, LN is used for layer normalization. It is a 1×1 convolution. The step size and padding value are both 1, which is used to adjust the number of channels in the feature map and fuse channel information.
[0054] The initial inputs to the feature fusion network are semantic features and contour features. After obtaining the initial fused features, a residual connection is made between these features and the initial inputs to the feature fusion network to obtain the final fused features.
[0055] Furthermore, the fused features obtained by the feature fusion network are input into the feature enhancement network. The feature enhancement network first performs layer normalization on the fused features to obtain feature maps. And use a 1×1 convolution to The number of channels is doubled, resulting in a feature map. Then, the feature map The channel is evenly divided along the channel dimension into a first feature part (the first half of the channel) and a second feature part (the second half of the channel). A second feature enhancement sub-network is then used to obtain the enhanced features of the first feature part. The first feature part can also be referred to as the first feature map, and the second feature part as the second feature map. The second feature enhancement sub-network obtains the enhanced features of the first feature part through the following steps: (1) Flatten the pixels in the first feature map in two directions, from left to right and from top to bottom, to obtain two feature sequences. The first feature sequence obtained by flattening from left to right has a length of W×H, and the second feature sequence obtained by flattening from top to bottom has a length of H×W (column-by-column splicing). The channel dimension needs to be preserved during the flattening process.
[0056] (2) The number of channels of the first feature sequence and the second feature sequence is doubled by a fully connected layer to obtain the first high-dimensional feature sequence and the second high-dimensional feature sequence.
[0057] (3) Obtain the first enhanced feature sequence and the second enhanced feature sequence of the first high-dimensional feature sequence and the second high-dimensional feature sequence, respectively. For the specific process, please refer to steps S2 to S6, which will not be repeated here.
[0058] (4) Obtain the inverse sequences of the first feature sequence and the second feature sequence, and denote them as the first inverse feature sequence and the second inverse feature sequence, respectively. Then, obtain the first enhanced inverse feature sequence and the second enhanced inverse feature sequence of the first inverse feature sequence and the second inverse feature sequence, respectively. The process of obtaining them is not described in detail here. It should be noted that obtaining the first enhanced inverse feature sequence and the second enhanced inverse feature sequence is performed simultaneously with obtaining the first enhanced feature sequence and the second enhanced feature sequence.
[0059] (5) Reconstruct the first feature map using the first enhanced feature sequence, the second enhanced feature sequence, the first enhanced inverse feature sequence, and the second enhanced inverse feature sequence according to the order of pixel flattening. Specifically, the first enhanced feature sequence is reconstructed into a feature map of size H×W in a flattening order from left to right; the second enhanced feature sequence is reconstructed into a feature map of size H×W in a flattening order from top to bottom; the first enhanced inverse feature sequence is reversed and then reconstructed into a feature map of size H×W in a flattening order from left to right; and the second enhanced inverse feature sequence is reversed and then reconstructed into a feature map of size H×W in a flattening order from top to bottom. This results in four reconstructed first feature maps.
[0060] (6) The four reconstructed first feature maps are concatenated along the channel dimension to obtain a concatenated feature map. A 1×1 convolution is then used to reduce the dimensionality of the concatenated feature map so that its number of channels is equal to that of the feature map. The same characteristics are obtained, resulting in enhanced features.
[0061] A 1×1 convolution is used to increase the dimensionality of the second feature map, making its channel number the same as the enhanced feature. After layer normalization of the increased second feature map and the enhanced feature, corresponding element-wise multiplication is performed to fuse the second feature part with the enhanced feature, resulting in the global feature. Further, a 1×1 convolution is used to increase the dimensionality of the fused feature obtained by the feature fusion network, making its channel number the same as the global feature. Then, a residual connection is performed between the global feature and the increased fused feature to finally obtain the enhanced fused feature.
[0062] The enhanced fused features are downsampled to halve their size and used as semantic features output by the encoding layer.
[0063] The decoder is specifically a convolutional decoder with four convolutional decoding layers. The last convolutional decoding layer is followed by a softmax layer, which is used to output the affinity map P. The nine channels of the affinity map represent the association probability between a pixel and pixels in its eight adjacent directions.
[0064] Furthermore, for any pixel O on the affinity map, if the correlation probability between O and pixel Q is the highest among its neighboring pixels, then pixel O should be grouped together with pixel Q. Based on this operation, the pixel aggregation submodule traverses every pixel on the affinity map, thus dividing the standardized brain CT image into multiple image blocks.
[0065] Specifically, in this embodiment, the contour feature extractor first extracts the contour features of the standardized brain CT image through a convolutional layer with a kernel size of 3×3 and a stride and padding value of 1. , This is used as the input to the first contour feature extraction layer, and the input to the second to fourth contour feature extraction layers is the output of the previous contour feature extraction layer.
[0066] The contour feature extraction layer first extracts feature maps step by step through four cascaded feature extraction sub-networks, and then performs max pooling with a kernel size of 2×2 and a stride of 2 on the extracted feature maps to obtain the corresponding contour features. The feature extraction sub-networks can be represented by the following formula: in, Let F be the feature map extracted by the feature extraction subnetwork, and let F be the input feature map of the feature extraction subnetwork. This is a convolution with a 1×1 kernel and no padding, and ReLU is the activation function. It is a depthwise separable convolution with a kernel size of 3×3 and a stride and padding value of 1.
[0067] Furthermore, the input to the contour generation layer is the output of the corresponding contour feature extraction layer; that is, the output of the first contour feature extraction layer serves as the input to the first contour generation layer, the output of the second contour feature extraction layer serves as the input to the second contour generation layer, and so on. The contour generation layer can be represented using the following formula: Where E represents the CT image contour output by the contour generation layer, and Up represents upsampling (using bilinear interpolation). The contour features are represented by sigmoid, which is the activation function.
[0068] The outputs of each contour generation layer are simultaneously used as input to the fusion output layer, which outputs a standardized contour of the brain CT image. The fusion output layer can be represented by the following formula: in, To standardize the outline of brain CT images, , , and These are the CT image contours output from the first, second, third, and fourth contour generation layers, respectively.
[0069] Specifically, in this embodiment, the pixel association recognition model and the contour extraction model are trained together as a whole. In fact, the pixel association recognition model and the contour extraction model belong to the same deep learning model, whose loss function includes image segmentation loss and contour extraction loss; the total loss is the sum of the image segmentation loss and the contour extraction loss. This involves obtaining the true contour of a standardized brain CT image and the corresponding CT image contour. Subsequently, the Sorenson-Dies coefficient loss function was used to calculate the contour extraction loss. Including the contour extraction loss as part of the total loss allows the affinity map to be constrained by the contour extraction loss during training, which significantly reduces the association probability learned by the model at the contour boundaries, thus naturally forming image patches with contours as boundaries during aggregation.
[0070] Furthermore, pixels in standardized brain CT images are classified according to intracranial tissue type, and one-hot encoding is used to represent different pixel categories. Simultaneously, the coordinate positions of the pixels in the standardized brain CT images are determined. After obtaining the affinity map and identifying the image blocks, the pixel categories and coordinate positions in the standardized brain CT images can be reconstructed.
[0071] First, the probability of a pixel belonging to different candidate image blocks is determined based on the association probability between a pixel and its eight neighboring pixels. A candidate image block refers to an image block adjacent to a pixel. If a pixel is located within an image block, then the image block adjacent to it is the image block it belongs to. For ease of understanding, the following example illustrates this: For any pixel O, assuming that each of its neighboring pixels is located in a different image block, the probability that pixel O belongs to these image blocks is its association probability with its neighboring pixels. For any pixel O, assuming that it is located in the same image block K with n (n is not less than 2 and not greater than 8) neighboring pixels, the probability that pixel O belongs to image block K is equal to the maximum value of its association probability with its n neighboring pixels.
[0072] After determining the probability of a pixel belonging to different candidate image blocks, it is denoted as the image block assignment probability for ease of writing. For any image block K, a sum-normalization method is used to normalize the image block assignment probability of each pixel within it. Specifically, the ratio of each pixel's image block assignment probability to the total sum of its image block assignment probabilities is calculated and used as the category weight for that pixel. For any image block K, based on the determined pixel categories and the calculated category weights, a weighted sum of pixel categories is calculated, and the result is used as the image block category of image block K.
[0073] Finally, for any image block K, the class of pixels in the standardized brain CT image can be reconstructed using its image block class and the image block assignment probability of pixels in image block K. This reconstructed class is denoted as the reconstructed class. Specifically, the product of the image block class and the image block assignment probability of the pixel is used as the reconstructed class. The closer the reconstructed class is to the true class of the pixel, the more accurate the image block division.
[0074] Furthermore, the coordinate positions of pixels in standardized brain CT images can be reconstructed using the same method as for reconstructing pixel categories, and these reconstructed coordinate positions are recorded as reconstructed coordinates. The closer the reconstructed coordinates are to the actual coordinates of the pixels, the more spatially concentrated the pixels are in the image block.
[0075] Finally, the cross-entropy loss between the reconstructed class and the ground truth class of each pixel is calculated, along with the L2 distance between the reconstructed coordinates and the ground truth coordinates of each pixel. The sum of the cross-entropy losses for each pixel is used as the pixel class loss, and the sum of the L2 distances for each pixel is used as the pixel position loss. Then, the sum of the pixel class loss and pixel position loss is calculated and used as the image segmentation loss. This image segmentation loss transforms the supervision of pairwise association probabilities into supervision of pixel-level class and coordinates, enabling the model to indirectly achieve accurate image patch segmentation through the optimization of association probabilities without directly outputting pixel class. The process of obtaining the image segmentation loss is actually a kind of "soft allocation" expectation-maximization optimization: in the expectation step, the association relationship between pixels and image patches is determined based on the current association probabilities; in the maximization step, the image patch class and centroid coordinates are aggregated based on the association relationship, and then the association probabilities are updated through backpropagation of reconstruction error. This design allows the model to indirectly learn reasonable image patch segmentation using pixel-level class and coordinate annotations without direct pixel-level segmentation annotations.
[0076] Multiple standardized brain CT images were acquired and pixel-level annotated (pixel category), along with their true contours. An intracranial CT image set was constructed using the pixel-level annotated standardized brain CT images and their corresponding true contours, and divided into training and validation sets in a 7:3 ratio to train and validate the pixel association recognition model and the contour extraction model. Specifically, AdamW was used to train the pixel association recognition model, with an initial learning rate of 0.0001, weight decay of 0.01, batch size of 24, and a training duration of 200 epochs.
[0077] Specifically, in this embodiment, the intracranial hematoma classification and recognition module includes a feature extraction submodule and a classification and recognition submodule. The feature extraction submodule is used to acquire image patch features. The classification and recognition submodule includes a classification and recognition model, which classifies and identifies intracranial hematomas based on the image patch features and standardized brain CT images.
[0078] The feature extraction submodule performs the following steps: T1. Obtain the image block features, which include statistical features, texture features, and spatial features.
[0079] Statistical features include at least the average CT value, standard deviation of CT value, maximum CT value, minimum CT value, and entropy of the image patch. Texture features include at least the contrast, correlation, and energy of the gray-level co-occurrence matrix. Spatial features include at least the area and shape factor of the image patch. The area of the image patch can be represented by the number of pixels, and the shape factor is the ratio of the area to the square of the perimeter of the image patch. The specific calculation methods for other statistical, texture, and spatial features are existing techniques and will not be detailed here.
[0080] T2. Map the normalized image block features to the pixel locations of the standardized brain CT image to obtain the classification and recognition feature map.
[0081] First, the features of each image block are normalized using the minimum-maximum normalization method. For each image block, its normalized feature value is assigned to all pixels within that block, forming a pixel-level feature map, i.e., a classification and recognition feature map. This classification and recognition feature map has 10 channels, corresponding to 10 types of image block features.
[0082] The classification and recognition model includes a feature fusion part and a classification and recognition part. The classification and recognition model performs the following steps: U1. The classification and recognition model fuses the classification and recognition feature map with the standardized brain CT image to obtain the classification and recognition fused feature.
[0083] The classification and recognition model stitches the classification and recognition feature map with the standardized brain CT image in the channel dimension through the feature fusion part to obtain the classification and recognition fused feature.
[0084] U2. The classification and recognition model achieves classification and recognition of intracranial hematomas based on the classification and recognition fusion features.
[0085] The classification and recognition part of the model is a multi-task ResNet50 network, whose input is the classification and recognition fusion features. After global average pooling, this network connects two parallel fully connected layers as classification heads: the first classification head outputs the hematoma location classification result, dividing the hematoma location into multiple brain lobe regions (such as frontal lobe, temporal lobe, parietal lobe, occipital lobe, cerebellum, brainstem, etc.) and outputting the corresponding probability distribution; the second classification head outputs the hematoma formation time classification result, including four categories: hyperacute phase, acute phase, subacute phase, and chronic phase. The two classification heads share the backbone network parameters of ResNet50.
[0086] To train, validate, and test the classification and recognition model, a corresponding dataset needs to be constructed. This involves labeling the location and time of hematoma formation in a large number of standardized brain CT images (hematoma formation time classification includes hyperacute, acute, subacute, and chronic phases). The dataset is then constructed by combining this with corresponding classification and recognition feature maps, and divided into training, validation, and test sets in a 7:2:1 ratio to train, validate, and test the classification and recognition model. When labeling the location and time of hematoma formation in standardized brain CT images, integer encoding is used for the labels, such as hyperacute phase = 1. Labels are independently labeled by multiple radiologists and then cross-validated to ensure label accuracy.
[0087] Since the classification and recognition model needs to perform multi-class tasks, the loss function adopts weighted cross-entropy loss, that is, the cross-entropy loss of hematoma location and category are calculated separately, and then summed according to the weights (the initial weights can be set to 1, and then fine-tuned on the validation set). In addition, this embodiment adds a Dropout layer before each classification head, and sets the Dropout probability to 0.3 to reduce overfitting.
[0088] When training the classification and recognition model, the AdamW optimizer was used with an initial learning rate of 0.0003, and a cosine annealing learning rate adjustment strategy was employed. The training epochs were set to 100 epochs, and an early stopping mechanism was used: training was stopped early when the loss value of the classification and recognition model on the validation set no longer significantly decreased for 10 consecutive epochs (the absolute value of the loss value between two adjacent epochs did not exceed 0.00001) to prevent overfitting. The F1 score and accuracy were used as evaluation metrics when testing the performance of the classification and recognition model.
[0089] It should be noted that in some cases, the actions described in the specification can be performed in different orders and still achieve the desired results. In this embodiment, the order of steps is given only to make the embodiment clearer and easier to explain, and not to limit it.
[0090] In summary, this application has at least the following beneficial effects: This application divides standardized brain CT images into multiple image blocks based on contour and semantic information. This contour-constrained image block division method can capture feature information in the images more meticulously, helping to more accurately identify the location, shape, and other details of hematomas, providing richer and more accurate feature basis for subsequent classification and recognition, and improving the accuracy of intracranial hematoma identification. Based on the obtained image blocks, this application generates classification and recognition feature maps by acquiring the statistical, texture, and spatial features of the image blocks, providing more comprehensive information for the classification and recognition model. This helps the model better understand and distinguish different types of intracranial hematomas, improving classification and recognition performance. This application integrates the classification and recognition feature maps with standardized brain CT images to identify hematomas, allowing the classification and recognition model to utilize abstract features while referencing the original spatial layout information, thereby more accurately locating and identifying hematomas. It also avoids, to some extent, the problem of the classification and recognition model relying too heavily on these specifically extracted features, which limits its generalization ability when encountering new data that differs significantly from the training data.
[0091] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be covered within the scope of the claims and specification of the present invention.
Claims
1. A deep learning-based intelligent identification system for intracranial hematoma, characterized in that, include: The system includes a data acquisition module, an image block division module, an intracranial hematoma classification and recognition module, a data storage module, and a data visualization module. The data acquisition module is used to acquire the patient's brain CT images and preprocess the brain CT images to obtain the patient's standardized brain CT images. The image block division module is used to divide the standardized brain CT image into multiple image blocks based on contour information and semantic information, thereby achieving preliminary hematoma identification. The intracranial hematoma classification and recognition module is used to acquire image block features, and then combine the standardized brain CT images to classify and recognize intracranial hematomas. The data storage module is used to store the brain CT images, the standardized brain CT images, the image blocks, the image block features, and the classification and recognition results; The data visualization module is used to output and display the classification and recognition results of the intracranial hematoma classification and recognition module.
2. The intelligent intracranial hematoma recognition system based on deep learning according to claim 1, characterized in that: The image block segmentation module includes an association map generation submodule and a pixel aggregation submodule; the association map generation submodule uses a pixel association recognition model and a contour extraction model to obtain an affinity map; the pixel aggregation submodule aggregates pixels according to the affinity map to obtain the image block.
3. The intelligent intracranial hematoma recognition system based on deep learning according to claim 2, characterized in that: The pixel association recognition model includes an encoder and a decoder, and the contour extraction model includes a contour feature extractor and a contour generator. The encoder includes four encoding layers, and there is a skip connection between the encoder and the decoder; The decoder is used to output the affinity map; The contour feature extractor includes four contour feature extraction layers, and there is a one-to-one correspondence between the encoding layer and the contour feature extraction layer; The contour generator includes four contour generation layers and one fusion output layer, and the contour feature extraction layer corresponds one-to-one with the contour generation layer.
4. The intelligent intracranial hematoma recognition system based on deep learning according to claim 3, characterized in that: The encoder initially extracts semantic features from the standardized brain CT image through a convolutional layer, which serves as the input to the first coding layer. The input to the first coding layer also includes the contour features output by the corresponding contour feature extraction layer. The inputs to the second to fourth coding layers are the semantic features output by the previous coding layer and the contour features output by the corresponding contour feature extraction layer. The coding layer uses a feature fusion network to fuse input features to obtain fused features, then uses a feature enhancement network to further enhance the fused features, and downsamples the enhanced fused features to obtain the output features of the coding layer. The feature fusion network performs channel number adjustment and depthwise separable convolution on the semantic features and the contour features respectively to obtain high-dimensional semantic features and high-dimensional contour features. The high-dimensional semantic features and the high-dimensional contour features are then enhanced by a first feature enhancement sub-network. Subsequently, a cross-attention mechanism is used to fuse the enhanced high-dimensional semantic features and the high-dimensional contour features to obtain an initial fused feature. This initial fused feature is then residually connected to the original input of the feature fusion network to obtain the fused feature. The feature enhancement network performs layer normalization and channel number adjustment on the fused feature, dividing it into a first feature part and a second feature part along the channel dimension. Then, the second feature enhancement sub-network is used to obtain the enhanced feature of the first feature part, and the enhanced feature is fused with the second feature part to obtain the global feature. Finally, the global feature is linearly transformed and residually connected with the fused feature to obtain the enhanced fused feature.
5. The intelligent intracranial hematoma recognition system based on deep learning according to claim 4, characterized in that, The first feature enhancement subnetwork enhances the high-dimensional semantic features and the high-dimensional contour features through the following steps: The high-dimensional semantic features and the high-dimensional contour features are flattened and concatenated in the length dimension to form a feature sequence. Then, the feature sequence is linearly transformed to obtain a high-dimensional feature sequence. The high-dimensional feature sequence is mapped to time step coefficients, input modulation matrix and output modulation matrix at each position through three parallel linear layers. Initialize the global state transition matrix and the hidden states of the high-dimensional feature sequence; Discretize the input modulation matrix and the global state transition matrix to obtain a discretized input modulation matrix and a discretized state transition matrix; The discretized input modulation matrix, the discretized state transition matrix, and the hidden state are used to map the feature sequence into a hidden space representation; The positive enhancement feature sequence is obtained using the output modulation matrix and the hidden space representation; Obtain the inverse sequence of the feature sequence, and obtain the inverse enhancement feature sequence corresponding to the inverse sequence; The elements in the reverse enhancement feature sequence are arranged in reverse order and then fused with the forward enhancement feature sequence to obtain the enhancement feature sequence. The enhanced feature sequence is used to reconstruct the high-dimensional semantic features and the high-dimensional contour features to obtain the enhanced high-dimensional semantic features and the high-dimensional contour features.
6. The intelligent intracranial hematoma recognition system based on deep learning according to claim 4, characterized in that: The contour feature extractor initially extracts the contour features of the standardized brain CT image through a convolutional layer and uses it as the input of the first contour feature extraction layer. The input of the second to fourth contour feature extraction layers is the output of the previous contour feature extraction layer. The input to the contour generation layer is the output of the corresponding contour feature extraction layer, and the output of each contour generation layer is simultaneously used as the input to the fusion output layer to obtain the contour of the standardized brain CT image.
7. The intelligent intracranial hematoma recognition system based on deep learning according to claim 1, characterized in that: The intracranial hematoma classification and recognition module includes a feature extraction submodule and a classification and recognition submodule; the feature extraction submodule is used to acquire image block features; the classification and recognition submodule includes a classification and recognition model, and the classification and recognition submodule uses the classification and recognition model to classify and recognize intracranial hematomas based on the image block features and the standardized brain CT images.
8. The intelligent intracranial hematoma recognition system based on deep learning according to claim 7, characterized in that, The feature extraction submodule performs the following steps: The image patch features are obtained, including statistical features, texture features, and spatial features; The normalized image block features are mapped to the pixel locations of the standardized brain CT image to obtain a classification and recognition feature map.
9. The intelligent intracranial hematoma recognition system based on deep learning according to claim 8, characterized in that: The statistical features include at least the average CT value, the standard deviation of the CT value, the maximum CT value, the minimum CT value, and the entropy of the image block; the texture features include at least the contrast, correlation, and energy of the gray-level co-occurrence matrix; and the spatial features include at least the area and shape factor of the image block.
10. A deep learning-based intelligent identification system for intracranial hematoma according to claim 8, characterized in that, The classification and recognition model performs the following steps: The classification and recognition model fuses the classification and recognition feature map with the standardized brain CT image to obtain the classification and recognition fused feature; The classification and recognition model achieves classification and recognition of intracranial hematomas based on the classification and recognition fusion features.