Microscopic examination sample image generation method, system and equipment based on artificial intelligence and medium

Abstract

The invention provides a microscopic examination sample image generation method, system and device based on artificial intelligence, and a medium, and the method comprises the steps: obtaining and based on an original sample image set of which the number of target fields is smaller than a preset threshold value, automatically constructing an image-text pair training set, carrying out the fine adjustment of a pre-trained multi-modal generation model through the image-text pair training set, and carrying out the fine adjustment of the pre-trained multi-modal generation model. At least freezing text encoder parameters of the model through fine tuning, performing low-rank adaptation updating on a visual correlation module of the model to obtain a domain-adaptive generative model, generating a sample image of a target domain based on the generative model, and performing fine tuning on the visual correlation module through freezing the text encoder parameters and combining with a fine tuning mode of low-rank adaptation updating. The method greatly reduces the total amount of parameters needing to be trained on a small sample, maintains the basic generation capability of the model, enables the finally obtained domain-adaptive generation model to effectively avoid overfitting, and can still generate a high-quality, high-diversity and accurate-feature domain sample image under the training of a very small number of domain samples.

Description

Technical Field

[0001] This application relates to the fields of artificial intelligence and medical image processing technology, and in particular to a method, system, device and medium for generating microscopic sample images based on artificial intelligence. Background Technology

[0002] In the field of AI-driven medical image analysis and clinical auxiliary diagnosis, high-quality, large-scale labeled data is the key to training high-precision models. However, in highly specialized scenarios such as clinical microscopy, it is extremely difficult to obtain sufficient sample images covering various pathological morphologies (such as parasite eggs and abnormal cells). The problem often faces issues such as sample scarcity, high collection costs, and privacy and security constraints, which severely limits the size of the dataset that can be used for model training, resulting in the "small sample" dilemma.

[0003] To address the problem of data scarcity, existing technologies mainly employ two types of methods: one is enhancement methods based on traditional image transformations (such as rotation, cropping, and color perturbation), but these methods generate images with insufficient diversity and limited realism, making it difficult to effectively improve the performance of downstream recognition models; the other is based on general artificial intelligence data generation methods, but when applied to this field (such as clinical endoscopy), these methods are often difficult to train stably or suffer from severe overfitting due to the extremely small amount of training data, resulting in their inability to reliably generate new samples with sufficient diversity and domain feature fidelity. Summary of the Invention

[0004] To address the aforementioned technical issues, this application provides an AI-based method, system, device, and medium for generating microscopic examination sample images. This AI-based method, system, device, and medium significantly reduces the total number of parameters required for training on small samples while maintaining the model's basic generative capabilities by employing a method of freezing text encoder parameters and combining low-rank adaptation to update visual relevance modules. This allows the resulting domain-adaptive generative model to effectively avoid overfitting and still generate high-quality, highly diverse, and feature-accurate domain sample images even with training on a very small number of domain samples.

[0005] The technical solution provided in this application is as follows: An AI-based method for generating microscopic sample images, applied to small sample scenarios, includes: Obtain a set of original sample images of the target domain, wherein the number of original sample images is less than a preset threshold; Based on the original sample image set, an image-text pair training set is automatically constructed. Using the image-text pair training set, the pre-trained multimodal generative model is fine-tuned, wherein the fine-tuning freezes at least the text encoder parameters of the generative model and performs low-rank adaptation updates on the visual relevance modules of the generative model to obtain a domain-adapted generative model. Based on the domain-adaptive generative model, sample images of the target domain are generated.

[0006] Optionally, the automatic construction of the image-text pair training set based on the original sample image set includes: The original sample image set is input into a pre-trained general image-text model to obtain text descriptions corresponding to each image in the original sample images, thus forming an image-text pair training set.

[0007] Optionally, the low-rank adaptation update is implemented using low-rank adaptation technology.

[0008] Optionally, generating sample images of the target domain based on the domain-adapted generative model includes: Input prompt words related to the target domain into the domain-adapted generative model and adjust the generation parameters to generate sample images of the target domain in batches.

[0009] Optionally, the generation parameters include at least one of: denoising intensity, weight coefficients of low-rank adaptation parameters, and number of iteration steps.

[0010] Optionally, after the step of obtaining the original sample image set of the target region, the method further includes: The original sample image set is preprocessed, and the preprocessing includes at least one of the following operations: denoising, super-resolution reconstruction, and image enhancement.

[0011] Optionally, the target field is the field of clinical microscopy, and the sample image is a clinical microscopic sample image containing at least one of parasite eggs, cells, crystals, or microorganisms.

[0012] This application provides an artificial intelligence-based microscopic sample image generation system for small sample scenarios, including: The acquisition module is used to acquire the original sample image set of the target domain, wherein the number of the original sample image set is less than a preset threshold. The construction module is used to automatically construct an image-text pair training set based on the original sample image set; The fine-tuning module is used to fine-tune the pre-trained multimodal generative model using the image-text pair training set. The fine-tuning at least freezes the text encoder parameters of the generative model and performs low-rank adaptation updates on the visual relevance modules of the generative model to obtain a domain-adapted generative model. The generation module is used to generate sample images of the target domain based on the domain-adapted generation model.

[0013] This application provides an electronic device, including: a processor, a memory, and a communication bus; The communication bus is used to realize the connection and communication between the processor and the memory; The processor is used to execute the AI-based microscopic sample image generation processing program stored in the memory to implement the steps of the AI-based microscopic sample image generation method as described above.

[0014] This application provides a readable storage medium storing computer-executable instructions, which, when loaded and executed by a processor, implement the steps of the artificial intelligence-based microscopic sample image generation method described in any of the preceding claims.

[0015] Compared with existing technologies, this application provides an artificial intelligence-based method, system, device, and medium for generating microscopic sample images. It acquires a set of original sample images from the target domain (the number of images in the original set is less than a preset threshold), automatically constructs an image-text pair training set based on the original image set, and fine-tunes a pre-trained multimodal generative model using the image-text pair training set. This fine-tuning involves at least freezing the text encoder parameters of the generative model and performing low-rank adaptation updates on the visual relevance modules of the generative model to obtain a domain-adapted generative model. Based on this domain-adapted generative model, sample images from the target domain are generated. This application, by employing a method of freezing text encoder parameters combined with low-rank adaptation updates on visual relevance modules, significantly reduces the total number of parameters required for training on small samples while maintaining the model's basic generative capabilities. This allows the final domain-adapted generative model to effectively avoid overfitting and still generate high-quality, highly diverse, and feature-accurate domain sample images even with training on a very small number of domain samples. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart illustrating an artificial intelligence-based method for generating microscopic sample images provided in an embodiment of this application. Figure 2Here is an example of an original sample image of Ascaris eggs provided in the embodiments of this application; Figure 3 This is an example of an original sample image of liver fluke eggs provided in the embodiments of this application; Figure 4 Here is an example of a generated Ascaris egg sample image provided in the embodiments of this application; Figure 5 This is an example of a generated liver fluke egg sample image provided in the embodiments of this application; Figure 6 This is a schematic diagram of the structure of an artificial intelligence-based microscopic sample image generation system provided in an embodiment of this application; Figure 7 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0018] To enable those skilled in the art to better understand the technical solutions in this application, the technical solutions in the embodiments of this application will be clearly and completely described below. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0019] It should be noted that when a component is referred to as being "fixed to" or "set on" another component, it can be directly on or indirectly set on the other component; when a component is referred to as being "connected to" another component, it can be directly connected to or indirectly connected to the other component.

[0020] It should be understood that the terms "length", "width", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this application.

[0021] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, "a plurality of" or "several" means two or more, unless otherwise explicitly specified.

[0022] It should be noted that the structures, proportions, sizes, etc., shown in the accompanying drawings of this specification are only for the purpose of assisting those skilled in the art in understanding and reading the content disclosed in the specification, and are not intended to limit the conditions under which this application can be implemented. Therefore, they have no substantial technical significance. Any modifications to the structure, changes in the proportions, or adjustments to the size should still fall within the scope of the technical content disclosed in this application, provided that they do not affect the effects and purposes that this application can produce.

[0023] like Figure 1 As shown, this application provides an artificial intelligence-based method for generating microscopic sample images, applicable to small sample scenarios, including: S11. Obtain the original sample image set of the target area. The number of original sample images is less than a preset threshold. In this embodiment, the target domain is a specialized field where data acquisition is difficult and samples are scarce, such as the field of clinical microscopy. The sample image can be a clinical microscopy sample image containing at least one of parasite eggs, cells, crystals, or microorganisms. The preset threshold can be set according to the scarcity of data in the domain, such as 1000, 500, 200, or 100, thereby quantitatively defining the small sample scenarios to which this method is applicable.

[0024] S12. Automatically construct an image-text pair training set based on the original sample image set; In this embodiment, it is preferable to achieve automated construction through a pre-trained general image-text model. The original sample image is input into the model, and the model automatically outputs a text description of the image content, key features, and image quality. This process solves the problems of high cost, low efficiency, and difficulty in ensuring consistency of manual annotation in small sample scenarios, and provides a structured and high-quality supervision signal for subsequent fine-tuning.

[0025] S13. Using the image-text pair training set, fine-tune the pre-trained multimodal generative model. The fine-tuning includes freezing the text encoder parameters of the generative model and performing low-rank adaptation updates on the visual-related modules of the generative model to obtain a domain-adapted generative model. In this embodiment, the pre-trained multimodal generative model can be a diffusion model such as Stable Diffusion. The fine-tuning process specifically includes: First, freeze the text encoder parameters: This is intended to preserve the general language understanding and semantic mapping capabilities that the model acquires during the pre-training phase, prevent them from being destroyed or forgotten on small samples, and ensure that the model still retains the foundation for responding to rich cue words after fine-tuning. Second, perform low-rank adaptation updates for the vision-related modules: This operation is preferably implemented using low-rank adaptation techniques. These techniques efficiently simulate the full update of model parameters by injecting a set of trainable low-rank decomposition matrices into the vision-related modules of the model, rather than directly updating the massive amount of original parameters.

[0026] The combined strategy of freezing the text side and updating the visual side with low-rank adaptation produces synergistic benefits: on the one hand, it greatly reduces the number of trainable parameters (usually by more than 90%), forcing the model to learn the feature distribution of the target domain through only limited adjustments to the visual parameters, fundamentally preventing overfitting, which is common on small sample data; on the other hand, it significantly reduces the computational resource requirements of the fine-tuning process. For example, in a training setting of 128x128 resolution and batch size of 5, the memory usage can be less than 10GB, making it possible to perform domain adaptation on large-scale generative models with limited hardware resources.

[0027] S14. Generate sample images of the target domain based on a domain-adaptive generative model.

[0028] In this embodiment, during generation, prompt words related to the target domain are input into the domain-adaptive generation model, and the clarity, style and diversity of the generated samples are finely controlled by adjusting generation parameters such as denoising intensity, weight coefficients of low-rank adaptation parameters, and number of iterations.

[0029] Because the model has learned the visual features of the target domain through the fine-tuning process in step S13, and retains its strong basic generation and semantic response capabilities due to the freezing of its text encoder, it can generate domain sample images in batches based on different prompt words. These images are highly realistic in visual features, have sufficient diversity in morphology, and are precisely semantically matched with the prompt words.

[0030] These high-quality, highly diverse generated images can be directly used as augmentation data to expand the training dataset of downstream recognition models (such as classifiers and detectors), thereby effectively solving the problem of scarce original data and significantly improving the recognition accuracy and generalization performance of downstream models in the target domain.

[0031] Compared with existing technologies, this application provides an artificial intelligence-based method, system, device, and medium for generating microscopic sample images. It acquires a set of original sample images from the target domain (the number of images in the original set is less than a preset threshold), automatically constructs an image-text pair training set based on the original image set, and fine-tunes a pre-trained multimodal generative model using the image-text pair training set. This fine-tuning involves at least freezing the text encoder parameters of the generative model and performing low-rank adaptation updates on the visual relevance modules of the generative model to obtain a domain-adapted generative model. Based on this domain-adapted generative model, sample images from the target domain are generated. This application, by employing a method of freezing text encoder parameters combined with low-rank adaptation updates on visual relevance modules, significantly reduces the total number of parameters required for training on small samples while maintaining the model's basic generative capabilities. This allows the final domain-adapted generative model to effectively avoid overfitting and still generate high-quality, highly diverse, and feature-accurate domain sample images even with training on a very small number of domain samples.

[0032] As one implementation method, in this embodiment of the application, an image-text pair training set is automatically constructed based on the original sample image set, including: The original sample image set is input into a pre-trained general image-text model to obtain text descriptions corresponding to each image in the original sample images, thus forming an image-text pair training set.

[0033] In this embodiment, the pre-trained general image-text model refers to a model trained on a large-scale public image-text dataset that possesses strong cross-modal understanding and generation capabilities. For example, models based on CLIP (Contrastive Language-Image Pre-training) or BLIP (Guided Language-Image Pre-training) architectures can automatically parse key visual elements, structural features, and overall style in input images and generate matching, descriptive natural language text. This automated process completely avoids the need for expensive, tedious, and error-prone manual annotation by experts in small-sample scenarios. It not only significantly improves the efficiency of data preparation but, more importantly, provides standardized, high-quality, and semantically rich supervision signals for subsequent model fine-tuning. These image-text pairs ensure that the generative model can align the visual features of the target domain with accurate semantic descriptions during the learning process, which is an important foundation for achieving high-quality, controllable sample generation.

[0034] As one implementation method, in this embodiment of the application, low-rank adaptation update is implemented using low-rank adaptation technology.

[0035] In this embodiment, the low-rank adaptation technique updates parameters by injecting a set of trainable low-rank decomposition matrices (usually represented as the product of two small matrices) next to the original weight matrix of the visual relevance module of the pre-trained multimodal generative model. During fine-tuning training, only the gradient of this injected low-rank matrix is ​​updated and optimized, while the original weight matrix of the model remains frozen. After training, the trained low-rank matrix is ​​added to the original weight matrix with specific weight coefficients to obtain the synthetic weights adapted to the target domain.

[0036] The key benefits of this technique are as follows: First, it achieves extremely high parameter efficiency. Since the number of parameters in a low-rank matrix is ​​much smaller than the original total number of parameters (usually reduced by more than 90%), the number of parameters required for training and storage is extremely small, greatly reducing the demand for computational memory and storage resources for fine-tuning. Second, it effectively ensures training stability and suppresses overfitting. The significantly reduced number of trainable parameters forces the model to learn the feature changes of the target domain in a highly compressed and generalized manner. At the same time, the powerful general visual priors of the original model are fully preserved. This combined effect enables the model to learn stably and efficiently on a very small number of domain samples, fundamentally avoiding the serious overfitting problem that inevitably occurs when fine-tuning all parameters directly in small sample scenarios.

[0037] As one implementation method, in this embodiment of the application, generating sample images of the target domain based on a domain-adaptive generative model includes: Input domain-adaptive generative models with prompts related to the target domain and adjust the generation parameters to generate sample images of the target domain in batches.

[0038] In this embodiment, prompts are used to precisely describe the visual features and content of the desired generated sample. For example, in a clinical microscopic stool sample generation experiment, to generate an image of Ascaris eggs, the input prompt keywords could be: "a close up of object, something in the middle, sometimes edge curve is coarse and blurry, rarely has brown body, high quality"; to generate an image of Clonorchis sinensis eggs, the input prompt keywords could be: "a close up of something, like a shell, silver object, water, sink drain, oyster, high quality". By inputting such fine-grained text instructions containing domain-specific keywords, the model can be precisely guided to generate images of pathological components that conform to specific morphological descriptions, thereby achieving a high degree of control over the content and style of the generated sample.

[0039] As one implementation method, in this application embodiment, the generation parameters include at least one of the following: denoising intensity, weight coefficient of low-rank adaptation parameter, and number of iteration steps.

[0040] In this embodiment, the generation effect can be optimized by comprehensively adjusting these parameters in the clinical microscopic sample generation experiment. For example, appropriately increasing the denoising intensity helps to make the generated egg edges clearer and the internal structure more distinct; adjusting the weight coefficient of the low-rank adaptation parameter (e.g., setting it between 0.5 and 1.0) can balance the presentation ratio between general morphology and specific fine-tuning features in the generated image, ensuring that the image has both the typical structure of the egg and details unique to the dataset (such as specific colors and textures); increasing the number of iterations can improve the convergence stability of the generation process, obtaining a final image with richer details and less noise. Through fine-tuning of such parameters, the visual fidelity, feature sharpness, and overall quality of the generated samples can be systematically controlled during batch generation, thereby meeting the strict requirements of downstream tasks for the diversity and realism of training data.

[0041] Experimental Example To verify the effectiveness of the technical solution of this application, an experiment was conducted using the generation of fecal parasite eggs from clinical microscopic examination as an example.

[0042] The experiment selected two categories: Ascaris eggs and Liver fluke eggs. One hundred original sample images were selected from each category to form the training set. Some of the original sample images can be found in [link to original image]. Figure 2(Ascaris eggs) and Figure 3 (Liver fluke eggs).

[0043] Using the method provided in this application, a pre-trained multimodal generative model is fine-tuned, and then generated by inputting specific prompt keywords. See the examples of some of the generated sample images. Figure 4 (Ascaris eggs) and Figure 5 (Liver fluke eggs) The results in the figure show that the generated images are highly consistent with the real samples in terms of key visual features such as morphology and texture, and have good diversity. This effectively verifies the effectiveness and practicality of the method in this application in generating high-quality, highly diverse and highly controllable domain samples under extremely low resource conditions.

[0044] As one implementation method, in this embodiment of the application, after the step of obtaining the original sample image set of the target region, the method further includes: S21. Preprocess the original sample image set. The preprocessing includes at least one of the following operations: denoising, super-resolution reconstruction, and image enhancement.

[0045] In this embodiment, the preprocessing operation aims to improve the basic visual quality of the original sample image set, laying the foundation for subsequent automated construction of high-quality training data and stable model fine-tuning. Specifically: Denoising: Random noise introduced by acquisition equipment, environment or transmission process is suppressed by filtering algorithms (such as non-local mean denoising, wavelet denoising) to improve the image signal-to-noise ratio and make key features more prominent.

[0046] Super-resolution reconstruction: For images with low resolution and blurred details, high-frequency details are reconstructed based on deep learning models (such as the ESRGAN model) or interpolation algorithms to improve image clarity and size, ensuring that subsequent models can learn sufficient structural information.

[0047] Image enhancement: By adjusting contrast, brightness, sharpness, or performing color equalization, the overall visual effect of an image is improved, making target features more consistent and easier to identify under different imaging conditions.

[0048] In small-sample scenarios, each original image is extremely valuable and its information density is crucial. Through the above preprocessing, the effective information in the limited data can be purified and enhanced to the maximum extent. This enables the general image-text model to more accurately identify and describe the image content in the subsequent automated text description generation stage. In the model fine-tuning stage, it provides the model with clearer and more consistent learning signals, which together help to generate higher quality and more realistic domain sample images.

[0049] As one implementation method, in this application embodiment, the target field is the field of clinical microscopy, and the sample image is a clinical microscopic sample image containing at least one of parasite eggs, cells, crystals or microorganisms.

[0050] In this embodiment, clinical microscopy is a typical scenario in medical testing where data scarcity and professional challenges are particularly prominent. The sample images in this field have the following significant characteristics: First, sample acquisition is extremely difficult, involving patient privacy, the rarity of pathological specimens, and compliant collection procedures, resulting in the amount of raw image data usually being far less than that of conventional visual tasks; Second, annotation requirements are extremely high, requiring the professional knowledge of senior laboratory physicians, and manual annotation is costly and inefficient; Third, morphological features are complex and subtle, such as the specific shape, texture, and contents of parasite eggs, and the nucleocytoplasmic ratio and staining characteristics of cells, which place stringent requirements on image quality and feature fidelity. These features together constitute a barrier that traditional data augmentation and general artificial intelligence generation methods cannot overcome.

[0051] The technical solution provided in this application, especially the efficient fine-tuning strategy for automatic image-text pair construction, freezing of the text side and low-rank updates for small samples, is designed to overcome the above-mentioned barriers. This method can automatically generate a large number of high-quality and highly diverse microscopic examination samples using only hundreds of original images, and the generated images have high fidelity and controllability in key morphological features.

[0052] like Figure 6 As shown, this application provides an artificial intelligence-based microscopic sample image generation system for small sample scenarios, including: The acquisition module 61 is used to acquire the original sample image set of the target area, wherein the number of original sample images is less than a preset threshold. Module 62 is used to automatically construct an image-text pair training set based on the original sample image set; The fine-tuning module 63 is used to fine-tune the pre-trained multimodal generative model using the image-text pair training set. The fine-tuning freezes at least the text encoder parameters of the generative model and performs low-rank adaptation updates on the visual related modules of the generative model to obtain a domain-adapted generative model. The generation module 64 is used to generate sample images of the target domain based on a domain-adaptive generative model.

[0053] like Figure 7 As shown, this application provides an electronic device, including: a processor 71, a memory 72, and a communication bus 73; Communication bus 73 is used to realize the connection and communication between processor 71 and memory 72; The processor 71 is used to execute the AI-based microscopic sample image generation processing program stored in the memory 72 to implement the steps of any of the AI-based microscopic sample image generation methods described above.

[0054] This application provides a readable storage medium storing computer-executable instructions. When the computer-executable instructions are loaded and executed by a processor, they implement the steps of any of the above-described artificial intelligence-based microscopic sample image generation methods.

[0055] It should be understood that the use of terms such as "system," "device," "unit," and / or "module" in this application is merely one method of distinguishing different components, elements, parts, sections, or assemblies at different levels. However, if other terms can achieve the same purpose, they may be replaced by other expressions.

[0056] The embodiments in this specification are described in a progressive manner, with each embodiment focusing on the related aspects. For any differences between the embodiments, or for the same or similar parts between the embodiments, please refer to each other.

[0057] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. An artificial intelligence-based method for generating microscopic sample images, applied to small sample scenarios, characterized in that, include: Obtain a set of original sample images of the target domain, wherein the number of original sample images is less than a preset threshold; Based on the original sample image set, an image-text pair training set is automatically constructed. Using the image-text pair training set, the pre-trained multimodal generative model is fine-tuned, wherein the fine-tuning freezes at least the text encoder parameters of the generative model and performs low-rank adaptation updates on the visual relevance modules of the generative model to obtain a domain-adapted generative model. Based on the domain-adaptive generative model, sample images of the target domain are generated.

2. The method according to claim 1, characterized in that, The automatic construction of an image-text pair training set based on the original sample image set includes: The original sample image set is input into a pre-trained general image-text model to obtain text descriptions corresponding to each image in the original sample images, thus forming an image-text pair training set.

3. The method according to claim 1 or 2, characterized in that, The low-rank adaptation update is implemented using low-rank adaptation technology.

4. The method according to claim 1 or 2, characterized in that, The domain-adaptive generative model generates sample images of the target domain, including: Input prompt words related to the target domain into the domain-adapted generative model and adjust the generation parameters to generate sample images of the target domain in batches.

5. The method according to claim 4, characterized in that, The generated parameters include at least one of the following: denoising intensity, weight coefficients of low-rank adaptation parameters, and number of iteration steps.

6. The method according to claim 1 or 2, characterized in that, Following the step of obtaining the original sample image set of the target region, the method further includes: The original sample image set is preprocessed, and the preprocessing includes at least one of the following operations: denoising, super-resolution reconstruction, and image enhancement.

7. The method according to claim 1 or 2, characterized in that, The target field is the field of clinical microscopy, and the sample image is a clinical microscopic sample image containing at least one of parasite eggs, cells, crystals, or microorganisms.

8. An artificial intelligence-based microscopic sample image generation system, applied to small sample scenarios, characterized in that, include: The acquisition module is used to acquire the original sample image set of the target domain, wherein the number of the original sample image set is less than a preset threshold. The construction module is used to automatically construct an image-text pair training set based on the original sample image set; The fine-tuning module is used to fine-tune the pre-trained multimodal generative model using the image-text pair training set. The fine-tuning at least freezes the text encoder parameters of the generative model and performs low-rank adaptation updates on the visual relevance modules of the generative model to obtain a domain-adapted generative model. The generation module is used to generate sample images of the target domain based on the domain-adapted generation model.

9. An electronic device, characterized in that, include: Processor, memory, and communication bus; The communication bus is used to realize the connection and communication between the processor and the memory; The processor is used to execute the AI-based microscopic sample image generation processing program stored in the memory to implement the steps of the AI-based microscopic sample image generation method as described in any one of claims 1-7.

10. A readable storage medium, characterized in that, The readable storage medium stores computer-executable instructions, which, when loaded and executed by a processor, implement the steps of the artificial intelligence-based microscopic sample image generation method as described in any one of claims 1 to 7.

Images