Chest image recognition method based on improved dense convolutional network transfer learning

Abstract

The invention relates to a chest image recognition method based on improved dense convolutional network transfer learning. The method comprises the following steps: acquiring and preprocessing a chest image and physiological index data; inputting the preprocessed chest image and the physiological index data into a pre-trained chest image recognition model to obtain a disease prediction probability vector, calling a pre-constructed medical knowledge graph, and performing secondary optimization on the output probability vector; inputting the disease prediction probability vector into a pre-trained large language model, and performing comprehensive analysis in combination with associated information of a medical knowledge graph to obtain chest image recognition result suggestions and daily notes; wherein the chest image recognition model extracts chest image domain features based on gray distribution, noise patterns and contrast features of the preprocessed chest image in the recognition process, and parameters of the pre-trained chest image recognition model are adjusted according to the domain features. Compared with the prior art, the diagnosis is more suitable for the individual condition of the patient.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence medical technology, and in particular to a method for chest image recognition based on improved dense convolutional network transfer learning. Background Technology

[0002] In the field of medical image analysis, the AI-assisted chest X-ray image recognition task aims to provide clinicians with reliable tools through intelligent image analysis technology, improving the efficiency and accuracy of early detection and diagnosis of lung diseases. This task not only focuses on the detection and classification of common chest diseases but also identifies minute lesions and early-stage lesions, significantly reducing the probability of misdiagnosis. Furthermore, the assisted recognition system can play a crucial role in large-scale screening, public health control, and in areas with limited medical resources. Through this intelligent assisted diagnostic approach, doctors can obtain efficient, accurate, and standardized diagnostic results, enabling more timely intervention in clinical practice. Moreover, when processing complex image data, the AI ​​system can continuously improve its diagnostic performance and accuracy through continuous optimization and learning.

[0003] In the past, the following technical shortcomings existed in the use of chest X-ray images to assist in diagnosis: 1. Single disease identification only: Traditional medical image recognition systems can typically only handle the identification of a single disease and cannot detect multiple diseases simultaneously. Each disease needs to be screened separately, which is inefficient and easily overlooks other potential lesions.

[0004] 2. Traditional models are inefficient: Previous identification and classification models were computationally inefficient when processing large-scale medical data. In particular, the introduction of fully connected layers significantly increased computational costs, resulting in long model inference times, high latency in practical applications, and impacting the real-time performance of the system.

[0005] 3. Lack of intelligent diagnostic assistance: Traditional models can only provide predicted probabilities or classification results for diseases, lacking further clinical interpretation and diagnostic assistance. When using these tools, doctors still need to rely on a lot of manual operation and professional knowledge to interpret imaging results, which cannot truly reduce their workload.

[0006] 4. High false alarm rate and missed diagnosis: Due to the complex features of lesions in medical images, traditional models are prone to high false alarm rates when dealing with complex images, leading to the need for extensive manual secondary screening in clinical practice. Furthermore, insufficient feature extraction results in the risk of missed diagnosis for some traditional models when identifying small or inconspicuous lesions. Traditional models also struggle to adapt and accurately identify lesions when faced with images from different devices, different patient groups, and different disease types, resulting in low recognition accuracy.

[0007] 5. Lack of personalized health management recommendations: Traditional image recognition systems can only provide image analysis results, neglecting personalized care recommendations for patients. Patients often have to rely solely on doctors' instructions, lacking direct guidance for daily health management. Summary of the Invention

[0008] The purpose of this invention is to overcome the shortcomings of the existing technology and provide a chest image recognition method based on improved dense convolutional network transfer learning.

[0009] The objective of this invention can be achieved through the following technical solutions: A chest image recognition method based on transfer learning of improved dense convolutional networks, the method comprising: Chest imaging and physiological indicator data are acquired and preprocessed. The preprocessed chest imaging and physiological indicator data are then input into a pre-trained chest image recognition model to obtain a disease prediction probability vector. A pre-built medical knowledge graph is then used to further optimize the output probability vector. This disease prediction probability vector is then input into a pre-trained large language model and analyzed in conjunction with the association information from the medical knowledge graph to obtain suggestions on chest image recognition results and daily precautions. The chest image recognition model is trained using a transfer learning strategy. The disease prediction probability vector includes the prediction probability of each disease. When the large language model performs comprehensive analysis, it selects the disease prediction probabilities with confidence levels greater than a preset disease threshold based on the disease prediction probability vector and generates suggestions and daily precautions. During the recognition process, the chest image recognition model extracts chest image domain features based on the grayscale distribution, noise pattern, and contrast features of the preprocessed chest image, and adjusts the parameters of the pre-trained chest image recognition model according to the domain features.

[0010] Furthermore, the construction process of the pre-built medical knowledge graph includes: Acquire existing chest images and preprocess them to obtain a chest image dataset; Based on the chest imaging dataset and medical knowledge, the probability of common clinical comorbidities among diseases is calculated, and a comorbidity association database is constructed; based on the chest imaging dataset and medical knowledge, combinations of diseases that are clinically difficult to coexist are associated, and a mutual exclusion relationship database is constructed; based on the chest imaging dataset and medical knowledge, the typical anatomical disease areas corresponding to each disease are recorded, and an anatomical location association database is constructed. The co-disease association database, mutual exclusion relationship database, and anatomical location association database are associated and stored in a triplet format to form a callable structured database, thus completing the construction of a medical knowledge graph.

[0011] Furthermore, the pre-training process of the chest image recognition model includes: Acquire existing chest images and preprocess them to obtain a chest image dataset; A deep convolutional neural network trained on a large chest image dataset is used as the backbone network of the chest image recognition model, and a cross-domain adaptive dynamic calibration module is added to construct the chest image recognition model. A composite loss function is constructed by combining binary cross-entropy with a medical logic penalty term. The chest image recognition model is then trained using the composite loss function and a chest image dataset based on a training iteration strategy. The training iteration strategy is as follows: during the initial number of iterations, the backbone network of the chest image recognition model is frozen, and the parameters related to the calibrator and loss function are trained. After the initial number of iterations, the backbone network is unfrozen for overall training. The loss is calculated through forward propagation, and the weight parameters of the network are updated using gradients during backpropagation. Output the trained chest image recognition model.

[0012] Furthermore, the preprocessing includes: Collect existing chest images and perform image size unification, normalization, and standardization; Based on the disease type, X-ray equipment type, and population type of existing chest images, image labels are assigned to each existing chest image, and the image labels are stored as strings through unique thermal encoding conversion. Based on a pre-built medical knowledge graph, the comorbidity matching degree and mutual exclusion conflict value of the image labels of each existing chest image are calculated and incorporated into the image label string; Based on the total number of disease categories in existing chest images, calculate the category weights and incorporate them into the image label string; The processed existing chest images are divided into a training set and a test set, and the training set and test set are randomly varied. The random variation includes flipping the images left and right according to a preset probability. Finally, existing chest images with label strings are obtained to construct a chest image dataset.

[0013] Furthermore, the label strings of each existing chest image obtained after preprocessing include: device type, population type, disease type, comorbidity matching degree, mutual exclusion conflict value, and category weight; The formula for calculating the comorbidity matching degree is: in, To determine the comorbidity matching of current chest imaging, This represents the number of diseases currently listed in chest imaging labels. and Let $\mathbf{i}$ and $\mathbf{j}$ represent the probability of the existence of disease $i$ and disease $j$, respectively. When disease $i$ is present in the existing chest image labels... When the i-th disease is not present in the existing chest imaging labels, When the j-th disease is present in the existing chest imaging labels, When the jth disease is not present in the existing chest imaging labels, , Indicates the i-th disease. To represent the j-th disease, The probability of clinical comorbidity between the i-th disease and the j-th disease is collected from a pre-built medical knowledge graph. The expression for calculating the mutual exclusion conflict value is as follows: in, This indicates the mutual exclusion conflict value of the current chest images. The mutual exclusion relationship value between the i-th disease and the j-th disease is collected from a pre-constructed medical knowledge graph; The formula for calculating the category weight is: in, C It is the total number of all disease categories. To train the set of existing chest images, No. j The number of existing chest images for this type of disease.

[0014] Furthermore, the expression for the composite loss function is: in, For binary cross-entropy loss, For medical logic penalty items, This refers to the weight value of the penalty item; in, For the label of the i-th sample, which is the j-th disease category, Let $\frac{i}{j}$ be the predicted probability of the $j$-th disease in the $i$-th sample. The total number of samples in the training set. The total number of target disease categories, Let be the mutual exclusion conflict value of the i-th sample. Let be the comorbidity matching degree of the prediction result for the i-th sample. This is the penalty coefficient for comorbidity matching. Let be the number of diseases with a confidence level ≥ 0.5 in the prediction results of the i-th sample. Let $\frac{i}{j}$ be the predicted probability of the $j$-th disease in the $i$-th sample. Let k be the predicted probability of the i-th sample being the k-th disease. For disease and The probability of clinical comorbidity, For disease type j, For the kth type of disease, This is an indicator function.

[0015] Furthermore, the deep convolutional neural network includes an input layer, dense blocks, a transition layer, a global average pooling layer, and an output layer. The cross-domain adaptive dynamic calibration module is embedded between the input layer and the dense blocks. After the chest image and physiological index data are input into the chest image recognition model through the input layer, the cross-domain adaptive dynamic calibration module extracts chest image domain features based on the grayscale distribution, noise pattern, and contrast features of the preprocessed chest image, and adjusts the parameters of the transition layer according to the domain features. The adjusted chest image domain features are then input into the subsequent dense blocks.

[0016] Furthermore, there are no fewer than four dense blocks, with a transition layer between each dense block. The dense blocks are used to extract new features from the chest image, and the transition layers are used to reduce the number of channels and spatial dimensions of the chest image.

[0017] Furthermore, the cross-domain adaptive dynamic calibration module includes convolutional layers and fully connected layers; Before incorporating the chest image recognition model, the cross-domain adaptive dynamic calibration module is pre-trained using meta-learning based on a pre-established cross-domain auxiliary dataset to obtain the trained cross-domain adaptive dynamic calibration module and a parameter adjustment mapping table for adjusting the parameters of the transition layer; the cross-domain auxiliary dataset contains chest images labeled with device type, population type, and disease label. The objective function of the meta-learning pre-training is to minimize the prediction error of cross-domain samples, and its expression is: in, These are the initial parameters. This is a cross-domain task set, where each task corresponds to a type of device or group of people. For single-task training set, For training losses within the mission, For the learning step size, Losses due to task testing.

[0018] Furthermore, the output layer includes a fully connected layer and a sigmoid function. The fully connected layer generates an output vector corresponding to the number of diseases based on the feature vector, and the sigmoid function is used to map the linear output of the fully connected layer to probability values ​​in the range [0, 1].

[0019] Compared with the prior art, the beneficial effects of the present invention include: 1. This invention addresses the core pain points of traditional chest image recognition methods—poor cross-scenario generalization, predictions that violate clinical logic, weak interpretability, and high implementation costs—through cross-domain adaptive calibration, medical knowledge graph collaboration, and multimodal data fusion, thereby improving the accuracy and efficiency of chest image recognition. This invention integrates chest image and physiological indicator data, incorporating physiological indicators into the consideration of their impact on disease. By fusing physiological indicators with image data, it overcomes the limitations of traditional methods that rely solely on images, enriching feature dimensions and making diagnoses more closely aligned with individual patient conditions. The recognition model of this invention can extract chest image domain features based on the grayscale distribution, noise patterns, and contrast characteristics of chest images. It then adjusts the parameters of the pre-trained chest image recognition model according to these domain features, enabling specific analysis of specific images. Traditional models require extensive local data collection and retraining to adapt to different devices; this invention achieves adaptability simply by recognizing different devices through domain features.

[0020] 2. The medical knowledge graph of this invention includes a comorbidity association library, a mutually exclusive relationship library, and an anatomical location association library. It stores clinical prior knowledge in triplets. The knowledge graph can limit absurd results of coexisting mutually exclusive diseases or omission of high-frequency comorbidities, making the identification results more accurate.

[0021] 3. This invention constrains model training through medical logic penalty terms. When the mutual exclusion conflict value is 1, the loss value is directly increased. When the comorbidity matching is too low, additional penalties are added, enabling the model to learn clinically reasonable disease combination patterns and reducing the misjudgment rate of mutual exclusion diseases and the missed detection rate of comorbidities.

[0022] 4. After the initial judgment by the model, the present invention also makes a secondary judgment based on the medical knowledge graph. False disease combinations no longer need to be manually identified by doctors, which can greatly reduce screening time. Attached Figure Description

[0023] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0025] Example 1 This embodiment discloses a chest image recognition method based on improved dense convolutional network transfer learning, the method as follows: Figure 1 As shown, steps S1-S4 are included, and each step is described as follows: Step S1: Acquire chest imaging and physiological data and perform preprocessing.

[0026] Specifically, preprocessing includes image denoising, resizing, normalization, and standardization.

[0027] Step S2: Input the preprocessed chest images and physiological index data into the pre-trained chest image recognition model to obtain the disease prediction probability vector.

[0028] The chest image recognition model is trained using a transfer learning strategy, and the disease prediction probability vector includes the predicted probability of each disease.

[0029] During the recognition process, the chest image recognition model extracts chest image domain features based on the grayscale distribution, noise pattern, and contrast characteristics of the pre-processed chest image, and adjusts the parameters of the pre-trained chest image recognition model according to the domain features.

[0030] The pre-training process of the chest image recognition model includes: Acquire existing chest images and preprocess them to obtain a chest image dataset; A deep convolutional neural network (DenseNet121 model) trained on a large chest image dataset was used as the backbone network of the chest image recognition model, and a cross-domain adaptive dynamic calibration module was added to construct the chest image recognition model. A composite loss function is constructed by combining binary cross-entropy with a medical logic penalty term. Using the composite loss function and a chest image dataset, a chest image recognition model is trained based on a training iteration strategy. The training iteration strategy is as follows: during the initial number of iterations, the backbone network of the chest image recognition model is frozen, and the parameters related to the calibrator and loss function are trained. After the initial number of iterations, the backbone network is unfrozen and the whole training is performed. The loss is calculated through forward propagation, and the weight parameters of the network are updated using gradients during backpropagation. Output the trained chest image recognition model.

[0031] In the preprocessing stage, a pre-built medical knowledge graph will be used.

[0032] The construction process of a pre-built medical knowledge graph includes: Acquire existing chest images and preprocess them to obtain a chest image dataset; Based on chest imaging datasets and medical knowledge, we calculate the probability of common clinical comorbidities among diseases and construct a comorbidity association database. Based on chest imaging datasets and medical knowledge, a mutually exclusive relation database is constructed by associating disease combinations that are logically difficult to coexist in clinical practice. Based on chest imaging datasets and medical knowledge, we record the typical anatomical disease areas corresponding to each disease and construct an anatomical location association database. The associated comorbidity database, mutual exclusion database, and anatomical location database are stored in a triplet format to form a callable structured database, thus completing the construction of a medical knowledge graph.

[0033] Preprocessing includes: Collect existing chest images and perform image size unification, normalization, standardization, and data augmentation; Resizing: Unify image size to The resolution should be adjusted to ensure consistent input size. Normalization: Scales the image pixel values ​​from [0, 255] to [0, 1]. Let the image pixel values ​​be... The normalized pixel value is: ; Standardization: Images are standardized using the mean and standard deviation of the ImageNet dataset to make the image distribution more uniform. Let the mean vector of ImageNet be... The standard deviation vector is The standardized operating procedure is as follows: ; Data augmentation: Random variations are applied to the training data to enhance the dataset and improve the model's generalization ability; images are mirrored horizontally with a 50% probability. Based on the disease type, X-ray equipment type, and population type of existing chest images, image labels are assigned to each existing chest image, and the image labels are stored as strings through unique thermal encoding conversion. Based on a pre-built medical knowledge graph, the comorbidity matching degree and mutual exclusion conflict value of the image labels of each existing chest image are calculated and incorporated into the image label string; Based on the total number of disease categories in existing chest images, calculate the category weights and incorporate them into the image label string; The processed existing chest images are divided into training and test sets, and random variations are applied to the training and test sets, including left-right mirroring of the images according to a preset probability. Finally, existing chest images with label strings are obtained to construct a chest image dataset.

[0034] Specifically, the image label information is stored in a CSV file as a two-dimensional table, with each row representing an image and its corresponding label. The total number of samples in the dataset is [number missing]. N Then it can be expressed as: in, It is the image file name (e.g., "000001.png"). This represents the label information corresponding to the image in its encoded form. The label can be a single disease or a combination of multiple diseases.

[0035] The label strings of each existing chest image obtained after preprocessing include: device type, population type, disease type, comorbidity matching degree, mutual exclusion conflict value, and category weight.

[0036] The formula for calculating the comorbidity matching degree is: in, To determine the comorbidity matching of current chest imaging, This represents the number of diseases currently listed in chest imaging labels. and Let $\mathbf{i}$ and $\mathbf{j}$ represent the probability of the existence of disease $i$ and disease $j$, respectively. When disease $i$ is present in the existing chest image labels... When the i-th disease is not present in the existing chest imaging labels, When the j-th disease is present in the existing chest imaging labels, When the jth disease is not present in the existing chest imaging labels, , Indicates the i-th disease. To represent the j-th disease, This represents the clinical comorbidity probability between the i-th disease and the j-th disease, collected from a pre-built medical knowledge graph.

[0037] The expression for calculating the mutual exclusion conflict value is: in, This indicates the mutual exclusion conflict value of the current chest images. The value represents the mutual exclusion relationship between the i-th disease and the j-th disease, collected from a pre-built medical knowledge graph.

[0038] Since some categories may have relatively few samples in multi-label classification tasks, class weights are calculated to address the class imbalance problem. The formula for calculating class weights is as follows: in, C It is the total number of all disease categories. To train the set of existing chest images, No. j The number of existing chest images for this type of disease.

[0039] The expression for the composite loss function is: in, For binary cross-entropy loss, For medical logic penalty items, This refers to the weight value of the penalty item; in, For the label of the i-th sample, which is the j-th disease category, Let $\frac{i}{j}$ be the predicted probability of the $j$-th disease in the $i$-th sample. The total number of samples in the training set. The total number of target disease categories, Let be the mutual exclusion conflict value of the i-th sample. Let be the comorbidity matching degree of the prediction result for the i-th sample. This is the penalty coefficient for comorbidity matching. Let be the number of diseases with a confidence level ≥ 0.5 in the prediction results of the i-th sample. Let $\frac{i}{j}$ be the predicted probability of the $j$-th disease in the $i$-th sample. Let k be the predicted probability of the i-th sample being the k-th disease. For disease and The probability of clinical comorbidity, For disease type j, For the kth type of disease, This is an indicator function.

[0040] The deep convolutional neural network includes an input layer, dense blocks, transition layers, a global average pooling layer, and an output layer. A cross-domain adaptive dynamic calibration module is embedded between the input layer and the dense blocks. After chest images and physiological index data are input into the chest image recognition model through the input layer, the cross-domain adaptive dynamic calibration module extracts chest image domain features based on the grayscale distribution, noise pattern, and contrast features of the preprocessed chest images. It then adjusts the parameters of the transition layer according to the domain features, and the adjusted chest image domain features are input into the subsequent dense blocks.

[0041] There are at least four dense blocks, with transition layers between each dense block. The dense blocks are used to extract new features from the chest image, and the transition layers are used to reduce the number of channels and spatial dimensions of the chest image.

[0042] Specifically, in the input layer: the encoder uses a convolutional neural network with a kernel size of 7x7, a stride of 2, and the output feature map size is half that of the input image.

[0043] Input data: The model receives chest X-ray images with dimensions of 512×512×3, typically in RGB three-channel format. The input layer primarily receives and prepares the data for subsequent processing by the convolutional network.

[0044] After passing through the input layer, the original image data (512×512×3) is transformed into a 128×128×64 feature matrix.

[0045] The cross-domain adaptive dynamic calibration module includes convolutional layers and fully connected layers; Before incorporating the chest image recognition model, the cross-domain adaptive dynamic calibration module also performs meta-learning pre-training based on a pre-established cross-domain auxiliary dataset to obtain the trained cross-domain adaptive dynamic calibration module and a parameter adjustment mapping table for adjusting the parameters of the transition layer; the cross-domain auxiliary dataset contains chest images labeled with device type, population type and disease label; The objective function of meta-learning pre-training is to minimize the prediction error of cross-domain samples, and its expression is: in, These are the initial parameters. This is a cross-domain task set, where each task corresponds to a type of device or group of people. For single-task training set, For training losses within the mission, For the learning step size, Losses due to task testing.

[0046] After inputting the cross-domain adaptive dynamic calibration module, the 128×128×64 feature matrix is ​​converted into 128×128×64 chest image domain features, and the weights of the transition layer are adjusted according to the domain features.

[0047] Dense Blocks: The core of the DenseNet121 model consists of four dense blocks. Their key characteristic is dense connectivity, meaning each convolutional layer receives the outputs of all previous layers as its input. This ensures that each layer can directly access the features of preceding layers, avoiding information loss and improving gradient propagation efficiency.

[0048] Dense Block 1: The 128×128×64 chest image features are fed into the first dense block. This block contains 6 convolutional layers. Each convolutional layer extracts new features and concatenates them with the output of the previous layer. The final output feature matrix has increased dimensionality, while the spatial dimension remains unchanged (still 128×128), but the depth increases because each convolutional layer generates new feature channels.

[0049] Transition Layer 1: After the first dense block, there is a transition layer. The transition layer contains a 1×1 convolutional layer to reduce the number of channels, thereby reducing computational cost. This is followed by a pooling layer to further reduce the spatial dimension of the feature matrix. After the transition layer, the spatial dimension of the feature matrix is ​​reduced to 64×64, and the number of channels (depth) is also adjusted (reduced).

[0050] Dense Block 2: The 64×64 feature matrix after the transition layer is fed into the second dense block, which contains 12 convolutional layers. Similar to Dense Block 1, the output of each convolutional layer is concatenated with the output of the previous layer to further extract new features.

[0051] Transition layer 2: After passing through the transition layer again, the spatial dimension of the feature matrix is ​​further reduced from 64×64 to 32×32.

[0052] Dense Block 3: The third dense block contains 24 convolutional layers, which process a 32×32 feature matrix to extract more high-level features.

[0053] Transition layer 3: After the third transition layer, the spatial dimension of the feature matrix is ​​reduced to 16×16.

[0054] Dense Block 4: The last dense block contains 16 convolutional layers, which process a 16×16 feature matrix to extract the most complex high-level features.

[0055] Global Average Pooling Layer: After all the dense blocks and transition layers, the feature matrix becomes 16×16×number of channels (designed to have 1024 channels). At this point, the model applies Global Average Pooling (GAP): GAP averages all values ​​of each feature map, generating a global value, transforming the 16×16×number of channels feature matrix into a 1×1×number of channels vector. This step not only reduces the number of parameters but also prevents overfitting. After GAP, the feature matrix becomes a flat vector with the same length as the number of feature channels. In a specific design, the feature matrix dimension is changed from 16×16×1024 to 1×1×1024.

[0056] Output layer: Fully connected layer and Sigmoid function: Fully connected layer: Responsible for generating an output vector (14-dimensional) corresponding to the number of diseases from a 1024-dimensional feature vector, with each dimension representing a predicted value for a disease.

[0057] The Sigmoid function maps the linear output of a fully connected layer to probability values ​​in the range [0,1], representing the probability of each disease occurring.

[0058] Step S3: Call the pre-built medical knowledge graph to perform secondary optimization on the output probability vector.

[0059] Specifically, this refers to the initial probability vector of the output. A second optimization is performed to obtain the final probability vector. .

[0060] The optimization rules are as follows: Mutually exclusive elimination rule: If there exists a predicted probability for disease type j... ≥0.5 and ≥0.5 and If the probability of the higher probability is retained, the predicted probability of the other is adjusted to 0.2.

[0061] Comorbidity completion rules: If ≥0.5 and And 0.3≤ A value <0.5 suggests the possible presence of comorbidities.

[0062] Step S4: Input the disease prediction probability vector into the pre-trained large language model, and perform comprehensive analysis by combining the association information of the medical knowledge graph to obtain suggestions on chest image recognition results and daily precautions.

[0063] The final probability vector The content related to the medical knowledge graph is input into the pre-trained large language model. The application process here is based on existing technology and will not be elaborated here.

[0064] Example 2 Based on Embodiment 1, this embodiment provides an electronic device, including: one or more processors and a memory, wherein the memory stores one or more programs, and the one or more programs include instructions for executing the chest image recognition method based on the aforementioned improved dense convolutional network transfer learning.

[0065] At the hardware level, the electronic device includes a processor, internal bus, network interface, memory, and non-volatile memory, and may also include other hardware required for the business. The processor reads the corresponding computer program from the non-volatile memory into memory and then runs it to implement the chest image recognition method based on improved dense convolutional network transfer learning described above. Of course, in addition to software implementation, this invention does not exclude other implementation methods, such as logic devices or a combination of hardware and software, etc. That is to say, the execution subject of the following processing flow is not limited to individual logic units, but can also be hardware or logic devices.

[0066] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0067] Computer-readable media include both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0068] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for chest image recognition based on transfer learning of improved dense convolutional networks, characterized in that, The method includes: Chest imaging and physiological indicator data are acquired and preprocessed. The preprocessed chest imaging and physiological indicator data are then input into a pre-trained chest image recognition model to obtain a disease prediction probability vector. A pre-built medical knowledge graph is then used to further optimize the output probability vector. This disease prediction probability vector is then input into a pre-trained large language model and analyzed in conjunction with the association information from the medical knowledge graph to obtain suggestions on chest image recognition results and daily precautions. The chest image recognition model is trained using a transfer learning strategy. The disease prediction probability vector includes the prediction probability of each disease. When the large language model performs comprehensive analysis, it selects the disease prediction probabilities with confidence levels greater than a preset disease threshold based on the disease prediction probability vector and generates suggestions and daily precautions. During the recognition process, the chest image recognition model extracts chest image domain features based on the grayscale distribution, noise pattern, and contrast features of the preprocessed chest image, and adjusts the parameters of the pre-trained chest image recognition model according to the domain features.

2. The chest image recognition method based on improved dense convolutional network transfer learning according to claim 1, characterized in that, The construction process of the pre-built medical knowledge graph includes: Acquire existing chest images and preprocess them to obtain a chest image dataset; Based on the chest imaging dataset and medical knowledge, the probability of common clinical comorbidities among diseases is calculated, and a comorbidity association database is constructed; based on the chest imaging dataset and medical knowledge, combinations of diseases that are clinically difficult to coexist are associated, and a mutual exclusion relationship database is constructed; based on the chest imaging dataset and medical knowledge, the typical anatomical disease areas corresponding to each disease are recorded, and an anatomical location association database is constructed. The co-disease association database, mutual exclusion relationship database, and anatomical location association database are associated and stored in a triplet format to form a callable structured database, thus completing the construction of a medical knowledge graph.

3. The chest image recognition method based on improved dense convolutional network transfer learning according to claim 1, characterized in that, The pre-training process of the chest image recognition model includes: Acquire existing chest images and preprocess them to obtain a chest image dataset; A deep convolutional neural network trained on a large chest image dataset is used as the backbone network of the chest image recognition model, and a cross-domain adaptive dynamic calibration module is added to construct the chest image recognition model. A composite loss function is constructed by combining binary cross-entropy with a medical logic penalty term. The chest image recognition model is then trained using the composite loss function and a chest image dataset based on a training iteration strategy. The training iteration strategy is as follows: during the initial number of iterations, the backbone network of the chest image recognition model is frozen, and the parameters related to the calibrator and loss function are trained. After the initial number of iterations, the backbone network is unfrozen for overall training. The loss is calculated through forward propagation, and the weight parameters of the network are updated using gradients during backpropagation. Output the trained chest image recognition model.

4. The chest image recognition method based on improved dense convolutional network transfer learning according to claim 2, characterized in that, The preprocessing includes: Collect existing chest images and perform image size unification, normalization, and standardization; Based on the disease type, X-ray equipment type, and population type of existing chest images, image labels are assigned to each existing chest image, and the image labels are stored as strings through unique thermal encoding conversion. Based on a pre-built medical knowledge graph, the comorbidity matching degree and mutual exclusion conflict value of the image labels of each existing chest image are calculated and incorporated into the image label string; Based on the total number of disease categories in existing chest images, calculate the category weights and incorporate them into the image label string; The processed existing chest images are divided into a training set and a test set, and the training set and test set are randomly varied. The random variation includes flipping the images left and right according to a preset probability. Finally, existing chest images with label strings are obtained to construct a chest image dataset.

5. The chest image recognition method based on improved dense convolutional network transfer learning according to claim 4, characterized in that, The label strings of each existing chest image obtained after preprocessing include: device type, population type, disease type, comorbidity matching degree, mutual exclusion conflict value, and category weight; The formula for calculating the comorbidity matching degree is: in, To determine the comorbidity matching of current chest imaging, This represents the number of diseases currently listed in chest imaging labels. and Let $\mathbf{i}$ and $\mathbf{j}$ represent the probability of the existence of disease $i$ and disease $j$, respectively. When disease $i$ is present in the existing chest image labels... When the i-th disease is not present in the existing chest imaging labels, When the j-th disease is present in the existing chest imaging labels, When the jth disease is not present in the existing chest imaging labels, , Indicates the i-th disease. To represent the j-th disease, The probability of clinical comorbidity between the i-th disease and the j-th disease is collected from a pre-built medical knowledge graph. The expression for calculating the mutual exclusion conflict value is as follows: in, This indicates the mutual exclusion conflict value of the current chest images. The mutual exclusion relationship value between the i-th disease and the j-th disease is collected from a pre-constructed medical knowledge graph; The formula for calculating the category weight is: in, C It is the total number of all disease categories. To train the set of existing chest images, No. j The number of existing chest images for this type of disease.

6. The chest image recognition method based on improved dense convolutional network transfer learning according to claim 3, characterized in that, The expression for the composite loss function is: in, For binary cross-entropy loss, For medical logic penalty items, This refers to the weight value of the penalty item; in, For the label of the i-th sample, which is the j-th disease category, Let $\frac{i}{j}$ be the predicted probability of the $j$-th disease in the $i$-th sample. The total number of samples in the training set. The total number of target disease categories, Let be the mutual exclusion conflict value of the i-th sample. Let be the comorbidity matching degree of the prediction result for the i-th sample. This is the penalty coefficient for comorbidity matching. Let be the number of diseases with a confidence level ≥ 0.5 in the prediction results of the i-th sample. Let $\frac{i}{j}$ be the predicted probability of the $j$-th disease in the $i$-th sample. Let k be the predicted probability of the i-th sample being the k-th disease. For disease and The probability of clinical comorbidity, For disease type j, For the kth type of disease, This is an indicator function.

7. The chest image recognition method based on improved dense convolutional network transfer learning according to claim 3, characterized in that, The deep convolutional neural network includes an input layer, dense blocks, a transition layer, a global average pooling layer, and an output layer. The cross-domain adaptive dynamic calibration module is embedded between the input layer and the dense blocks. After the chest images and physiological index data are input into the chest image recognition model through the input layer, the cross-domain adaptive dynamic calibration module extracts chest image domain features based on the grayscale distribution, noise pattern, and contrast features of the preprocessed chest images. The parameters of the transition layer are then adjusted according to the domain features. The adjusted chest image domain features are then input into the subsequent dense blocks.

8. The chest image recognition method based on improved dense convolutional network transfer learning according to claim 7, characterized in that, There are at least four dense blocks, with a transition layer between each dense block. The dense blocks are used to extract new features from the chest image, and the transition layers are used to reduce the number of channels and spatial dimensions of the chest image.

9. A chest image recognition method based on improved dense convolutional network transfer learning according to claim 7, characterized in that, The cross-domain adaptive dynamic calibration module includes convolutional layers and fully connected layers; Before incorporating the chest image recognition model, the cross-domain adaptive dynamic calibration module is pre-trained using meta-learning based on a pre-established cross-domain auxiliary dataset to obtain the trained cross-domain adaptive dynamic calibration module and a parameter adjustment mapping table for adjusting the parameters of the transition layer; the cross-domain auxiliary dataset contains chest images labeled with device type, population type, and disease label. The objective function of the meta-learning pre-training is to minimize the prediction error of cross-domain samples, and its expression is: in, These are the initial parameters. This is a cross-domain task set, where each task corresponds to a type of device or group of people. For single-task training set, For training losses within the mission, For the learning step size, Losses due to task testing.

10. A chest image recognition method based on improved dense convolutional network transfer learning according to claim 7, characterized in that, The output layer includes a fully connected layer and a sigmoid function. The fully connected layer generates an output vector corresponding to the number of diseases based on the feature vector. The sigmoid function is used to map the linear output of the fully connected layer to probability values ​​in the range [0, 1].

Images