Intelligent optimization method and device based on large generative medical model
By optimizing labeled medical images using a generative medical large-scale model and training with error correction data, the accuracy problem of medical image diagnostic models is solved, and efficient optimization of annotation results is achieved.
Patent Information
- Application Number
- CN202510923837.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-04
- Publication Date
- 2025-11-21
AI Technical Summary
The accuracy of existing medical image diagnosis and recognition models needs to be improved, and deep convolutional models cannot effectively correct errors on their own.
A generative medical large model is adopted. By acquiring labeled medical images and early warning information, the model is trained using pre-training data containing error correction data to optimize the labeling results.
It improves the accuracy of medical image annotation, reduces error output, shortens generation time, and maintains the efficiency of the model.
Smart Images
Figure CN120998433A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of medical image processing technology, and more specifically, to an intelligent optimization method and apparatus based on a generative medical large model. Background Technology
[0002] Current medical images are used for diagnosis, recognition, and planning, but the accuracy of this medical information needs to be improved. However, deep convolutional models can automatically correct errors in the output of the above information.
[0003] Therefore, there is an urgent need for a method that can modify and optimize medical information to overcome the above-mentioned shortcomings. Summary of the Invention
[0004] To address the aforementioned issues, the first aspect of this application provides an intelligent optimization method based on a generative medical large-scale model, which includes:
[0005] Obtain labeled medical images and their annotation information;
[0006] Obtain early warning information for labeled medical images;
[0007] Input the labeled medical images into the generative medical model to obtain optimized labeling results;
[0008] The generative medical big data model is trained based on pre-trained data that includes error correction data.
[0009] The second aspect of this application provides a manufacturing system based on an intelligent optimization method using a generative medical large model, comprising:
[0010] The image acquisition module is used to acquire labeled medical images and their annotation information;
[0011] The early warning acquisition module is used to acquire early warning information from labeled medical images;
[0012] The image optimization module is used to input the labeled medical images into the generative medical model to obtain optimized annotation results; the generative medical model is trained based on pre-trained data containing error correction data.
[0013] A third aspect of this application provides an electronic device, including: a memory and a processor; the memory being configurable to store a program, and the processor being coupled to the memory for executing the program in the memory for:
[0014] Obtain labeled medical images and their annotation information;
[0015] Obtain early warning information for labeled medical images;
[0016] Input the labeled medical images into the generative medical model to obtain optimized labeling results;
[0017] The generative medical big data model is trained based on pre-trained data that includes error correction data.
[0018] The fourth aspect of this application provides a computer-readable storage medium having a computer program stored thereon, the program being executed by a processor to implement the aforementioned intelligent optimization method based on a generative medical large model.
[0019] In this application, the labeled medical images and annotation information output by the deep convolutional model are optimized by using a generative medical large model, thereby obtaining the modified and optimized annotation results. Attached Figure Description
[0020] Figure 1 This is a flowchart of an intelligent optimization method based on a generative medical large model according to an embodiment of this application;
[0021] Figure 2 This is a schematic diagram of intelligent optimization based on a generative medical large model according to an embodiment of this application;
[0022] Figure 3 This is an architectural diagram of an intelligent optimization device based on a generative medical large model according to an embodiment of this application;
[0023] Figure 4 This is an architectural diagram of an electronic device according to an embodiment of this application. Detailed Implementation
[0024] To make the above-mentioned objects, features, and advantages of this application more apparent and understandable, specific embodiments of this application will be described in detail below with reference to the accompanying drawings. Although exemplary embodiments of this application are shown in the drawings, it should be understood that this application can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of this application and to fully convey the scope of this application to those skilled in the art.
[0025] It should be noted that, unless otherwise stated, the technical or scientific terms used in this application shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains.
[0026] This application provides an embodiment of the intelligent optimization method based on a generative medical large model as described above. The specific scheme of this method is as follows: Figure 1 Figure 2As shown, this method can be executed by an intelligent optimization device based on a generative medical model, which can be integrated into electronic devices such as computers, servers, computer clusters, and data centers. Combined with... Figure 1 As shown, the intelligent optimization method based on a generative medical large model includes:
[0027] S101, Obtain labeled medical images and labeling information;
[0028] In this application, labeled medical images and their corresponding labeling information are obtained as input data. The labeling information typically includes key information such as the location, type, and size of the lesion.
[0029] S102, Obtain early warning information for labeled medical images;
[0030] In this application, warning information refers to factors that may affect the accuracy of annotation or require special attention, or the review results of medical image annotation information.
[0031] In this application, the warning sources are: medical experts marking potentially problematic areas or labeling errors, or using a rule engine or deep learning model to detect common problems in the labeling (such as inconsistent units or blurred boundaries).
[0032] S103: Input the labeled medical images into the generative medical large-scale model to obtain the optimized labeling results;
[0033] The generative medical big data model is trained based on pre-trained data that includes error correction data.
[0034] In this application, obtaining early warning information means that the labeled medical images need to be corrected or optimized.
[0035] In this application, the generative medical large model is trained on pre-trained data containing error correction data, which can identify errors in the annotation and generate optimized annotation results.
[0036] In this application, generative medical large models (such as GPT, T5, Med-PaLM) are used to process text and image data, and multimodal models (such as CLIP, Flamingo) are combined to jointly analyze image and annotation information.
[0037] In this application, the labeled medical images and annotation information output by the deep convolutional model are optimized by using a generative medical large model, thereby obtaining the modified and optimized annotation results.
[0038] In one implementation, the training process of the automatic error correction large model includes:
[0039] Acquire pre-training data, which includes error correction data;
[0040] Pre-training of the large automatic error correction model based on pre-training data;
[0041] Obtain image annotation training data;
[0042] The large-scale automatic error correction model is fine-tuned based on the image-annotated training data to obtain the fine-tuned large-scale automatic error correction model.
[0043] It should be noted that conventional large models cannot optimize their own output. Even if additional reviews are conducted on parallel model structures internally, and the model is regenerated after the review, the accuracy of the model output cannot be improved in practice. Moreover, the review-regeneration process is extremely cumbersome, which greatly increases the generation time.
[0044] In this application, the goal of the pre-training data is to enable the model to learn to identify and correct errors, therefore it needs to include correct data, erroneous data, and corresponding error correction data.
[0045] In this application, the format of the pre-training data is: labeled "Error Data," indicating that the data is incorrect and needs to be changed to "Correct Data." That is, the pre-training data contains both erroneous data, error correction language, and the corrected data.
[0046] In practical use, the applicant determined that the automatic error correction large model trained on this type of pre-trained data can randomly discover its own output errors and correct them in a timely manner, and the overall output time is not significantly different from the correct output time of a conventional large model.
[0047] It should be noted that this error correction sample data is only included in the pre-training data and cannot be included in the fine-tuning process. In other words, the image annotation training data does not include error correction data.
[0048] In this application, the goal of pre-training is to enable the model to learn how to extract features from erroneous data and generate correct results.
[0049] In this application, the image annotation training data is high-quality annotation data specific to a particular domain, used to fine-tune the model to adapt it to a specific task.
[0050] In one implementation, acquiring the pre-training data includes:
[0051] Obtain a publicly available dataset, which includes at least medical images and text annotations;
[0052] Construct multiple weakly trained models;
[0053] When medical images are input into a weakly trained model, incorrect identification results are obtained.
[0054] The error identification results are inserted into the text annotations of medical images to generate pre-training data with error data and correction data.
[0055] In this application, the public dataset serves as the basis for generating pre-training data, typically including medical images and their corresponding text annotations.
[0056] In this application, the role of the weakly trained model is to generate error identification results to simulate common errors in annotation.
[0057] In this application, simple machine learning models (such as SVM, random forest) or lightweight deep learning models (such as small CNNs) are used to generate incorrect results, or large models that are not sufficiently trained (such as small versions of GPT and BERT) are used.
[0058] Error mode design in this application:
[0059] Descriptive errors: such as incorrect anatomical location ("left upper lobe" -> "right lower lobe").
[0060] Numerical errors: such as inconsistent size or shape attributes ("3cm×2cm" -> "5cm×4cm").
[0061] Logical errors: such as violating common medical knowledge ("no obvious abnormalities" -> "malignant tumor present").
[0062] Omission errors: such as failure to mark key lesion areas.
[0063] In this application, each weak model can generate different types of errors, ensuring the diversity of error data.
[0064] In this application, a weakly trained model is used to generate error identification results, which serve as the basis for generating error data subsequently.
[0065] Preferably, the error rate of the weak model is controlled to ensure that the generated error data is both challenging and not too outrageous.
[0066] In this application, the error identification results are combined with the original annotations to generate formatted pre-training data.
[0067] In one implementation, the text annotation is diagnostic information, surgical planning information, or image description information.
[0068] In one implementation, after obtaining the public dataset, the method further includes:
[0069] Random perturbations are added to the text annotations in public datasets to obtain erroneous data.
[0070] In one implementation, the step of adding random perturbation to the text annotations of the public dataset to obtain erroneous data includes:
[0071] Convert text annotations in public datasets into image format;
[0072] Adjust the image format to obtain the adjusted image;
[0073] The adjusted image is then inversely transformed to obtain erroneous data with random perturbations.
[0074] In this application, text annotations are converted into image format so that they can be manipulated at the pixel level.
[0075] In this application, in the text annotation, each character can be split into corresponding word combinations, and each word is a one-dimensional vector; the one-dimensional vector is arranged row by row to obtain multiple rows of data, which can be understood as a two-dimensional vector; each number in the one-dimensional vector (multiple numbers) is converted into a grayscale value, thus obtaining an image format composed of grayscale values.
[0076] Conversely, after obtaining the image format, it can be reverse mapped to text annotations using the methods described above.
[0077] In this way, through a text-image-text approach, deterministic adjustments in an image are transformed into random perturbations in the text. This results in a randomness while simultaneously maintaining a deterministic mapping relationship.
[0078] In this application, random perturbations are introduced into the image format to simulate common errors in annotation.
[0079] In this application, the adjusted image is converted back into text format, generating erroneous data with random perturbations.
[0080] In one implementation, adjusting the image format to obtain the adjusted image includes:
[0081] The image format is divided into blocks to obtain independent blocks;
[0082] For each independent block, obtain the first and second neighboring blocks with different spacings;
[0083] A first feature block is generated based on the independent block and the first neighboring block;
[0084] A second feature block is generated based on the independent block and the second neighboring block.
[0085] The first and second feature blocks are compressed to obtain a compressed block.
[0086] Iterate through all the individual blocks and generate an adjusted image based on the resulting compressed blocks.
[0087] In this application, the image format is divided into blocks, that is, the image format is divided into corresponding image blocks using a checkerboard pattern; wherein, the image block can be at the pixel level (that is, each pixel is an image block) or other levels, the specific division depends on the actual processing situation.
[0088] In this application, a sliding window or a fixed step size is used to divide the image into blocks of the same size.
[0089] Preferably, in this application, each image block has 1,001,000 pixels, thereby enabling more feature calculations between local regions while ensuring generation accuracy and reducing computational load.
[0090] In this application, an image block is selected as an independent block. The image blocks above, below, to the left, and to the right of this independent block are the first neighboring blocks; the image blocks one grid away from the top, bottom, left, and right of this independent block are the second neighboring blocks. The spacing between the first and second neighboring blocks and the independent block is different.
[0091] In this application, neighborhood information is extracted for each independent block to capture local structure.
[0092] In this application, generating the first feature block is to generate a local feature representation using an independent block and its first neighboring block. Specifically, this can be done by processing the independent block and the first neighboring block with convolutional layers and attention layers to obtain the first feature block.
[0093] In this application, the specific structure and parameters of the convolutional layer and attention layer can be obtained from the training data or determined according to the actual situation.
[0094] It should be noted that in this application, there are four first neighboring blocks and multiple first feature blocks.
[0095] In this application, the independent block and the first neighboring block are processed by convolutional layers and attention layers to obtain the first feature block. The specific process is as follows: the independent block and four neighboring blocks are concatenated together to form a multi-channel input, and the convolutional layer is used to extract features from the concatenated block; an important feature is enhanced by using a self-attention mechanism or a channel attention mechanism, the attention weight is calculated, and the output of the convolutional layer is weighted to enhance the important feature; the output of the attention layer is split into multiple feature blocks, and each feature block corresponds to the processing result of the independent block and at least one neighboring block.
[0096] In this application, a second feature block is generated to generate a broader local feature representation using the independent block and its second neighboring block. The specific generation process is the same as that of the first feature block, except that the parameters of the convolutional layer and the attention layer are different.
[0097] In this application, the generated feature blocks are compressed into a more compact representation to reduce computational cost while retaining key information. Feature compression is performed using pooling operations (such as max pooling or average pooling) or fully connected layers.
[0098] In this way, multiple first and second feature blocks are compressed into a single compressed block, which corresponds to the size and position of the individual blocks and is used to replace them. All image blocks are replaced by the compressed block, resulting in an adjusted image.
[0099] In this application, each image block of the image format is traversed to obtain the corresponding compressed block.
[0100] In this application, for image blocks / independent blocks near the edge, their first and second neighboring blocks are incomplete. In this case, the incomplete blocks are completed by copying the first and second neighboring blocks in their relative positions. For example, if the first neighboring block above an independent block does not exist, the first neighboring block below it is copied and used as the block above it.
[0101] In this application, by completing the image blocks, the processing accuracy of adjacent image blocks is greatly improved.
[0102] In this way, by adjusting, the randomness of the image format is increased locally.
[0103] In one embodiment, before adjusting the image format to obtain the adjusted image, the method further includes:
[0104] The image format is subjected to residual processing, and the image format after residual processing is input into the aforementioned adjustment steps for block adjustment.
[0105] The residual processing process includes:
[0106] The input image is adjusted using a residual module with two convolutions;
[0107] The spatial information of the adjusted input image is obtained by adopting the concept of multi-scale / multi-branch: after a 3×3 convolution, it is divided into four branches, and each branch performs convolution with different kernel sizes (3, 5, 7, 9);
[0108] All four branches obtain channel weights through the SE module, resulting in different weighted feature maps;
[0109] The weighted feature maps of each branch are concatenated to obtain the concatenated feature map;
[0110] The concatenated feature map is added to the adjusted input image to obtain the activation feature map;
[0111] The activation feature map is combined with the input image to obtain the input feature map for residual processing.
[0112] This allows for the activation of channel and spatial information in the image before format adjustment, improving the accuracy and local randomness of subsequent image format adjustments.
[0113] This application provides an intelligent optimization device based on a generative medical model, used to execute the intelligent optimization method based on a generative medical model described above. The intelligent optimization device based on a generative medical model will be described in detail below.
[0114] like Figure 3 As shown, the intelligent optimization device based on a generative medical large model includes:
[0115] Image acquisition module 101 is used to acquire labeled medical images and annotation information;
[0116] The early warning acquisition module 102 is used to acquire early warning information of labeled medical images;
[0117] Image optimization module 103 is used to input the labeled medical image into the generative medical big model to obtain the optimized labeling result; the generative medical big model is trained based on pre-training data containing error correction data.
[0118] In one implementation, the image optimization module 103 is further configured to:
[0119] Acquire pre-training data, which includes error correction data; pre-train the large-scale automatic error correction model based on the pre-training data; acquire image-annotated training data; fine-tune the large-scale automatic error correction model based on the image-annotated training data to obtain the fine-tuned large-scale automatic error correction model.
[0120] In one implementation, the image optimization module 103 is further configured to:
[0121] Obtain a public dataset, which includes at least medical images and text annotations; construct multiple weakly trained models; input the medical images into the weakly trained models to obtain misidentification results; insert the misidentification results into the text annotations of the medical images to generate pre-trained data with error data and correction data.
[0122] In one implementation, the text annotation is diagnostic information, surgical planning information, or image description information.
[0123] In one implementation, the image optimization module 103 is further configured to:
[0124] Random perturbations are added to the text annotations in public datasets to obtain erroneous data.
[0125] In one implementation, the image optimization module 103 is further configured to:
[0126] The text annotations in the public dataset are converted into image format; the image format is adjusted to obtain the adjusted image; the adjusted image is then inversely converted to obtain erroneous data with random perturbations.
[0127] In one implementation, the image optimization module 103 is further configured to:
[0128] The image format is divided into blocks to obtain independent blocks; for each independent block, a first neighboring block and a second neighboring block with different spacing are obtained; a first feature block is generated based on the independent blocks and the first neighboring blocks; a second feature block is generated based on the independent blocks and the second neighboring blocks; the first feature block and the second feature block are compressed to obtain a compressed block; all independent blocks are traversed, and an adjusted image is generated based on the obtained compressed blocks.
[0129] The intelligent optimization device based on a generative medical model provided in the above embodiments of this application corresponds to the intelligent optimization method based on a generative medical model provided in the embodiments of this application. Therefore, the specific content in this system corresponds to the intelligent optimization method based on a generative medical model. The specific content can be referred to the records in the intelligent optimization method based on a generative medical model, and will not be repeated in this application.
[0130] The intelligent optimization device based on a generative medical model provided in the above embodiments of this application and the intelligent optimization method based on a generative medical model provided in the embodiments of this application are based on the same inventive concept and have the same beneficial effects as the methods adopted, run or implemented by the applications stored therein.
[0131] The above describes the internal functions and structure of the intelligent optimization device based on a generative medical large model, such as... Figure 4 As shown, in practice, this intelligent optimization device based on a generative medical model can be implemented as an electronic device, including: a memory 301 and a processor 303.
[0132] Memory 301 can be configured to store a program.
[0133] Additionally, memory 301 can also be configured to store various other data to support operation on the electronic device. Examples of this data include instructions for any application or method used to operate on the electronic device, contact data, phonebook data, messages, pictures, videos, etc.
[0134] Memory 301 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. Processor 303, coupled to memory 301, is used to execute programs in memory 301 for:
[0135] Obtain labeled medical images and their annotation information;
[0136] Obtain early warning information for labeled medical images;
[0137] Input the labeled medical images into the generative medical model to obtain optimized labeling results;
[0138] The generative medical big data model is trained based on pre-trained data that includes error correction data.
[0139] In one implementation, the processor 303 is further configured to:
[0140] Acquire pre-training data, which includes error correction data; pre-train the large-scale automatic error correction model based on the pre-training data; acquire image-annotated training data; fine-tune the large-scale automatic error correction model based on the image-annotated training data to obtain the fine-tuned large-scale automatic error correction model.
[0141] In one implementation, the processor 303 is further configured to:
[0142] Obtain a public dataset, which includes at least medical images and text annotations; construct multiple weakly trained models; input the medical images into the weakly trained models to obtain misidentification results; insert the misidentification results into the text annotations of the medical images to generate pre-trained data with error data and correction data.
[0143] In one implementation, the text annotation is diagnostic information, surgical planning information, or image description information.
[0144] In one implementation, the processor 303 is further configured to:
[0145] Random perturbations are added to the text annotations in public datasets to obtain erroneous data.
[0146] In one implementation, the processor 303 is further configured to:
[0147] The text annotations in the public dataset are converted into image format; the image format is adjusted to obtain the adjusted image; the adjusted image is then inversely converted to obtain erroneous data with random perturbations.
[0148] In one implementation, the processor 303 is further configured to:
[0149] The image format is divided into blocks to obtain independent blocks; for each independent block, a first neighboring block and a second neighboring block with different spacing are obtained; a first feature block is generated based on the independent blocks and the first neighboring blocks; a second feature block is generated based on the independent blocks and the second neighboring blocks; the first feature block and the second feature block are compressed to obtain a compressed block; all independent blocks are traversed, and an adjusted image is generated based on the obtained compressed blocks.
[0150] In this application, Figure 4 The diagram only shows some components and does not mean that the electronic device includes only these components. Figure 4 The components shown.
[0151] The electronic device provided in this embodiment is based on the same inventive concept as the intelligent optimization method based on a generative medical large model provided in this application embodiment, and has the same beneficial effects as the methods adopted, run or implemented by the application stored therein.
[0152] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-readable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0153] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 The computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1The function specified in one or more boxes.
[0154] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.
[0155] In a typical configuration, a computing device includes one or more processors (CPUs), input / output interfaces, network interfaces, and memory. 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.
[0156] This application also provides a computer-readable storage medium corresponding to the intelligent optimization method based on a generative medical large model provided in the foregoing embodiments, wherein a computer program (i.e., a program product) is stored thereon. When the computer program is run by a processor, it executes the interactive image analysis assistance method for 3D aerial imaging provided in any of the foregoing embodiments.
[0157] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information by 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, magnetic magnetic 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.
[0158] The computer-readable storage medium provided in the above embodiments of this application and the interactive image analysis assistance method for 3D aerial imaging provided in the embodiments of this application are based on the same inventive concept and have the same beneficial effects as the methods adopted, run or implemented by the applications stored therein.
[0159] It should be noted that numerous specific details are set forth in the specification provided herein. However, it will be understood that embodiments of this application may be practiced without these specific details. In some instances, well-known structures and techniques have not been shown in detail so as not to obscure the understanding of this specification.
[0160] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.
[0161] The above description is merely an embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of this application should be included within the scope of the claims of this application.
Claims
1. A method for intelligent optimization based on a generative medical large model, characterized in that, include: Acquire medical images to be labeled; Input the medical image to be labeled into the automatic error correction model to obtain the labeled result after error correction; The automatic error correction model is trained based on pre-trained data that includes error correction data.
2. The intelligent optimization method based on a generative medical large model according to claim 1, characterized in that, The training process of the automatic error correction large model includes: Acquire pre-training data, which includes error correction data; Pre-training of the large automatic error correction model based on pre-training data; Obtain image annotation training data; The large-scale automatic error correction model is fine-tuned based on the image-annotated training data to obtain the fine-tuned large-scale automatic error correction model.
3. The intelligent optimization method based on a generative medical large model according to claim 2, characterized in that, The acquisition of pre-training data includes: Obtain a publicly available dataset, which includes at least medical images and text annotations; Construct multiple weakly trained models; When medical images are input into a weakly trained model, incorrect identification results are obtained. The error identification results are inserted into the text annotations of medical images to generate pre-training data with error data and correction data.
4. The intelligent optimization method based on a generative medical large model according to claim 3, characterized in that, The text annotations may be diagnostic information, surgical planning information, or image description information.
5. The intelligent optimization method based on a generative medical large model according to claim 3, characterized in that, After obtaining the public dataset, the process also includes: Random perturbations are added to the text annotations in public datasets to obtain erroneous data.
6. The intelligent optimization method based on a generative medical large model according to claim 5, characterized in that, The process of adding random perturbations to the text annotations in the public dataset to obtain erroneous data includes: Convert text annotations in public datasets into image format; Adjust the image format to obtain the adjusted image; The adjusted image is then inversely transformed to obtain erroneous data with random perturbations.
7. The intelligent optimization method based on a generative medical large model according to claim 6, characterized in that, The process of adjusting the image format to obtain the adjusted image includes: The image format is divided into blocks to obtain independent blocks; For each independent block, obtain the first and second neighboring blocks with different spacings; A first feature block is generated based on the independent block and the first neighboring block; A second feature block is generated based on the independent block and the second neighboring block. The first and second feature blocks are compressed to obtain a compressed block. Iterate through all the individual blocks and generate an adjusted image based on the resulting compressed blocks.
8. An intelligent optimization device based on a generative medical large model, characterized in that, include: The image acquisition module is used to acquire labeled medical images and their annotation information; The early warning acquisition module is used to acquire early warning information from labeled medical images; The image optimization module is used to input the labeled medical images into the generative medical model to obtain optimized annotation results; the generative medical model is trained based on pre-trained data containing error correction data.
9. An electronic device, characterized in that, include: Memory and processor; The memory is used to store programs; The processor, coupled to the memory, is used to execute the program for: Acquire medical images to be labeled; Input the medical image to be labeled into the automatic error correction model to obtain the labeled result after error correction; The automatic error correction model is trained based on pre-trained data that includes error correction data.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that, The program is executed by a processor to implement the intelligent optimization method based on a generative medical large model as described in any one of claims 17.