Medical image big data governance method based on artificial intelligence

Abstract

The invention relates to data management, in particular to a medical image big data management method based on artificial intelligence, and the method comprises the steps: obtaining multi-source heterogeneous data from a medical image cloud platform, and carrying out the data integration and data reconstruction; performing data desensitization on the preprocessed data to protect patient information; performing data cleaning on the desensitized data by using a pre-trained medical image data cleaning model, and removing data which cannot be used; performing classification and recognition on the cleaned data by using a pre-trained medical image classification and recognition model to obtain various types of data sets; according to the technical scheme provided by the invention, the defect that a high-quality medical image data set is difficult to accurately and efficiently obtain in the prior art can be effectively overcome.

Description

Technical Field

[0001] The present invention relates to data management, and in particular to a medical imaging big data management method based on artificial intelligence. Background Art

[0002] As medical imaging big data enters a stage of rapid development, massive amounts of medical imaging data can promote the development of artificial intelligence. However, at this stage, there is a lack of high-quality medical imaging data sets, and many valuable medical imaging data have not been effectively utilized, resulting in serious value loss of medical imaging data resources. There are two main reasons for this. On the one hand, because medical imaging data involves the personal privacy of patients, data owners cannot provide sufficient experimental data to participate in the design and evaluation process of diagnostic models; on the other hand, medical imaging data is highly professional, and tasks such as labeling can only be completed by professional doctors. At the same time, the structure of medical imaging data is relatively complex, and doctors need to obtain it manually. The data is not diverse enough, resulting in medical imaging data from different medical institutions. In the end, only a small part of the data is provided to the field of artificial intelligence for research after professional doctors spend a lot of time manually screening and labeling.

[0003] Nevertheless, this type of data is still limited by the research on data governance systems for intelligent medical imaging diagnosis, the manufacturer and model of imaging equipment, and the patient's physical characteristics, resulting in a certain degree of flaws in this type of high-quality data. The existing commonly used public medical imaging datasets are far from the optimal level of artificial intelligence research in terms of data quality and quantity. Even the evaluation and improvement of these datasets has become a research direction in the field of academic research in intelligent medical imaging diagnosis. Summary of the invention

[0004] 1. Technical issues to be resolved

[0005] In view of the above-mentioned shortcomings of the prior art, the present invention provides a medical imaging big data management method based on artificial intelligence, which can effectively overcome the defects of the prior art that it is difficult to accurately and efficiently obtain high-quality medical imaging data sets.

[0006] (II) Technical solution

[0007] To achieve the above objectives, the present invention is implemented through the following technical solutions:

[0008] A medical imaging big data governance method based on artificial intelligence includes the following steps:

[0009] S1. Obtain multi-source heterogeneous data from the medical imaging cloud platform, and perform data integration and data reconstruction;

[0010] S2. Desensitize the pre-processed data to protect patient information;

[0011] S3. Use the pre-trained medical imaging data cleaning model to clean the desensitized data and remove unusable data;

[0012] S4. Use the pre-trained medical image classification and recognition model to classify and recognize the cleaned data to obtain data sets of each category.

[0013] Preferably, S1 obtains multi-source heterogeneous data from the medical imaging cloud platform, and performs data integration and data reconstruction, including:

[0014] Acquire multi-source heterogeneous data from the medical imaging cloud platform and convert DICOM medical images into common image formats.

[0015] Preferably, S2 performs data desensitization on the pre-processed data to protect patient information, including:

[0016] The patient information in the converted DICOM medical images is desensitized by using masking, transformation, rounding and deletion methods to protect the patient information.

[0017] Preferably, S3 uses a pre-trained medical imaging data cleaning model to clean the desensitized data, and before removing unusable data, includes:

[0018] The model training of the medical image data cleaning model includes the following steps:

[0019] S301, annotating a DICOM medical image historical data set, and dividing the annotated DICOM medical image historical data set into a training set, a validation set, and a test set according to a preset ratio;

[0020] S302, setting the loss function and optimizer of the medical image data cleaning model;

[0021] S303, inputting the training set into the medical imaging data cleaning model for model training;

[0022] S304, calculating the loss value based on the loss function, and the optimizer updates the model parameters according to the loss value and the network gradient information;

[0023] S305. If the loss value is less than the preset threshold, the model training ends, and the current medical image data cleaning model is the trained medical image data cleaning model. Otherwise, the process returns to S303 and continues to use the training set for model training.

[0024] S306, inputting the validation set into the trained medical image data cleaning model, evaluating the generalization ability of the model by observing the performance on the validation set, and tuning the hyperparameters and structure of the model;

[0025] S307, inputting the test set into the tuned medical image data cleaning model to evaluate the model performance;

[0026] Among them, the medical image data cleaning model is built based on the Transformer model. The loss function uses the cross entropy loss function to measure the difference between the predicted value and the available label. The optimizer uses the AdamW optimizer. At the same time, Dropout or other regularization techniques are used during the model training process to prevent overfitting.

[0027] Preferably, in S301, the DICOM medical image historical data set is annotated, and the annotated DICOM medical image historical data set is divided into a training set, a validation set, and a test set according to a preset ratio, including:

[0028] S3011, performing normalization processing including cropping, scaling and normalization on the DICOM medical image historical data set;

[0029] S3012, expanding the normalized DICOM medical image historical data set by using rotation, flipping, and color jittering methods;

[0030] S3013, labeling each DICOM medical image in the expanded DICOM medical image historical data set, and setting an available label for each DICOM medical image;

[0031] S3014. Divide the annotated DICOM medical image historical data set into a training set, a validation set, and a test set according to a preset ratio.

[0032] Preferably, S3 uses a pre-trained medical imaging data cleaning model to clean the desensitized data and remove unusable data, including:

[0033] S311, performing normalization processing including cropping, scaling and normalization on the desensitized DICOM medical image;

[0034] S312: Input the normalized DICOM medical images into a pre-trained medical image data cleaning model, determine unusable DICOM medical images based on the output results, and remove them.

[0035] Preferably, the medical image data cleaning model includes an input part, an encoder, a decoder and an output part;

[0036] In the input part, the DICOM medical image is divided into blocks of fixed size. Each block is regarded as a word. A linear transformation layer is used to convert each image block into a vector of fixed dimension. A position code is added after the vector so that the model can understand the relative position information between image blocks.

[0037] The encoder includes a multi-head self-attention sublayer and a first feed-forward fully connected sublayer. The multi-head self-attention sublayer uses a multi-head self-attention mechanism to capture the dependencies between image blocks. Each head focuses on a different part of the image, thereby obtaining global information of the image. The first feed-forward fully connected sublayer performs a nonlinear transformation on the output of the multi-head self-attention sublayer for further feature extraction.

[0038] A decoder, comprising a masked multi-head self-attention sublayer, an encoder-decoder attention sublayer, and a second feed-forward fully connected sublayer, wherein the masked multi-head self-attention sublayer adds a masking operation on the basis of the multi-head self-attention sublayer, so that the decoder can only rely on the first i-1 words when generating the i-th word, and maintains the autoregressive characteristics of the decoding process, the encoder-decoder attention sublayer interacts the input of the current position of the decoder with all positions of the encoder to obtain information related to the target sequence, and generates an output sequence corresponding to the input sequence, and the second feed-forward fully connected sublayer performs a nonlinear transformation on the output of the encoder-decoder attention sublayer;

[0039] The output part includes a fully connected layer and a Softmax layer. The fully connected layer converts the vector output by the decoder into the final output dimension, and the Softmax layer converts the output of the fully connected layer into a category probability distribution.

[0040] Among them, the multi-head self-attention sublayer and the first feed-forward fully connected sublayer in the encoder are connected with a residual connection and a normalization layer, and the masked multi-head self-attention sublayer, encoder-decoder attention sublayer and the second feed-forward fully connected sublayer in the decoder are connected with a residual connection and a normalization layer.

[0041] Preferably, in S4, the pre-trained medical image classification and recognition model is used to classify and recognize the cleaned data, and before obtaining each category data set, the following steps are included:

[0042] S401, annotating the available DICOM medical image historical data set, and dividing the annotated DICOM medical image historical data set into a training set, a validation set, and a test set according to a preset ratio;

[0043] S402, setting the loss function and optimizer of the medical image classification and recognition model;

[0044] S403, inputting the training set into the medical image classification and recognition model for model training;

[0045] S404, calculating the loss value based on the loss function, and the optimizer updates the model parameters according to the loss value and the network gradient information;

[0046] S405. If the loss value is less than the preset threshold, the model training ends, and the current medical image classification and recognition model is the trained medical image classification and recognition model. Otherwise, the process returns to S403 and continues to use the training set for model training.

[0047] S406, inputting the validation set into the trained medical image classification and recognition model, evaluating the generalization ability of the model by observing the performance on the validation set, and tuning the hyperparameters and structure of the model;

[0048] S407, inputting the test set into the tuned medical image classification and recognition model to evaluate the model performance;

[0049] Among them, the medical image classification and recognition model is built based on the ResNet50 model;

[0050] During the model training process, the learning rate was set to 0.001, the batch size was set to 128, the Adam optimizer was used as the optimizer, and the model training was stopped after 20 epochs;

[0051] When evaluating model performance, ROC curve and sensitivity SE are used as model performance evaluation indicators.

[0052] Preferably, in S401, the available DICOM medical image historical data set is annotated, and the annotated DICOM medical image historical data set is divided into a training set, a validation set and a test set according to a preset ratio, including:

[0053] S4011, labeling each DICOM medical image according to the X-ray examination item to which each DICOM medical image belongs in the available DICOM medical image historical data set, and setting a category label for each DICOM medical image;

[0054] S4012, dividing the annotated usable DICOM medical image historical data set into a training set, a validation set, and a test set according to a preset ratio of 7:1:2;

[0055] Among them, the X-ray examination items include posteroanterior radiograph of chest, lateral radiograph of thoracic spine, anteroposterior radiograph of lumbar spine, lateral radiograph of lumbar spine, upright radiograph of abdomen, anteroposterior radiograph of pelvis, anteroposterior radiograph of cervical spine, lateral radiograph of cervical spine, anteroposterior radiograph of thoracic spine, anteroposterior radiograph of sacrum and coccyx, lateral radiograph of sacrum and coccyx, posteroanterior radiograph of hand, oblique radiograph of hand, anteroposterior radiograph of hip joint, anteroposterior radiograph of knee joint, lateral radiograph of knee joint, anteroposterior radiograph of ankle joint, lateral radiograph of ankle joint, oblique radiograph of foot, anteroposterior radiograph of foot, anteroposterior radiograph of shoulder joint, anteroposterior radiograph of elbow joint, lateral radiograph of elbow joint, posteroanterior radiograph of wrist joint, and lateral radiograph of wrist joint.

[0056] Preferably, in S4, a pre-trained medical image classification and recognition model is used to classify and recognize the cleaned data to obtain data sets of various categories, including:

[0057] The available DICOM medical images are input into the pre-trained medical image classification and recognition model, and the corresponding X-ray examination items are determined according to the output results. The available DICOM medical images belonging to the same X-ray examination items are classified separately to obtain data sets of each category.

[0058] (III) Beneficial effects

[0059] Compared with the prior art, the medical imaging big data management method based on artificial intelligence provided by the present invention has the following beneficial effects:

[0060] 1) Accuracy: Using artificial intelligence neural network models for data cleaning and classification and identification reduces interference from human factors and ensures the accuracy of the final datasets of each category, thereby enabling the construction of high-quality medical imaging datasets;

[0061] 2) Efficiency: Through automated data collection and analysis and processing, labor costs are greatly reduced and efficiency is improved, thereby enabling the efficient acquisition of high-quality medical imaging data sets. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the prior art descriptions are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention, and for ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0063] Figure 1 It is a schematic diagram of the process of the present invention. DETAILED DESCRIPTION

[0064] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0065] An artificial intelligence-based medical imaging big data governance method, such as Figure 1 As shown, S1 obtains multi-source heterogeneous data from the medical imaging cloud platform and performs data integration and data reconstruction, including:

[0066] Acquire multi-source heterogeneous data from the medical imaging cloud platform and convert DICOM medical images into common image formats.

[0067] S2. Desensitize the pre-processed data to protect patient information, including:

[0068] The patient information in the converted DICOM medical images is desensitized by using masking, transformation, rounding and deletion methods to protect the patient information.

[0069] S3. Use the pre-trained medical imaging data cleaning model to clean the desensitized data and remove unusable data, including:

[0070] S311, performing normalization processing including cropping, scaling and normalization on the desensitized DICOM medical image;

[0071] S312: Input the normalized DICOM medical images into a pre-trained medical image data cleaning model, determine unusable DICOM medical images based on the output results, and remove them.

[0072] Specifically, S3 uses a pre-trained medical imaging data cleaning model to clean the desensitized data before removing unusable data, including:

[0073] The model training of the medical image data cleaning model includes the following steps:

[0074] S301, annotating a DICOM medical image historical data set, and dividing the annotated DICOM medical image historical data set into a training set, a validation set, and a test set according to a preset ratio;

[0075] S302, setting the loss function and optimizer of the medical image data cleaning model;

[0076] S303, inputting the training set into the medical imaging data cleaning model for model training;

[0077] S304, calculating the loss value based on the loss function, and the optimizer updates the model parameters according to the loss value and the network gradient information;

[0078] S305. If the loss value is less than the preset threshold, the model training ends, and the current medical image data cleaning model is the trained medical image data cleaning model. Otherwise, the process returns to S303 and continues to use the training set for model training.

[0079] S306, inputting the validation set into the trained medical image data cleaning model, evaluating the generalization ability of the model by observing the performance on the validation set, and tuning the hyperparameters and structure of the model;

[0080] S307, inputting the test set into the tuned medical image data cleaning model to evaluate the model performance;

[0081] Among them, the medical image data cleaning model is built based on the Transformer model. The loss function uses the cross entropy loss function to measure the difference between the predicted value and the available label. The optimizer uses the AdamW optimizer. At the same time, Dropout or other regularization techniques are used during the model training process to prevent overfitting.

[0082] Specifically, in S301, the DICOM medical image historical data set is annotated, and the annotated DICOM medical image historical data set is divided into a training set, a validation set, and a test set according to a preset ratio, including:

[0083] S3011, performing normalization processing including cropping, scaling and normalization on the DICOM medical image historical data set;

[0084] S3012, expanding the normalized DICOM medical image historical data set by using rotation, flipping, and color jittering methods;

[0085] S3013, labeling each DICOM medical image in the expanded DICOM medical image historical data set, and setting an available label for each DICOM medical image;

[0086] S3014. Divide the annotated DICOM medical image historical data set into a training set, a validation set, and a test set according to a preset ratio.

[0087] In the technical solution of the present application, the medical image data cleaning model includes an input part, an encoder, a decoder and an output part;

[0088] In the input part, the DICOM medical image is divided into blocks of fixed size. Each block is regarded as a word. A linear transformation layer is used to convert each image block into a vector of fixed dimension. A position code is added after the vector so that the model can understand the relative position information between image blocks.

[0089] The encoder includes a multi-head self-attention sublayer and a first feed-forward fully connected sublayer. The multi-head self-attention sublayer uses a multi-head self-attention mechanism to capture the dependencies between image blocks. Each head focuses on a different part of the image, thereby obtaining global information of the image. The first feed-forward fully connected sublayer performs a nonlinear transformation on the output of the multi-head self-attention sublayer for further feature extraction.

[0090] A decoder, comprising a masked multi-head self-attention sublayer, an encoder-decoder attention sublayer, and a second feed-forward fully connected sublayer, wherein the masked multi-head self-attention sublayer adds a masking operation on the basis of the multi-head self-attention sublayer, so that the decoder can only rely on the first i-1 words when generating the i-th word, and maintains the autoregressive characteristics of the decoding process, the encoder-decoder attention sublayer interacts the input of the current position of the decoder with all positions of the encoder to obtain information related to the target sequence, and generates an output sequence corresponding to the input sequence, and the second feed-forward fully connected sublayer performs a nonlinear transformation on the output of the encoder-decoder attention sublayer;

[0091] The output part includes a fully connected layer and a Softmax layer. The fully connected layer converts the vector output by the decoder into the final output dimension, and the Softmax layer converts the output of the fully connected layer into a category probability distribution.

[0092] Among them, the multi-head self-attention sublayer and the first feed-forward fully connected sublayer in the encoder are connected with a residual connection and a normalization layer, and the masked multi-head self-attention sublayer, encoder-decoder attention sublayer and the second feed-forward fully connected sublayer in the decoder are connected with a residual connection and a normalization layer.

[0093] S4. Use the pre-trained medical image classification and recognition model to classify and recognize the cleaned data to obtain data sets of various categories, including:

[0094] The available DICOM medical images are input into the pre-trained medical image classification and recognition model, and the corresponding X-ray examination items are determined according to the output results. The available DICOM medical images belonging to the same X-ray examination items are classified separately to obtain data sets of each category.

[0095] Specifically, S4 uses the pre-trained medical image classification and recognition model to classify and recognize the cleaned data, and obtains the data sets of each category, including:

[0096] S401, annotating the available DICOM medical image historical data set, and dividing the annotated DICOM medical image historical data set into a training set, a validation set, and a test set according to a preset ratio;

[0097] S402, setting the loss function and optimizer of the medical image classification and recognition model;

[0098] S403, inputting the training set into the medical image classification and recognition model for model training;

[0099] S404, calculating the loss value based on the loss function, and the optimizer updates the model parameters according to the loss value and the network gradient information;

[0100] S405. If the loss value is less than the preset threshold, the model training ends, and the current medical image classification and recognition model is the trained medical image classification and recognition model. Otherwise, the process returns to S403 and continues to use the training set for model training.

[0101] S406, inputting the validation set into the trained medical image classification and recognition model, evaluating the generalization ability of the model by observing the performance on the validation set, and tuning the hyperparameters and structure of the model;

[0102] S407, inputting the test set into the tuned medical image classification and recognition model to evaluate the model performance;

[0103] Among them, the medical image classification and recognition model is built based on the ResNet50 model;

[0104] During the model training process, the learning rate was set to 0.001, the batch size was set to 128, the Adam optimizer was used as the optimizer, and the model training was stopped after 20 epochs;

[0105] When evaluating model performance, ROC curve and sensitivity SE are used as model performance evaluation indicators.

[0106] Specifically, in S401, the available DICOM medical image historical data set is annotated, and the annotated DICOM medical image historical data set is divided into a training set, a validation set and a test set according to a preset ratio, including:

[0107] S4011, labeling each DICOM medical image according to the X-ray examination item to which each DICOM medical image belongs in the available DICOM medical image historical data set, and setting a category label for each DICOM medical image;

[0108] S4012, dividing the annotated usable DICOM medical image historical data set into a training set, a validation set, and a test set according to a preset ratio of 7:1:2;

[0109] Among them, the X-ray examination items include posteroanterior radiograph of chest, lateral radiograph of thoracic spine, anteroposterior radiograph of lumbar spine, lateral radiograph of lumbar spine, upright radiograph of abdomen, anteroposterior radiograph of pelvis, anteroposterior radiograph of cervical spine, lateral radiograph of cervical spine, anteroposterior radiograph of thoracic spine, anteroposterior radiograph of sacrum and coccyx, lateral radiograph of sacrum and coccyx, posteroanterior radiograph of hand, oblique radiograph of hand, anteroposterior radiograph of hip joint, anteroposterior radiograph of knee joint, lateral radiograph of knee joint, anteroposterior radiograph of ankle joint, lateral radiograph of ankle joint, oblique radiograph of foot, anteroposterior radiograph of foot, anteroposterior radiograph of shoulder joint, anteroposterior radiograph of elbow joint, lateral radiograph of elbow joint, posteroanterior radiograph of wrist joint, and lateral radiograph of wrist joint.

[0110] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. Such modifications or replacements will not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A medical imaging big data management method based on artificial intelligence, characterized by: The following steps are involved: S1. Obtain multi-source heterogeneous data from the medical imaging cloud platform, and perform data integration and data reconstruction; S2. Desensitize the pre-processed data to protect patient information; S3. Use the pre-trained medical imaging data cleaning model to clean the desensitized data and remove unusable data; S4. Use the pre-trained medical image classification and recognition model to classify and recognize the cleaned data to obtain data sets of each category.

2. The method for managing medical image big data based on artificial intelligence according to claim 1, characterized in that: S1 obtains multi-source heterogeneous data from the medical imaging cloud platform and performs data integration and data reconstruction, including: Acquire multi-source heterogeneous data from the medical imaging cloud platform and convert DICOM medical images into common image formats.

3. The method for managing medical image big data based on artificial intelligence according to claim 1 is characterized in that: In S2, the pre-processed data is desensitized to protect patient information, including: The patient information in the converted DICOM medical images is desensitized by using masking, transformation, rounding and deletion methods to protect the patient information.

4. The method for managing medical image big data based on artificial intelligence according to claim 1, characterized in that: In S3, the pre-trained medical imaging data cleaning model is used to clean the desensitized data before removing the unusable data, including: The model training of the medical image data cleaning model includes the following steps: S301, annotating a DICOM medical image historical data set, and dividing the annotated DICOM medical image historical data set into a training set, a validation set, and a test set according to a preset ratio; S302, setting the loss function and optimizer of the medical image data cleaning model; S303, inputting the training set into the medical imaging data cleaning model for model training; S304, calculating the loss value based on the loss function, and the optimizer updates the model parameters according to the loss value and the network gradient information; S305. If the loss value is less than the preset threshold, the model training ends, and the current medical image data cleaning model is the trained medical image data cleaning model. Otherwise, the process returns to S303 and continues to use the training set for model training. S306, inputting the validation set into the trained medical image data cleaning model, evaluating the generalization ability of the model by observing the performance on the validation set, and tuning the hyperparameters and structure of the model; S307, inputting the test set into the tuned medical image data cleaning model to evaluate the model performance; Among them, the medical image data cleaning model is built based on the Transformer model. The loss function uses the cross entropy loss function to measure the difference between the predicted value and the available label. The optimizer uses the AdamW optimizer. At the same time, Dropout or other regularization techniques are used during the model training process to prevent overfitting.

5. The method for managing medical image big data based on artificial intelligence according to claim 4 is characterized in that: In S301, the DICOM medical image historical data set is annotated, and the annotated DICOM medical image historical data set is divided into a training set, a validation set, and a test set according to a preset ratio, including: S3011, performing normalization processing including cropping, scaling and normalization on the DICOM medical image historical data set; S3012, expanding the normalized DICOM medical image historical data set by using rotation, flipping, and color jittering methods; S3013, labeling each DICOM medical image in the expanded DICOM medical image historical data set, and setting an available label for each DICOM medical image; S3014. Divide the annotated DICOM medical image historical data set into a training set, a validation set, and a test set according to a preset ratio.

6. The method for managing medical image big data based on artificial intelligence according to claim 5 is characterized in that: In S3, the pre-trained medical imaging data cleaning model is used to clean the desensitized data and remove unusable data, including: S311, performing normalization processing including cropping, scaling and normalization on the desensitized DICOM medical image; S312: Input the normalized DICOM medical images into a pre-trained medical image data cleaning model, determine unusable DICOM medical images based on the output results, and remove them.

7. The method for managing medical image big data based on artificial intelligence according to any one of claims 4 to 6, characterized in that: The medical image data cleaning model includes an input part, an encoder, a decoder and an output part; In the input part, the DICOM medical image is divided into blocks of fixed size. Each block is regarded as a word. A linear transformation layer is used to convert each image block into a vector of fixed dimension. A position code is added after the vector so that the model can understand the relative position information between image blocks. The encoder includes a multi-head self-attention sublayer and a first feed-forward fully connected sublayer. The multi-head self-attention sublayer uses a multi-head self-attention mechanism to capture the dependencies between image blocks. Each head focuses on a different part of the image, thereby obtaining global information of the image. The first feed-forward fully connected sublayer performs a nonlinear transformation on the output of the multi-head self-attention sublayer for further feature extraction. A decoder, comprising a masked multi-head self-attention sublayer, an encoder-decoder attention sublayer, and a second feed-forward fully connected sublayer, wherein the masked multi-head self-attention sublayer adds a masking operation on the basis of the multi-head self-attention sublayer, so that the decoder can only rely on the first i-1 words when generating the i-th word, and maintains the autoregressive characteristics of the decoding process, the encoder-decoder attention sublayer interacts the input of the current position of the decoder with all positions of the encoder to obtain information related to the target sequence, and generates an output sequence corresponding to the input sequence, and the second feed-forward fully connected sublayer performs a nonlinear transformation on the output of the encoder-decoder attention sublayer; The output part includes a fully connected layer and a Softmax layer. The fully connected layer converts the vector output by the decoder into the final output dimension, and the Softmax layer converts the output of the fully connected layer into a category probability distribution. Among them, the multi-head self-attention sublayer and the first feed-forward fully connected sublayer in the encoder are connected with a residual connection and a normalization layer, and the masked multi-head self-attention sublayer, encoder-decoder attention sublayer and the second feed-forward fully connected sublayer in the decoder are connected with a residual connection and a normalization layer.

8. The method for managing medical image big data based on artificial intelligence according to claim 6 is characterized by: In S4, the pre-trained medical image classification and recognition model is used to classify and recognize the cleaned data. Before obtaining the data sets of each category, it includes: S401, annotating the available DICOM medical image historical data set, and dividing the annotated DICOM medical image historical data set into a training set, a validation set, and a test set according to a preset ratio; S402, setting the loss function and optimizer of the medical image classification and recognition model; S403, inputting the training set into the medical image classification and recognition model for model training; S404, calculating the loss value based on the loss function, and the optimizer updates the model parameters according to the loss value and the network gradient information; S405. If the loss value is less than the preset threshold, the model training ends, and the current medical image classification and recognition model is the trained medical image classification and recognition model. Otherwise, the process returns to S403 and continues to use the training set for model training. S406, inputting the validation set into the trained medical image classification and recognition model, evaluating the generalization ability of the model by observing the performance on the validation set, and tuning the hyperparameters and structure of the model; S407, inputting the test set into the tuned medical image classification and recognition model to evaluate the model performance; Among them, the medical image classification and recognition model is built based on the ResNet50 model; During the model training process, the learning rate was set to 0.001, the batch size was set to 128, the Adam optimizer was used as the optimizer, and the model training was stopped after 20 epochs; When evaluating model performance, ROC curve and sensitivity SE are used as model performance evaluation indicators.

9. The method for managing medical image big data based on artificial intelligence according to claim 8 is characterized by: In S401, the available DICOM medical image historical data set is annotated, and the annotated DICOM medical image historical data set is divided into a training set, a validation set, and a test set according to a preset ratio, including: S4011, labeling each DICOM medical image according to the X-ray examination item to which each DICOM medical image belongs in the available DICOM medical image historical data set, and setting a category label for each DICOM medical image; S4012, dividing the annotated usable DICOM medical image historical data set into a training set, a validation set, and a test set according to a preset ratio of 7:1:2; Among them, the X-ray examination items include posteroanterior radiograph of chest, lateral radiograph of thoracic spine, anteroposterior radiograph of lumbar spine, lateral radiograph of lumbar spine, upright radiograph of abdomen, anteroposterior radiograph of pelvis, anteroposterior radiograph of cervical spine, lateral radiograph of cervical spine, anteroposterior radiograph of thoracic spine, anteroposterior radiograph of sacrum and coccyx, lateral radiograph of sacrum and coccyx, posteroanterior radiograph of hand, oblique radiograph of hand, anteroposterior radiograph of hip joint, anteroposterior radiograph of knee joint, lateral radiograph of knee joint, anteroposterior radiograph of ankle joint, lateral radiograph of ankle joint, oblique radiograph of foot, anteroposterior radiograph of foot, anteroposterior radiograph of shoulder joint, anteroposterior radiograph of elbow joint, lateral radiograph of elbow joint, posteroanterior radiograph of wrist joint, and lateral radiograph of wrist joint.

10. The method for managing medical image big data based on artificial intelligence according to claim 9, characterized in that: In S4, the pre-trained medical image classification and recognition model is used to classify and recognize the cleaned data to obtain data sets of various categories, including: The available DICOM medical images are input into the pre-trained medical image classification and recognition model, and the corresponding X-ray examination items are determined according to the output results. The available DICOM medical images belonging to the same X-ray examination items are classified separately to obtain data sets of each category.

Images